[agi] Re: [agi] P≠NP
Does anyone have any comments on this proof? I don't have the mathematical background to tell if it is correct. But it seems related to the idea from algorithmic information theory that the worst case complexity for any algorithm is equal to the average case for compressed inputs. Then to show that P != NP you would show that SAT (specifically 9-SAT) with compressed inputs has exponential average case complexity. That is not quite the approach the paper takes, probably because compression is not computable. -- Matt Mahoney, matmaho...@yahoo.com From: Kaj Sotala xue...@gmail.com To: agi agi@v2.listbox.com Sent: Thu, August 12, 2010 2:18:13 AM Subject: [agi] Re: [agi] P≠NP 2010/8/12 John G. Rose johnr...@polyplexic.com BTW here is the latest one: http://www.win.tue.nl/~gwoegi/P-versus-NP/Deolalikar.pdf See also: http://www.ugcs.caltech.edu/~stansife/pnp.html - brief summary of the proof Discussion about whether it's correct: http://rjlipton.wordpress.com/2010/08/08/a-proof-that-p-is-not-equal-to-np/ http://rjlipton.wordpress.com/2010/08/09/issues-in-the-proof-that-p?np/ http://rjlipton.wordpress.com/2010/08/10/update-on-deolalikars-proof-that-p≠np/ http://rjlipton.wordpress.com/2010/08/11/deolalikar-responds-to-issues-about-his-p≠np-proof/ http://news.ycombinator.com/item?id=1585850 Wiki page summarizing a lot of the discussion, as well as collecting many of the links above: http://michaelnielsen.org/polymath1/index.php?title=Deolalikar%27s_P!%3DNP_paper#Does_the_argument_prove_too_much.3F --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com * Open Link in New Tab * Download --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] Help requested: Making a list of (non-robotic) AGI low hanging fruit apps
Wouldn't it depend on the other researcher's area of expertise? -- Matt Mahoney, matmaho...@yahoo.com From: Ben Goertzel b...@goertzel.org To: agi agi@v2.listbox.com Sent: Sat, August 7, 2010 9:10:23 PM Subject: [agi] Help requested: Making a list of (non-robotic) AGI low hanging fruit apps Hi, A fellow AGI researcher sent me this request, so I figured I'd throw it out to you guys I'm putting together an AGI pitch for investors and thinking of low hanging fruit applications to argue for. I'm intentionally not involving any mechanics (robots, moving parts, etc.). I'm focusing on voice (i.e. conversational agents) and perhaps vision-based systems. Hellen Keller AGI, if you will :) Along those lines, I'd like any ideas you may have that would fall under this description. I need to substantiate the case for such AGI technology by making an argument for high-value apps. All ideas are welcome. All serious responses will be appreciated!! Also, I would be grateful if we could keep this thread closely focused on direct answers to this question, rather than digressive discussions on Helen Keller, the nature of AGI, the definition of AGI versus narrow AI, the achievability or unachievability of AGI, etc. etc. If you think the question is bad or meaningless or unclear or whatever, that's fine, but please start a new thread with a different subject line to make your point. If the discussion is useful, my intention is to mine the answers into a compact list to convey to him Thanks! Ben G --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] Epiphany - Statements of Stupidity
Mike Tintner wrote: What will be the SIMPLEST thing that will mark the first sign of AGI ? - Given that there are zero but zero examples of AGI. Machines have already surpassed human intelligence. If you don't think so, try this IQ test. http://mattmahoney.net/iq/ Or do you prefer to define intelligence as more like a human? In that case I agree that AGI will never happen. No machine will ever be more like a human than a human. I really don't care how you define it. Either way, computers are profoundly affecting the way people interact with each other and with the world. Where is the threshold when machines do most of our thinking for us? Who cares as long as the machines still give us the feeling that we are in charge. -- Matt Mahoney, matmaho...@yahoo.com From: Mike Tintner tint...@blueyonder.co.uk To: agi agi@v2.listbox.com Sent: Fri, August 6, 2010 5:57:33 AM Subject: Re: [agi] Epiphany - Statements of Stupidity sTEVE:I have posted plenty about statements of ignorance, our probable inability to comprehend what an advanced intelligence might be thinking, What will be the SIMPLEST thing that will mark the first sign of AGI ? - Given that there are zero but zero examples of AGI. Don't you think it would be a good idea to begin at the beginning? With initial AGI? Rather than advanced AGI? agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] Comments On My Skepticism of Solomonoff Induction
Jim, see http://www.scholarpedia.org/article/Algorithmic_probability I think this answers your questions. -- Matt Mahoney, matmaho...@yahoo.com From: Jim Bromer jimbro...@gmail.com To: agi agi@v2.listbox.com Sent: Fri, August 6, 2010 2:18:09 PM Subject: Re: [agi] Comments On My Skepticism of Solomonoff Induction I meant: Did Solomonoff's original idea use randomization to determine the bits of the programs that are used to produce the prior probabilities? I think that the answer to that is obviously no. The randomization of the next bit would used in the test of the prior probabilities as done using a random sampling. He probably found that students who had some familiarity with statistics would initially assume that the prior probability was based on some subset of possible programs as would be expected from a typical sample, so he gave this statistics type of definition to emphasize the extent of what he had in mind. I asked this question just to make sure that I understood what Solomonoff Induction was, because Abram had made some statement indicating that I really didn't know. Remember, this particular branch of the discussion was originally centered around the question of whether Solomonoff Induction would be convergent, even given a way around the incomputability of finding only those programs that halted. So while the random testing of the prior probabilities is of interest to me, I wanted to make sure that there is no evidence that Solomonoff Induction is convergent. I am not being petty about this, but I also needed to make sure that I understood what Solomonoff Induction is. I am interested in hearing your ideas about your variation of Solomonoff Induction because your convergent series, in this context, was interesting. Jim Bromer On Fri, Aug 6, 2010 at 6:50 AM, Jim Bromer jimbro...@gmail.com wrote: Jim: So, did Solomonoff's original idea involve randomizing whether the next bit would be a 1 or a 0 in the program? Abram: Yep. I meant, did Solomonoff's original idea involve randomizing whether the next bit in the program's that are originally used to produce the prior probabilities involve the use of randomizing whether the next bit would be a 1 or a 0? I have not been able to find any evidence that it was. I thought that my question was clear but on second thought I guess it wasn't. I think that the part about the coin flips was only a method to express that he was interested in the probability that a particular string would be produced from all possible programs, so that when actually testing the prior probability of a particular string the program that was to be run would have to be randomly generated. Jim Bromer On Wed, Aug 4, 2010 at 10:27 PM, Abram Demski abramdem...@gmail.com wrote: Jim, Your function may be convergent but it is not a probability. True! All the possibilities sum to less than 1. There are ways of addressing this (ie, multiply by a normalizing constant which must also be approximated in a convergent manner), but for the most part adherents of Solomonoff induction don't worry about this too much. What we care about, mostly, is comparing different hyotheses to decide which to favor. The normalizing constant doesn't help us here, so it usually isn't mentioned. You said that Solomonoff's original construction involved flipping a coin for the next bit. What good does that do? Your intuition is that running totally random programs to get predictions will just produce garbage, and that is fine. The idea of Solomonoff induction, though, is that it will produce systematically less garbage than just flipping coins to get predictions. Most of the garbage programs will be knocked out of the running by the data itself. This is supposed to be the least garbage we can manage without domain-specific knowledge This is backed up with the proof of dominance, which we haven't talked about yet, but which is really the key argument for the optimality of Solomonoff induction. And how does that prove that his original idea was convergent? The proofs of equivalence between all the different formulations of Solomonoff induction are something I haven't cared to look into too deeply. Since his idea is incomputable, there are no algorithms that can be run to demonstrate what he was talking about so the basic idea is papered with all sorts of unverifiable approximations. I gave you a proof of convergence for one such approximation, and if you wish I can modify it to include a normalizing constant to ensure that it is a probability measure. It would be helpful to me if your criticisms were more specific to that proof. So, did Solomonoff's original idea involve randomizing whether the next bit would be a 1 or a 0 in the program? Yep. Even ignoring the halting problem what kind of result would that give? Well, the general idea is this. An even
Re: [agi] Walker Lake
Samantha Atkins wrote: No it hasn't. People want public surveillance. Guess I am not people then. Then why are you posting your response to a public forum instead of replying by encrypted private email? People want their words to be available to the world. I don't think the global brain needs to know exactly how often I have sex or with whom or in what varieties. Do you? A home surveillance system needs to know who is in your house and whether they belong there. If it is intelligent then it will know that you prefer not to have video of you having sex broadcast on the internet. At the same time, it has to recognize what you are doing. Public surveillance is less objectionable because it will be two-way and can't be abused. If someone searches for information about you, then you get a notification of who it was and what they learned. I describe how this works in http://mattmahoney.net/agi2.html No humans will control it and it is not going to be that expensive. Humans will eventually lose control of anything that is smarter than them. But we should delay that as long as possible by making the required threshold the organized intelligence of all humanity, and make that organization as efficient as possible. The cost is on the order of $1 quadrillion because the knowledge that AGI needs is mostly in billions of human brains and there is no quick way to extract it. -- Matt Mahoney, matmaho...@yahoo.com From: Samantha Atkins sjatk...@gmail.com To: agi agi@v2.listbox.com Sent: Tue, August 3, 2010 6:49:34 AM Subject: Re: [agi] Walker Lake Matt Mahoney wrote: Steve Richfield wrote: How about an international ban on the deployment of all unmanned and automated weapons? How about a ban on suicide bombers to level the playing field? 1984 has truly arrived. No it hasn't. People want public surveillance. Guess I am not people then. Actually I think surveillance is inevitable given current and all but certain future tech. However, I recognize that human beings today, and especially their governments, are not remotely ready for it. To be ready for it at the very least the State would have to consider a great number of things none of its business to attempt to legislate for or against. As it is with the current incredible number of arcane laws on the books it would be very easy to see the already ridiculously large prison population of the US double. Also, please note that full surveillance means no successful rebellion no matter how bad the powers that be become and how ineffectual the means that let remain legal are to change things. Ever. It is also necessary for AGI. In order for machines to do what you want, they have to know what you know. It is not necessary to have every waking moment surveilled in order to have AGI know what we want. In order for a global brain to use that knowledge, it has to be public. I don't think the global brain needs to know exactly how often I have sex or with whom or in what varieties. Do you? AGI has to be a global brain because it is too expensive to build any other way, and because it would be too dangerous ifthe whole world didn't control it. No humans will control it and it is not going to be that expensive. - samantha agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] Walker Lake
Steve Richfield wrote: Disaster scenarios aside, what would YOU have YOUR AGI do to navigate this future? It won't be my AGI. If it were, I would be a despot and billions of people would suffer, just like if any other person ruled the world with absolute power. We will be much better off if everyone has a voice, and we have an AGI that makes that voice available to everyone else. -- Matt Mahoney, matmaho...@yahoo.com From: Steve Richfield steve.richfi...@gmail.com To: agi agi@v2.listbox.com Sent: Mon, August 2, 2010 10:03:27 PM Subject: Re: [agi] Walker Lake Matt, On Mon, Aug 2, 2010 at 1:10 PM, Matt Mahoney matmaho...@yahoo.com wrote: Steve Richfield wrote: How about an international ban on the deployment of all unmanned and automated weapons? How about a ban on suicide bombers to level the playing field? Of course we already have that. Unfortunately, one begets the other. Hence, we seem to have a choice, neither or both. I vote for neither. 1984 has truly arrived. No it hasn't. People want public surveillance. I'm not sure what you mean by public surveillance. Monitoring private phone calls? Monitoring otherwise unused web cams? Monitoring your output when you use the toilet? Where, if anywhere, do YOU draw the line? It is also necessary for AGI. In order for machines to do what you want, they have to know what you know. Unfortunately, knowing everything, any use of this information will either be to my benefit, or my detriment. Do you foresee any way to limit use to only beneficial use? BTW, decades ago I developed the plan of, when my kids got in some sort of trouble in school or elsewhere, to represent their interests as well as possible, regardless of whether I agreed with them or not. This worked EXTREMELY well for me, and for several other families who have tried this. The point is that to successfully represent their interests, I had to know what was happening. Potential embarrassment and explainability limited the kids' actions. I wonder if the same would work for AGIs? In order for a global brain to use that knowledge, it has to be public. Again, where do you draw the line between public and private? AGI has to be a global brain because it is too expensive to build any other way, and because it would be too dangerous if the whole world didn't control it. I'm not sure what you mean by control. Here is the BIG question in my own mind, that I have asked in various ways, so far without any recognizable answer: There are plainly lots of things wrong with our society. We got here by doing what we wanted, and by having our representatives do what we wanted them to do. Clearly some social re-engineering is in our future, if we are to thrive in the foreseeable future. All changes are resisted by some, and I suspect that some needed changes will be resisted by most, and perhaps nearly everyone. Disaster scenarios aside, what would YOU have YOUR AGI do to navigate this future? To help guide your answer, I see that the various proposed systems of ethics would prevent breaking the eggs needed to make a good futuristic omelet. I suspect that completely democratic systems have run their course. Against this is letting AGI loose has its own unfathomable hazards. I've been hanging around here for quite a while, and I don't yet see any success path to work toward. I'm on your side in that any successful AGI would have to have the information and the POWER to succeed, akin to Colossus, the Forbin Project, which I personally see as more of a success story than a horror scenario. Absent that, AGIs will only add to our present problems. What is the success path that you see? Steve agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] AGI Int'l Relations
Steve Richfield wrote: I would feel a **LOT** better if someone explained SOME scenario to eventually emerge from our current economic mess. What economic mess? http://www.google.com/publicdata?ds=wb-wdictype=lstrail=falsenselm=hmet_y=ny_gdp_mktp_cdscale_y=linind_y=falserdim=countryidim=country:USAtdim=truetstart=-31561920tunit=Ytlen=48hl=endl=en Unemployment appears to be permanent and getting worse, When you pay people not to work, they are less inclined to work. -- Matt Mahoney, matmaho...@yahoo.com From: Steve Richfield steve.richfi...@gmail.com To: agi agi@v2.listbox.com Sent: Mon, August 2, 2010 11:54:25 AM Subject: Re: [agi] AGI Int'l Relations Jan I can see that I didn't state one of my points clearly enough... On Sun, Aug 1, 2010 at 3:04 PM, Jan Klauck jkla...@uni-osnabrueck.de wrote: My simple (and completely unacceptable) cure for this is to tax savings, to force the money back into the economy. You have either consumption or savings. The savings are put back into the economy in form of credits to those who invest the money. Our present economic problem is that those credits aren't being turned over fast enough to keep the economic engine running well. At present, with present systems in place, there is little motivation to quickly turn over one's wealth, and lots of motivation to very carefully protect it. The result is that most of the wealth of the world is just sitting there in various accounts, and is NOT being spent/invested on various business propositions to benefit the population of the world. We need to do SOMETHING to get the wealth out of the metaphorical mattresses and back into the economy. Taxation is about the only effective tool that the government hasn't already dulled beyond utility. However, taxation doesn't stand a chance without the cooperation of other countries to do the same. There seems to be enough lobbying power in the hands of those with the money to stop any such efforts, or at least to leave enough safe havens to make such efforts futile. I would feel a **LOT** better if someone explained SOME scenario to eventually emerge from our current economic mess. Unemployment appears to be permanent and getting worse, as does the research situation. All I hear are people citing stock prices and claiming that the economy is turning around, when I see little connection between stock prices and on-the-street economy. This is an IR problems of monumental proportions. What would YOU do about it? Steve agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] Walker Lake
Steve Richfield wrote: How about an international ban on the deployment of all unmanned and automated weapons? How about a ban on suicide bombers to level the playing field? 1984 has truly arrived. No it hasn't. People want public surveillance. It is also necessary for AGI. In order for machines to do what you want, they have to know what you know. In order for a global brain to use that knowledge, it has to be public. AGI has to be a global brain because it is too expensive to build any other way, and because it would be too dangerous if the whole world didn't control it. -- Matt Mahoney, matmaho...@yahoo.com From: Steve Richfield steve.richfi...@gmail.com To: agi agi@v2.listbox.com Sent: Mon, August 2, 2010 10:40:20 AM Subject: [agi] Walker Lake Sometime when you are flying between the northwest US to/from Las Vegas, look out your window as you fly over Walker Lake in eastern Nevada. At the south end you will see a system of roads leading to tiny buildings, all surrounded by military security. From what I have been able to figure out, you will find the U.S. arsenal of chemical and biological weapons housed there. No, we are not now making these weapons, but neither are we disposing of them. Similarly, there has been discussion of developing advanced military technology using AGI and other computer-related methods. I believe that these efforts are fundamentally anti-democratic, as they allow a small number of people to control a large number of people. Gone are the days when people voted with their swords. We now have the best government that money can buy monitoring our every email, including this one, to identify anyone resisting such efforts. 1984 has truly arrived. This can only lead to a horrible end to freedom, with AGIs doing their part and more. Like chemical and biological weapons, unmanned and automated weapons should be BANNED. Unfortunately, doing so would provide a window of opportunity for others to deploy them. However, if we make these and stick them in yet another building at the south end of Walker Lake, we would be ready in case other nations deploy such weapons. How about an international ban on the deployment of all unmanned and automated weapons? The U.S. won't now even agree to ban land mines. At least this would restore SOME relationship between popular support and military might. Doesn't it sound ethical to insist that a human being decide when to end another human being's life? Doesn't it sound fair to require the decision maker to be in harm's way, especially when the person being killed is in or around their own home? Doesn't it sound unethical to add to the present situation? When deployed on a large scale, aren't these WMDs? Steve agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] Re: Shhh!
Jim, you are thinking out loud. There is no such thing as trans-infinite. How
about posting when you actually solve the problem.
-- Matt Mahoney, matmaho...@yahoo.com
From: Jim Bromer jimbro...@gmail.com
To: agi agi@v2.listbox.com
Sent: Mon, August 2, 2010 9:06:53 AM
Subject: [agi] Re: Shhh!
I think I can write an abbreviated version, but there would only be a few
people
in the world who would both believe me and understand why it would work.
On Mon, Aug 2, 2010 at 8:53 AM, Jim Bromer jimbro...@gmail.com wrote:
I can write an algorithm that is capable of describing ('reaching') every
possible irrational number - given infinite resources. The infinite is not a
number-like object, it is an active form of incrementation or concatenation.
So
I can write an algorithm that can write every finite state of every possible
number. However, it would take another algorithm to 'prove' it. Given an
irrational number, this other algorithm could find the infinite incrementation
for every digit of the given number. Each possible number (including
the incrementation of those numbers that cannot be represented in truncated
form) is embedded within a single infinite infinite incrementation of digits
that is produced by the algorithm, so the second algorithm would have to
calculate where you would find each digit of the given irrational number by
increment. But the thing is, both functions would be computable and provable.
(I haven't actually figured the second algorithm out yet, but it is not
a difficult problem.)
This means that the Trans-Infinite Is Computable. But don't tell anyone about
this, it's a secret.
agi | Archives | Modify Your Subscription
---
agi
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/
Modify Your Subscription:
https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c
Powered by Listbox: http://www.listbox.com
Re: [agi] AGI Alife
Ian Parker wrote: Matt Mahoney has costed his view of AGI. I say that costs must be recoverable as we go along. Matt, don't frighten people with a high estimate of cost. Frighten people instead with the bill they are paying now for dumb systems. It is not my intent to scare people out of building AGI, but rather to be realistic about its costs. Building machines that do what we want is a much harder problem than building intelligent machines. Machines surpassed human intelligence 50 years ago. But getting them to do useful work is still a $60 trillion per year problem. It's going to happen, but not as quickly as one might hope. -- Matt Mahoney, matmaho...@yahoo.com From: Ian Parker ianpark...@gmail.com To: agi agi@v2.listbox.com Sent: Wed, July 28, 2010 6:54:05 AM Subject: Re: [agi] AGI Alife On 27 July 2010 21:06, Jan Klauck jkla...@uni-osnabrueck.de wrote: Second observation about societal punishment eliminating free loaders. The fact of the matter is that *freeloading* is less of a problem in advanced societies than misplaced unselfishness. Fact of the matter, hm? Freeloading is an inherent problem in many social configurations. 9/11 brought down two towers, freeloading can bring down an entire country. There are very considerable knock on costs. There is the mushrooming cost of security This manifests itself in many ways. There is the cost of disruption to air travel. If someone rides on a plane without a ticket no one's life is put at risk. There are the military costs, it costs $1m per year to keep a soldier in Afghanistan. I don't know how much a Taliban fighter costs, but it must be a lot less. Clearly any reduction in these costs would be welcomed. If someone were to come along in the guise of social simulation and offer a reduction in these costs the research would pay for itself many times over. What you are interested in. This may be a somewhat unpopular thing to say, but money is important. Matt Mahoney has costed his view of AGI. I say that costs must be recoverable as we go along. Matt, don't frighten people with a high estimate of cost. Frighten people instead with the bill they are paying now for dumb systems. simulations seem :- 1) To be better done by Calculus. You usually use both, equations and heuristics. It depends on the problem, your resources, your questions, the people working with it a.s.o. That is the way things should be done. I agree absolutely. We could in fact take steepest descent (Calculus) and GAs and combine them together in a single composite program. This would in fact be quite a useful exercise. We would also eliminate genes that simply dealt with Calculus and steepest descent. I don't know whether it is useful to think in topological terms. - Ian Parker --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] Re: Huge Progress on the Core of AGI
David Jones wrote: I should also mention that I ran into problems mainly because I was having a hard time deciding how to identify objects and determine what is really going on in a scene. I think that your approach makes the problem harder than it needs to be (not that it is easy). Natural language processing is hard, so researchers in an attempt to break down the task into simpler parts, focused on steps like lexical analysis, parsing, part of speech resolution, and semantic analysis. While these problems went unsolved, Google went directly to a solution by skipping them. Likewise, parsing an image into physically separate objects and then building a 3-D model makes the problem harder, not easier. Again, look at the whole picture. You input an image and output a response. Let the system figure out which features are important. If your goal is to count basketball passes, then it is irrelevant whether the AGI recognizes that somebody is wearing a gorilla suit. -- Matt Mahoney, matmaho...@yahoo.com From: David Jones davidher...@gmail.com To: agi agi@v2.listbox.com Sent: Sat, July 24, 2010 2:25:49 PM Subject: Re: [agi] Re: Huge Progress on the Core of AGI Abram, I should also mention that I ran into problems mainly because I was having a hard time deciding how to identify objects and determine what is really going on in a scene. This adds a whole other layer of complexity to hypotheses. It's not just about what is more predictive of the observations, it is about deciding what exactly you are observing in the first place. (although you might say its the same problem). I ran into this problem when my algorithm finds matches between items that are not the same. Or it may not find any matches between items that are the same, but have changed. So, how do you decide whether it is 1) the same object, 2) a different object or 3) the same object but it has changed. And how do you decide its relationship to something else... is it 1) dependently attached 2) semi-dependently attached(can move independently, but only in certain ways. Yet also moves dependently) 3) independent 4) sometimes dependent 5) was dependent, but no longer is, 6) was dependent on something else, but then was independent, but now is dependent on something new. These hypotheses are different ways of explaining the same observations, but are complicated by the fact that we aren't sure of the identity of the objects we are observing in the first place. Multiple hypotheses may fit the same observations, and its hard to decide why one is simpler or better than the other. The object you were observing at first may have disappeared. A new object may have appeared at the same time (this is why screenshots are a bit malicious). Or the object you were observing may have changed. In screenshots, sometimes the objects that you are trying to identify as different never appear at the same time because they always completely occlude each other. So, that can make it extremely difficult to decide whether they are the same object that has changed or different objects. Such ambiguities are common in AGI. It is unclear to me yet how to deal with them effectively, although I am continuing to work hard on it. I know its a bit of a mess, but I'm just trying to demonstrate the trouble I've run into. I hope that makes it more clear why I'm having so much trouble finding a way of determining what hypothesis is most predictive and simplest. Dave On Thu, Jul 22, 2010 at 10:23 PM, Abram Demski abramdem...@gmail.com wrote: David, What are the different ways you are thinking of for measuring the predictiveness? I can think of a few different possibilities (such as measuring number incorrect vs measuring fraction incorrect, et cetera) but I'm wondering which variations you consider significant/troublesome/etc. --Abram On Thu, Jul 22, 2010 at 7:12 PM, David Jones davidher...@gmail.com wrote: It's certainly not as simple as you claim. First, assigning a probability is not always possible, nor is it easy. The factors in calculating that probability are unknown and are not the same for every instance. Since we do not know what combination of observations we will see, we cannot have a predefined set of probabilities, nor is it any easier to create a probability function that generates them for us. That is just as exactly what I meant by quantitatively define the predictiveness... it would be proportional to the probability. Second, if you can define a program ina way that is always simpler when it is smaller, then you can do the same thing without a program. I don't think it makes any sense to do it this way. It is not that simple. If it was, we could solve a large portion of agi easily. On Thu, Jul 22, 2010 at 3:16 PM, Matt Mahoney matmaho...@yahoo.com wrote: David Jones wrote: But, I am amazed at how difficult it is to quantitatively define more
Re: [agi] Comments On My Skepticism of Solomonoff Induction
Jim Bromer wrote: Solomonoff Induction may require a trans-infinite level of complexity just to run each program. Trans-infinite is not a mathematically defined term as far as I can tell. Maybe you mean larger than infinity, as in the infinite set of real numbers is larger than the infinite set of natural numbers (which is true). But it is not true that Solomonoff induction requires more than aleph-null operations. (Aleph-null is the size of the set of natural numbers, the smallest infinity). An exact calculation requires that you test aleph-null programs for aleph-null time steps each. There are aleph-null programs because each program is a finite length string, and there is a 1 to 1 correspondence between the set of finite strings and N, the set of natural numbers. Also, each program requires aleph-null computation in the case that it runs forever, because each step in the infinite computation can be numbered 1, 2, 3... However, the total amount of computation is still aleph-null because each step of each program can be described by an ordered pair (m,n) in N^2, meaning the n'th step of the m'th program, where m and n are natural numbers. The cardinality of N^2 is the same as the cardinality of N because there is a 1 to 1 correspondence between the sets. You can order the ordered pairs as (1,1), (1,2), (2,1), (1,3), (2,2), (3,1), (1,4), (2,3), (3,2), (4,1), (1,5), etc. See http://en.wikipedia.org/wiki/Countable_set#More_formal_introduction Furthermore you may approximate Solomonoff induction to any desired precision with finite computation. Simply interleave the execution of all programs as indicated in the ordering of ordered pairs that I just gave, where the programs are ordered from shortest to longest. Take the shortest program found so far that outputs your string, x. It is guaranteed that this algorithm will approach and eventually find the shortest program that outputs x given sufficient time, because this program exists and it halts. In case you are wondering how Solomonoff induction is not computable, the problem is that after this algorithm finds the true shortest program that outputs x, it will keep running forever and you might still be wondering if a shorter program is forthcoming. In general you won't know. -- Matt Mahoney, matmaho...@yahoo.com From: Jim Bromer jimbro...@gmail.com To: agi agi@v2.listbox.com Sent: Sat, July 24, 2010 3:59:18 PM Subject: Re: [agi] Comments On My Skepticism of Solomonoff Induction Solomonoff Induction may require a trans-infinite level of complexity just to run each program. Suppose each program is iterated through the enumeration of its instructions. Then, not only do the infinity of possible programs need to be run, many combinations of the infinite programs from each simulated Turing Machine also have to be tried. All the possible combinations of (accepted) programs, one from any two or more of the (accepted) programs produced by each simulated Turing Machine, have to be tried. Although these combinations of programs from each of the simulated Turing Machine may not all be unique, they all have to be tried. Since each simulated Turing Machine would produce infinite programs, I am pretty sure that this means that Solmonoff Induction is, by definition, trans-infinite. Jim Bromer On Thu, Jul 22, 2010 at 2:06 PM, Jim Bromer jimbro...@gmail.com wrote: I have to retract my claim that the programs of Solomonoff Induction would be trans-infinite. Each of the infinite individual programs could be enumerated by their individual instructions so some combination of unique individual programs would not correspond to a unique program but to the enumerated program that corresponds to the string of their individual instructions. So I got that one wrong. Jim Bromer agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] Re: Huge Progress on the Core of AGI
Mike Tintner wrote: Huh, Matt? What examples of this holistic scene analysis are there (or are you thinking about)? I mean a neural model with increasingly complex features, as opposed to an algorithmic 3-D model (like video game graphics in reverse). Of course David rejects such ideas ( http://practicalai.org/Prize/Default.aspx ) even though the one proven working vision model uses it. -- Matt Mahoney, matmaho...@yahoo.com From: Mike Tintner tint...@blueyonder.co.uk To: agi agi@v2.listbox.com Sent: Sat, July 24, 2010 6:16:07 PM Subject: Re: [agi] Re: Huge Progress on the Core of AGI Huh, Matt? What examples of this holistic scene analysis are there (or are you thinking about)? From: Matt Mahoney Sent: Saturday, July 24, 2010 10:25 PM To: agi Subject: Re: [agi] Re: Huge Progress on the Core of AGI David Jones wrote: I should also mention that I ran into problems mainly because I was having a hard time deciding how to identify objects and determine what is really going on in a scene. I think that your approach makes the problem harder than it needs to be (not that it is easy). Natural language processing is hard, so researchers in an attempt to break down the task into simpler parts, focused on steps like lexical analysis, parsing, part of speech resolution, and semantic analysis. While these problems went unsolved, Google went directly to a solution by skipping them. Likewise, parsing an image into physically separate objects and then building a 3-D model makes the problem harder, not easier. Again, look at the whole picture. You input an image and output a response. Let the system figure out which features are important. If your goal is to count basketball passes, then it is irrelevant whether the AGI recognizes that somebody is wearing a gorilla suit. agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] Re: Huge Progress on the Core of AGI
Mike Tintner wrote: Which is? The one right behind your eyes. -- Matt Mahoney, matmaho...@yahoo.com From: Mike Tintner tint...@blueyonder.co.uk To: agi agi@v2.listbox.com Sent: Sat, July 24, 2010 9:00:42 PM Subject: Re: [agi] Re: Huge Progress on the Core of AGI Matt: I mean a neural model with increasingly complex features, as opposed to an algorithmic 3-D model (like video game graphics in reverse). Of course David rejects such ideas ( http://practicalai.org/Prize/Default.aspx ) even though the one proven working vision model uses it. Which is? and does what? (I'm starting to consider that vision and visual perception - or perhaps one should say common sense, since no sense in humans works independent of the others - may well be considerably *more* complex than language. The evolutionary time required to develop our common sense perception and conception of the world was vastly greater than that required to develop language. And we are as a culture merely in our babbling infancy in beginning to understand how sensory images work and are processed). agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] Pretty worldchanging
The video says it has 2 GB of memory. I assume that's SSD and there is no disk. It's actually not hard to find a computer for $35. People are always throwing away old computers that still work. -- Matt Mahoney, matmaho...@yahoo.com From: Mike Tintner tint...@blueyonder.co.uk To: agi agi@v2.listbox.com Sent: Fri, July 23, 2010 9:50:44 AM Subject: [agi] Pretty worldchanging this strikes me as socially worldchanging if it works - potentially leading to you-ain't-see-nothing-yet changes in world education ( commerce) levels over the next decade: http://www.physorg.com/news199083092.html Any comments on its technical massproduction viability ? agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] Re: Huge Progress on the Core of AGI
David Jones wrote: But, I am amazed at how difficult it is to quantitatively define more predictive and simpler for specific problems. It isn't hard. To measure predictiveness, you assign a probability to each possible outcome. If the actual outcome has probability p, you score a penalty of log(1/p) bits. To measure simplicity, use the compressed size of the code for your prediction algorithm. Then add the two scores together. That's how it is done in the Calgary challenge http://www.mailcom.com/challenge/ and in my own text compression benchmark. -- Matt Mahoney, matmaho...@yahoo.com From: David Jones davidher...@gmail.com To: agi agi@v2.listbox.com Sent: Thu, July 22, 2010 3:11:46 PM Subject: Re: [agi] Re: Huge Progress on the Core of AGI Because simpler is not better if it is less predictive. On Thu, Jul 22, 2010 at 1:21 PM, Abram Demski abramdem...@gmail.com wrote: Jim, Why more predictive *and then* simpler? --Abram On Thu, Jul 22, 2010 at 11:49 AM, David Jones davidher...@gmail.com wrote: An Update I think the following gets to the heart of general AI and what it takes to achieve it. It also provides us with evidence as to why general AI is so difficult. With this new knowledge in mind, I think I will be much more capable now of solving the problems and making it work. I've come to the conclusion lately that the best hypothesis is better because it is more predictive and then simpler than other hypotheses (in that order more predictive... then simpler). But, I am amazed at how difficult it is to quantitatively define more predictive and simpler for specific problems. This is why I have sometimes doubted the truth of the statement. In addition, the observations that the AI gets are not representative of all observations! This means that if your measure of predictiveness depends on the number of certain observations, it could make mistakes! So, the specific observations you are aware of may be unrepresentative of the predictiveness of a hypothesis relative to the truth. If you try to calculate which hypothesis is more predictive and you don't have the critical observations that would give you the right answer, you may get the wrong answer! This all depends of course on your method of calculation, which is quite elusive to define. Visual input from screenshots, for example, can be somewhat malicious. Things can move, appear, disappear or occlude each other suddenly. So, without sufficient knowledge it is hard to decide whether matches you find between such large changes are because it is the same object or a different object. This may indicate that bias and preprogrammed experience should be introduced to the AI before training. Either that or the training inputs should be carefully chosen to avoid malicious input and to make them nice for learning. This is the correspondence problem that is typical of computer vision and has never been properly solved. Such malicious input also makes it difficult to learn automatically because the AI doesn't have sufficient experience to know which changes or transformations are acceptable and which are not. It is immediately bombarded with malicious inputs. I've also realized that if a hypothesis is more explanatory, it may be better. But quantitatively defining explanatory is also elusive and truly depends on the specific problems you are applying it to because it is a heuristic. It is not a true measure of correctness. It is not loyal to the truth. More explanatory is really a heuristic that helps us find hypothesis that are more predictive. The true measure of whether a hypothesis is better is simply the most accurate and predictive hypothesis. That is the ultimate and true measure of correctness. Also, since we can't measure every possible prediction or every last prediction (and we certainly can't predict everything), our measure of predictiveness can't possibly be right all the time! We have no choice but to use a heuristic of some kind. So, its clear to me that the right hypothesis is more predictive and then simpler. But, it is also clear that there will never be a single measure of this that can be applied to all problems. I hope to eventually find a nice model for how to apply it to different problems though. This may be the reason that so many people have tried and failed to develop general AI. Yes, there is a solution. But there is no silver bullet that can be applied to all problems. Some methods are better than others. But I think another major reason of the failures is that people think they can predict things without sufficient information. By approaching the problem this way, we compound the need for heuristics and the errors they produce because we simply don't have sufficient information to make a good decision with limited evidence. If approached correctly, the right solution would solve many more
Re: [agi] How do we hear music
deepakjnath wrote: Why do we listen to a song sung in different scale and yet identify it as the same song.? Does it have something to do with the fundamental way in which we store memory? For the same reason that gray looks green on a red background. You have more neurons that respond to differences in tones than to absolute frequencies. -- Matt Mahoney, matmaho...@yahoo.com From: deepakjnath deepakjn...@gmail.com To: agi agi@v2.listbox.com Sent: Thu, July 22, 2010 3:59:57 PM Subject: [agi] How do we hear music Why do we listen to a song sung in different scale and yet identify it as the same song.? Does it have something to do with the fundamental way in which we store memory? cheers, Deepak agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] Comments On My Skepticism of Solomonoff Induction
Jim Bromer wrote: Please give me a little more explanation why you say the fundamental method is that the probability of a string x is proportional to the sum of all programs M that output x weighted by 2^-|M|. Why is the M in a bracket? By |M| I mean the length of the program M in bits. Why 2^-|M|? Because each bit means you can have twice as many programs, so they should count half as much. Being uncomputable doesn't make it wrong. The fact that there is no general procedure for finding the shortest program that outputs a string doesn't mean that you can never find it, or that for many cases you can't approximate it. You apply Solomonoff induction all the time. What is the next bit in these sequences? 1. 0101010101010101010101010101010 2. 11001001110110101010001 In sequence 1 there is an obvious pattern with a short description. You can find a short program that outputs 0 and 1 alternately forever, so you predict the next bit will be 1. It might not be the shortest program, but it is enough that alternate 0 and 1 forever is shorter than alternate 0 and 1 15 times followed by 00 that you can confidently predict the first hypothesis is more likely. The second sequence is not so obvious. It looks like random bits. With enough intelligence (or computation) you might discover that the sequence is a binary representation of pi, and therefore the next bit is 0. But the fact that you might not discover the shortest description does not invalidate the principle. It just says that you can't always apply Solomonoff induction and get the number you want. Perhaps http://en.wikipedia.org/wiki/Kolmogorov_complexity will make this clear. -- Matt Mahoney, matmaho...@yahoo.com From: Jim Bromer jimbro...@gmail.com To: agi agi@v2.listbox.com Sent: Thu, July 22, 2010 5:06:12 PM Subject: Re: [agi] Comments On My Skepticism of Solomonoff Induction On Wed, Jul 21, 2010 at 8:47 PM, Matt Mahoney matmaho...@yahoo.com wrote: The fundamental method is that the probability of a string x is proportional to the sum of all programs M that output x weighted by 2^-|M|. That probability is dominated by the shortest program, but it is equally uncomputable either way. Also, please point me to this mathematical community that you claim rejects Solomonoff induction. Can you find even one paper that refutes it? You give a precise statement of the probability in general terms, but then say that it is uncomputable. Then you ask if there is a paper that refutes it. Well, why would any serious mathematician bother to refute it since you yourself acknowledge that it is uncomputable and therefore unverifiable and therefore not a mathematical theorem that can be proven true or false? It isn't like you claimed that the mathematical statement is verifiable. It is as if you are making a statement and then ducking any responsibility for it by denying that it is even an evaluation. You honestly don't see the irregularity? My point is that the general mathematical community doesn't accept Solomonoff Induction, not that I have a paper that refutes it, whatever that would mean. Please give me a little more explanation why you say the fundamental method is that the probability of a string x is proportional to the sum of all programs M that output x weighted by 2^-|M|. Why is the M in a bracket? On Wed, Jul 21, 2010 at 8:47 PM, Matt Mahoney matmaho...@yahoo.com wrote: Jim Bromer wrote: The fundamental method of Solmonoff Induction is trans-infinite. The fundamental method is that the probability of a string x is proportional to the sum of all programs M that output x weighted by 2^-|M|. That probability is dominated by the shortest program, but it is equally uncomputable either way. How does this approximation invalidate Solomonoff induction? Also, please point me to this mathematical community that you claim rejects Solomonoff induction. Can you find even one paper that refutes it? -- Matt Mahoney, matmaho...@yahoo.com From: Jim Bromer jimbro...@gmail.com To: agi agi@v2.listbox.com Sent: Wed, July 21, 2010 3:08:13 PM Subject: Re: [agi] Comments On My Skepticism of Solomonoff Induction I should have said, It would be unwise to claim that this method could stand as an ideal for some valid and feasible application of probability. Jim Bromer On Wed, Jul 21, 2010 at 2:47 PM, Jim Bromer jimbro...@gmail.com wrote: The fundamental method of Solmonoff Induction is trans-infinite. Suppose you iterate through all possible programs, combining different programs as you go. Then you have an infinite number of possible programs which have a trans-infinite number of combinations, because each tier of combinations can then be recombined to produce a second, third, fourth,... tier of recombinations. Anyone who claims that this method is the ideal for a method of applied probability is unwise
Re: [agi] The Collective Brain
Mike Tintner wrote: The fantasy of a superAGI machine that can grow individually without a vast society supporting it, is another one of the wild fantasies of AGI-ers and Singularitarians that violate truly basic laws of nature. Individual brains cannot flourish individually in the real world, only societies of brains (and bodies) can. I agree. It is the basis of my AGI design, to supplement a global brain with computers. http://mattmahoney.net/agi2.html -- Matt Mahoney, matmaho...@yahoo.com From: Mike Tintner tint...@blueyonder.co.uk To: agi agi@v2.listbox.com Sent: Tue, July 20, 2010 1:50:45 PM Subject: [agi] The Collective Brain http://www.ted.com/talks/matt_ridley_when_ideas_have_sex.html?utm_source=newsletter_weekly_2010-07-20utm_campaign=newsletter_weeklyutm_medium=email Good lecture worth looking at about how trade - exchange of both goods and ideas - has fostered civilisation. Near the end introduces a v. important idea - the collective brain. In other words, our apparently individual intelligence is actually a collective intelligence. Nobody he points out actually knows how to make a computer mouse, although that may seem counterintuitive - it's an immensely complex piece of equipment, simple as it may appear, that engages the collective, interdependent intelligence and productive efforts of vast numbers of people. When you start thinking like that, you realise that there is v. little we know how to do, esp of an intellectual nature, individually, without the implicit and explicit collaboration of vast numbers of people and sectors of society. The fantasy of a superAGI machine that can grow individually without a vast society supporting it, is another one of the wild fantasies of AGI-ers and Singularitarians that violate truly basic laws of nature. Individual brains cannot flourish individually in the real world, only societies of brains (and bodies) can. (And of course computers can do absolutely nothing or in any way survive without their human masters - even if it may appear that way, if you don't look properly at their whole operation) agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] Comments On My Skepticism of Solomonoff Induction
Jim Bromer wrote: The question was asked whether, given infinite resources could Solmonoff Induction work. I made the assumption that it was computable and found that it wouldn't work. On what infinitely powerful computer did you do your experiment? My conclusion suggests, that the use of Solmonoff Induction as an ideal for compression or something like MDL is not only unsubstantiated but based on a massive inability to comprehend the idea of a program that runs every possible program. It is sufficient to find the shortest program consistent with past results, not all programs. The difference is no more than the language-dependent constant. Legg proved this in the paper that Ben and I both pointed you to. Do you dispute his proof? I guess you don't, because you didn't respond the last 3 times this was pointed out to you. I am comfortable with the conclusion that the claim that Solomonoff Induction is an ideal for compression or induction or anything else is pretty shallow and not based on careful consideration. I am comfortable with the conclusion that the world is flat because I have a gut feeling about it and I ignore overwhelming evidence to the contrary. There is a chance that I am wrong So why don't you drop it? -- Matt Mahoney, matmaho...@yahoo.com From: Jim Bromer jimbro...@gmail.com To: agi agi@v2.listbox.com Sent: Tue, July 20, 2010 3:10:40 PM Subject: Re: [agi] Comments On My Skepticism of Solomonoff Induction The question was asked whether, given infinite resources could Solmonoff Induction work. I made the assumption that it was computable and found that it wouldn't work. It is not computable, even with infinite resources, for the kind of thing that was claimed it would do. (I believe that with a governance program it might actually be programmable) but it could not be used to predict (or compute the probability of) a subsequent string given some prefix string. Not only is the method impractical it is theoretically inane. My conclusion suggests, that the use of Solmonoff Induction as an ideal for compression or something like MDL is not only unsubstantiated but based on a massive inability to comprehend the idea of a program that runs every possible program. I am comfortable with the conclusion that the claim that Solomonoff Induction is an ideal for compression or induction or anything else is pretty shallow and not based on careful consideration. There is a chance that I am wrong, but I am confident that there is nothing in the definition of Solmonoff Induction that could be used to prove it. Jim Bromer agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] Comments On My Skepticism of Solomonoff Induction
Jim Bromer wrote: The fundamental method of Solmonoff Induction is trans-infinite. The fundamental method is that the probability of a string x is proportional to the sum of all programs M that output x weighted by 2^-|M|. That probability is dominated by the shortest program, but it is equally uncomputable either way. How does this approximation invalidate Solomonoff induction? Also, please point me to this mathematical community that you claim rejects Solomonoff induction. Can you find even one paper that refutes it? -- Matt Mahoney, matmaho...@yahoo.com From: Jim Bromer jimbro...@gmail.com To: agi agi@v2.listbox.com Sent: Wed, July 21, 2010 3:08:13 PM Subject: Re: [agi] Comments On My Skepticism of Solomonoff Induction I should have said, It would be unwise to claim that this method could stand as an ideal for some valid and feasible application of probability. Jim Bromer On Wed, Jul 21, 2010 at 2:47 PM, Jim Bromer jimbro...@gmail.com wrote: The fundamental method of Solmonoff Induction is trans-infinite. Suppose you iterate through all possible programs, combining different programs as you go. Then you have an infinite number of possible programs which have a trans-infinite number of combinations, because each tier of combinations can then be recombined to produce a second, third, fourth,... tier of recombinations. Anyone who claims that this method is the ideal for a method of applied probability is unwise. Jim Bromer agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] Of definitions and tests of AGI
Mike, I think we all agree that we should not have to tell an AGI the steps to solving problems. It should learn and figure it out, like the way that people figure it out. The question is how to do that. We know that it is possible. For example, I could write a chess program that I could not win against. I could write the program in such a way that it learns to improve its game by playing against itself or other opponents. I could write it in such a way that initially does not know the rules for chess, but instead learns the rules by being given examples of legal and illegal moves. What we have not yet been able to do is scale this type of learning and problem solving up to general, human level intelligence. I believe it is possible, but it will require lots of training data and lots of computing power. It is not something you could do on a PC, and it won't be cheap. -- Matt Mahoney, matmaho...@yahoo.com From: Mike Tintner tint...@blueyonder.co.uk To: agi agi@v2.listbox.com Sent: Mon, July 19, 2010 9:07:53 PM Subject: Re: [agi] Of definitions and tests of AGI The issue isn't what a computer can do. The issue is how you structure the computer's or any agent's thinking about a problem. Programs/Turing machines are only one way of structuring thinking/problemsolving - by, among other things, giving the computer a method/process of solution. There is an alternative way of structuring a computer's thinking, which incl., among other things, not giving it a method/ process of solution, but making it rather than a human programmer do the real problemsolving. More of that another time. From: Matt Mahoney Sent: Tuesday, July 20, 2010 1:38 AM To: agi Subject: Re: [agi] Of definitions and tests of AGI Creativity is the good feeling you get when you discover a clever solution to a hard problem without knowing the process you used to discover it. I think a computer could do that. -- Matt Mahoney, matmaho...@yahoo.com From: Mike Tintner tint...@blueyonder.co.uk To: agi agi@v2.listbox.com Sent: Mon, July 19, 2010 2:08:28 PM Subject: Re: [agi] Of definitions and tests of AGI Yes that's what people do, but it's not what programmed computers do. The useful formulation that emerges here is: narrow AI (and in fact all rational) problems have *a method of solution* (to be equated with general method) - and are programmable (a program is a method of solution) AGI (and in fact all creative) problems do NOT have *a method of solution* (in the general sense) - rather a one.off *way of solving the problem* has to be improvised each time. AGI/creative problems do not in fact have a method of solution, period. There is no (general) method of solving either the toy box or the build-a-rock-wall problem - one essential feature which makes them AGI. You can learn, as you indicate, from *parts* of any given AGI/creative solution, and apply the lessons to future problems - and indeed with practice, should improve at solving any given kind of AGI/creative problem. But you can never apply a *whole* solution/way to further problems. P.S. One should add that in terms of computers, we are talking here of *complete, step-by-step* methods of solution. From: rob levy Sent: Monday, July 19, 2010 5:09 PM To: agi Subject: Re: [agi] Of definitions and tests of AGI And are you happy with: AGI is about devising *one-off* methods ofproblemsolving (that only apply to the individual problem, and cannot bere-used - at least not in their totality) Yes exactly, isn't that what people do? Also, I think that being able to recognize where past solutions can be generalized and where past solutions can be varied and reused is a detail of how intelligence works that is likely to be universal. vs narrow AI is about applying pre-existing*general* methods of problemsolving (applicable to whole classes ofproblems)? From: rob levy Sent: Monday, July 19, 2010 4:45 PM To: agi Subject: Re: [agi] Of definitions and tests ofAGI Well, solving ANY problem is a little too strong. This isAGI, not AGH (artificial godhead), though AGH could be an unintendedconsequence ;). So I would rephrase solving any problem as being ableto come up with reasonable approaches and strategies to any problem (just ashumans are able to do). On Mon, Jul 19, 2010 at 11:32 AM, Mike Tintner tint...@blueyonder.co.uk wrote: Whaddya mean by solve the problem of how to solve problems? Develop a universal approach to solving any problem? Or find a method of solving a class of problems? Or what? From: rob levy Sent: Monday, July 19, 2010 1:26 PM To: agi Subject: Re: [agi] Of definitions and tests of AGI However, I see that there are no validdefinitions of AGI that explain what AGI is generally , and why
Re: [agi] Of definitions and tests of AGI
Creativity is the good feeling you get when you discover a clever solution to a hard problem without knowing the process you used to discover it. I think a computer could do that. -- Matt Mahoney, matmaho...@yahoo.com From: Mike Tintner tint...@blueyonder.co.uk To: agi agi@v2.listbox.com Sent: Mon, July 19, 2010 2:08:28 PM Subject: Re: [agi] Of definitions and tests of AGI Yes that's what people do, but it's not what programmed computers do. The useful formulation that emerges here is: narrow AI (and in fact all rational) problems have *a method of solution* (to be equated with general method) - and are programmable (a program is a method of solution) AGI (and in fact all creative) problems do NOT have *a method of solution* (in the general sense) - rather a one.off *way of solving the problem* has to be improvised each time. AGI/creative problems do not in fact have a method of solution, period. There is no (general) method of solving either the toy box or the build-a-rock-wall problem - one essential feature which makes them AGI. You can learn, as you indicate, from *parts* of any given AGI/creative solution, and apply the lessons to future problems - and indeed with practice, should improve at solving any given kind of AGI/creative problem. But you can never apply a *whole* solution/way to further problems. P.S. One should add that in terms of computers, we are talking here of *complete, step-by-step* methods of solution. From: rob levy Sent: Monday, July 19, 2010 5:09 PM To: agi Subject: Re: [agi] Of definitions and tests of AGI And are you happy with: AGI is about devising *one-off* methods ofproblemsolving (that only apply to the individual problem, and cannot bere-used - at least not in their totality) Yes exactly, isn't that what people do? Also, I think that being able to recognize where past solutions can be generalized and where past solutions can be varied and reused is a detail of how intelligence works that is likely to be universal. vs narrow AI is about applying pre-existing*general* methods of problemsolving (applicable to whole classes ofproblems)? From: rob levy Sent: Monday, July 19, 2010 4:45 PM To: agi Subject: Re: [agi] Of definitions and tests ofAGI Well, solving ANY problem is a little too strong. This isAGI, not AGH (artificial godhead), though AGH could be an unintendedconsequence ;). So I would rephrase solving any problem as being ableto come up with reasonable approaches and strategies to any problem (just ashumans are able to do). On Mon, Jul 19, 2010 at 11:32 AM, Mike Tintner tint...@blueyonder.co.uk wrote: Whaddya mean by solve the problem of how to solve problems? Develop a universal approach to solving any problem? Or find a method of solving a class of problems? Or what? From: rob levy Sent: Monday, July 19, 2010 1:26 PM To: agi Subject: Re: [agi] Of definitions and tests of AGI However, I see that there are no validdefinitions of AGI that explain what AGI is generally , and why thesetests are indeed AGI. Google - there are v. few defs. of AGI or Strong AI,period. I like Fogel's idea that intelligence is the ability to solve the problem of how to solve problems in new and changing environments. I don't think Fogel's method accomplishes this, but the goal he expresses seems to be the goal of AGI as I understand it. Rob agi | Archives | Modify Your Subscription agi | Archives | Modify Your Subscription agi | Archives | Modify Your Subscription agi | Archives | Modify Your Subscription agi | Archives | Modify Your Subscription agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] Of definitions and tests of AGI
http://www.loebner.net/Prizef/loebner-prize.html -- Matt Mahoney, matmaho...@yahoo.com From: David Jones davidher...@gmail.com To: agi agi@v2.listbox.com Sent: Sun, July 18, 2010 3:10:12 PM Subject: Re: [agi] Of definitions and tests of AGI If you can't convince someone, clearly something is wrong with it. I don't think a test is the right way to do this. Which is why I haven't commented much. When you understand how to create AGI, it will be obvious that it is AGI or that it is what you intend it to be. You'll then understand how what you have built fits into the bigger scheme of things. There is no such point at which you can say something is AGI and not AGI. Intelligence is a very subjective thing that really depends on your goals. Someone will always say it is not good enough. But if it really works, people will quickly realize it based on results. What you want is to develop a system that can learn about the world or its environment in a general way so that it can solve arbitrary problems, be able to plan in general ways, act in general ways and perform the types of goals you want it to perform. Dave On Sun, Jul 18, 2010 at 3:03 PM, deepakjnath deepakjn...@gmail.com wrote: So if I have a system that is close to AGI, I have no way of really knowing it right? Even if I believe that my system is a true AGI there is no way of convincing the others irrefutably that this system is indeed a AGI not just an advanced AI system. I have read the toy box problem and rock wall problem, but not many people will still be convinced I am sure. I wanted to know that if there is any consensus on a general problem which can be solved and only solved by a true AGI. Without such a test bench how will we know if we are moving closer or away from our quest. There is no map. Deepak On Sun, Jul 18, 2010 at 11:50 PM, Mike Tintner tint...@blueyonder.co.uk wrote: I realised that what is needed is a *joint* definition *and* range of tests of AGI. Benamin Johnston has submitted one valid test - the toy box problem. (See archives). I have submitted another still simpler valid test - build a rock wall from rocks given, (or fill an earth hole with rocks). However, I see that there are no valid definitions of AGI that explain what AGI is generally , and why these tests are indeed AGI. Google - there are v. few defs. of AGI or Strong AI, period. The most common: AGI is human-level intelligence - is an embarrassing non-starter - what distinguishes human intelligence? No explanation offered. The other two are also inadequate if not as bad: Ben's solves a variety of complex problems in a variety of complex environments. Nope, so does a multitasking narrow AI. Complexity does not distinguish AGI. Ditto Pei's - something to do with insufficient knowledge and resources... Insufficient is open to narrow AI interpretations and reducible to mathematically calculable probabilities.or uncertainties. That doesn't distinguish AGI from narrow AI. The one thing we should all be able to agree on (but who can be sure?) is that: ** an AGI is a general intelligence system, capable of independent learning** i.e. capable of independently learning new activities/skills with minimal guidance or even, ideally, with zero guidance (as humans and animals are) - and thus acquiring a general, all-round range of intelligence.. This is an essential AGI goal - the capacity to keep entering and mastering new domains of both mental and physical skills WITHOUT being specially programmed each time - that crucially distinguishes it from narrow AI's, which have to be individually programmed anew for each new task. Ben's AGI dog exemplified this in a v simple way - the dog is supposed to be able to learn to fetch a ball, with only minimal instructions, as real dogs do - they can learn a whole variety of new skills with minimal instruction. But I am confident Ben's dog can't actually do this. However, the independent learning def. while focussing on the distinctive AGI goal, still is not detailed enough by itself. It requires further identification of the **cognitive operations** which distinguish AGI, and wh. are exemplified by the above tests. [I'll stop there for interruptions/comments continue another time]. P.S. Deepakjnath, It is vital to realise that the overwhelming majority of AGI-ers do not * want* an AGI test - Ben has never gone near one, and is merely typical in this respect. I'd put almost all AGI-ers here in the same league as the US banks, who only want mark-to-fantasy rather than mark-to-market tests of their assets. agi | Archives | Modify Your Subscription -- cheers, Deepak agi | Archives | Modify Your Subscription agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member
Re: [agi] Comments On My Skepticism of Solomonoff Induction
Jim Bromer wrote: The definition of all possible programs, like the definition of all possible mathematical functions, is not a proper mathematical problem that can be comprehended in an analytical way. Finding just the shortest program is close enough because it dominates the probability. Or which step in the proof of theorem 1.7.2 in http://www.vetta.org/documents/disSol.pdf do you disagree with? You have been saying that you think Solomonoff induction is wrong, but offering no argument except your own intuition. So why should we care? -- Matt Mahoney, matmaho...@yahoo.com From: Jim Bromer jimbro...@gmail.com To: agi agi@v2.listbox.com Sent: Sun, July 18, 2010 9:09:36 PM Subject: Re: [agi] Comments On My Skepticism of Solomonoff Induction Abram, I was going to drop the discussion, but then I thought I figured out why you kept trying to paper over the difference. Of course, our personal disagreement is trivial; it isn't that important. But the problem with Solomonoff Induction is that not only is the output hopelessly tangled and seriously infinite, but the input is as well. The definition of all possible programs, like the definition of all possible mathematical functions, is not a proper mathematical problem that can be comprehended in an analytical way. I think that is the part you haven't totally figured out yet (if you will excuse the pun). Total program space, does not represent a comprehensible computational concept. When you try find a way to work out feasible computable examples it is not enough to limit the output string space, you HAVE to limit the program space in the same way. That second limitation makes the entire concept of total program space, much too weak for our purposes. You seem to know this at an intuitive operational level, but it seems to me that you haven't truly grasped the implications. I say that Solomonoff Induction is computational but I have to use a trick to justify that remark. I think the trick may be acceptable, but I am not sure. But the possibility that the concept of all possible programs, might be computational doesn't mean that that it is a sound mathematical concept. This underlies the reason that I intuitively came to the conclusion that Solomonoff Induction was transfinite. However, I wasn't able to prove it because the hypothetical concept of all possible program space, is so pretentious that it does not lend itself to mathematical analysis. I just wanted to point this detail out because your implied view that you agreed with me but total program space was mathematically well-defined did not make any sense. Jim Bromer agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] NL parsing
That that that Buffalo buffalo that Buffalo buffalo buffalo buffalo that Buffalo buffalo that Buffalo buffalo buffalo. -- Matt Mahoney, matmaho...@yahoo.com - Original Message From: Mike Tintner tint...@blueyonder.co.uk To: agi agi@v2.listbox.com Sent: Fri, July 16, 2010 11:05:51 AM Subject: Re: [agi] NL parsing Or if you want to be pedantic about caps, the speaker is identifying 3 buffaloes from Buffalo, 2 from elsewhere. Anyone got any other readings? -- From: Jiri Jelinek jjelinek...@gmail.com Sent: Friday, July 16, 2010 3:12 PM To: agi agi@v2.listbox.com Subject: [agi] NL parsing Believe it or not, this sentence is grammatically correct and has meaning: 'Buffalo buffalo Buffalo buffalo buffalo buffalo Buffalo buffalo.' source: http://www.mentalfloss.com/blogs/archives/13120 :-) --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] Comments On My Skepticism of Solomonoff Induction
Jim Bromer wrote: Since you cannot fully compute every string that may be produced at a certain iteration, you cannot make the claim that you even know the probabilities of any possible string before infinity and therefore your claim that the sum of the probabilities can be computed is not provable. But I could be wrong. Could be. Theorem 1.7.2 in http://www.vetta.org/documents/disSol.pdf proves that finding just the shortest program that outputs x gives you a probability for x close to the result you would get if you found all of the (infinite number of) programs that output x. Either number could be used for Solomonoff induction because the difference is bounded only by the choice of language. -- Matt Mahoney, matmaho...@yahoo.com From: Jim Bromer jimbro...@gmail.com To: agi agi@v2.listbox.com Sent: Thu, July 15, 2010 8:18:13 AM Subject: Re: [agi] Comments On My Skepticism of Solomonoff Induction On Wed, Jul 14, 2010 at 7:46 PM, Abram Demski abramdem...@gmail.com wrote: Jim, There is a simple proof of convergence for the sum involved in defining the probability of a given string in the Solomonoff distribution: At its greatest, a particular string would be output by *all* programs. In this case, its sum would come to 1. This puts an upper bound on the sum. Since there is no subtraction, there is a lower bound at 0 and the sum monotonically increases as we take the limit. Knowing these facts, suppose it *didn't* converge. It must then increase without bound, since it cannot fluctuate back and forth (it can only go up). But this contradicts the upper bound of 1. So, the sum must stop at 1 or below (and in fact we can prove it stops below 1, though we can't say where precisely without the infinite computing power required to compute the limit). --Abram I believe that Solomonoff Induction would be computable given infinite time and infinite resources (the Godel Theorem fits into this category) but some people disagree for reasons I do not understand. If it is not computable then it is not a mathematical theorem and the question of whether the sum of probabilities equals 1 is pure fantasy. If it is computable then the central issue is whether it could (given infinite time and infinite resources) be used to determine the probability of a particular string being produced from all possible programs. The question about the sum of all the probabilities is certainly an interesting question. However, the problem of making sure that the function was actually computable would interfere with this process of determining the probability of each particular string that can be produced. For example, since some strings would be infinite, the computability problem makes it imperative that the infinite strings be partially computed at an iteration (or else the function would be hung up at some particular iteration and the infinite other calculations could not be considered computable). My criticism is that even though I believe the function may be theoretically computable, the fact that each particular probability (of each particular string that is produced) cannot be proven to approach a limit through mathematical analysis, and since the individual probabilities will fluctuate with each new string that is produced, one would have to know how to reorder the production of the probabilities in order to demonstrate that the individual probabilities do approach a limit. If they don't, then the claim that this function could be used to define the probabilities of a particular string from all possible program is unprovable. (Some infinite calculations fluctuate infinitely.) Since you do not have any way to determine how to reorder the infinite probabilities a priori, your algorithm would have to be able to compute all possible reorderings to find the ordering and filtering of the computations that would produce evaluable limits. Since there are trans infinite rearrangements of an infinite list (I am not sure that I am using the term 'trans infinite' properly) this shows that the conclusion that the theorem can be used to derive the desired probabilities is unprovable through a variation of Cantor's Diagonal Argument, and that you can't use Solomonoff Induction the way you have been talking about using it. Since you cannot fully compute every string that may be produced at a certain iteration, you cannot make the claim that you even know the probabilities of any possible string before infinity and therefore your claim that the sum of the probabilities can be computed is not provable. But I could be wrong. Jim Bromer agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id
Re: [agi] How do we Score Hypotheses?
Hypotheses are scored using Bayes law. Let D be your observed data and H be your hypothesis. Then p(H|D) = p(D|H)p(H)/p(D). Since p(D) is constant, you can remove it and rank hypotheses by p(D|H)p(H). p(H) can be estimated using the minimum description length principle or Solomonoff induction. Ideally, p(H) = 2^-|H| where |H| is the length (in bits) of the description of the hypothesis. The value is language dependent, so this method is not perfect. -- Matt Mahoney, matmaho...@yahoo.com From: David Jones davidher...@gmail.com To: agi agi@v2.listbox.com Sent: Thu, July 15, 2010 10:22:44 AM Subject: Re: [agi] How do we Score Hypotheses? It is no wonder that I'm having a hard time finding documentation on hypothesis scoring. Few can agree on how to do it and there is much debate about it. I noticed though that a big reason for the problems is that explanatory reasoning is being applied to many diverse problems. I think, like I mentioned before, that people should not try to come up with a single universal rule set for applying explanatory reasoning to every possible problem. So, maybe that's where the hold up is. I've been testing my ideas out on complex examples. But now I'm going to go back to simplified model testing (although not as simple as black squares :) ) and work my way up again. Dave On Wed, Jul 14, 2010 at 12:59 PM, David Jones davidher...@gmail.com wrote: Actually, I just realized that there is a way to included inductive knowledge and experience into this algorithm. Inductive knowledge and experience about a specific object or object type can be exploited to know which hypotheses in the past were successful, and therefore which hypothesis is most likely. By choosing the most likely hypothesis first, we skip a lot of messy hypothesis comparison processing and analysis. If we choose the right hypothesis first, all we really have to do is verify that this hypothesis reveals in the data what we expect to be there. If we confirm what we expect, that is reason enough not to look for other hypotheses because the data is explained by what we originally believed to be likely. We only look for additional hypotheses when we find something unexplained. And even then, we don't look at the whole problem. We only look at what we have to to explain the unexplained data. In fact, we could even ignore the unexplained data if we believe, from experience, that it isn't pertinent. I discovered this because I'm analyzing how a series of hypotheses are navigated when analyzing images. It seems to me that it is done very similarly to way we do it. We sort of confirm what we expect and try to explain what we don't expect. We try out hypotheses in a sort of trial and error manor and see how each hypothesis affects what we find in the image. If we confirm things because of the hypothesis, we are likely to keep it. We keep going, navigating the tree of hypotheses, conflicts and unexpected observations until we find a good hypothesis. Something like that. I'm attempting to construct an algorithm for doing this as I analyze specific problems. Dave On Wed, Jul 14, 2010 at 10:22 AM, David Jones davidher...@gmail.com wrote: What do you mean by definitive events? I guess the first problem I see with my approach is that the movement of the window is also a hypothesis. I need to analyze it in more detail and see how the tree of hypotheses affects the hypotheses regarding the es on the windows. What I believe is that these problems can be broken down into types of hypotheses, types of events and types of relationships. then those types can be reasoned about in a general way. If possible, then you have a method for reasoning about any object that is covered by the types of hypotheses, events and relationships that you have defined. How to reason about specific objects should not be preprogrammed. But, I think the solution to this part of AGI is to find general ways to reason about a small set of concepts that can be combined to describe specific objects and situations. There are other parts to AGI that I am not considering yet. I believe the problem has to be broken down into separate pieces and understood before putting it back together into a complete system. I have not covered inductive learning for example, which would be an important part of AGI. I have also not yet incorporated learned experience into the algorithm, which is also important. The general AI problem is way too complicated to consider all at once. I simply can't solve hypothesis generation, comparison and disambiguation while at the same time solving induction and experience-based reasoning. It becomes unwieldly. So, I'm starting where I can and I'll work my way up to the full complexity of the problem. I don't really understand what you mean here: The central unsolved problem, in my view, is: How can hypotheses
Re: [agi] What is the smallest set of operations that can potentially define everything and how do you combine them ?
Actually, Fibonacci numbers can be computed without loops or recursion.
int fib(int x) {
return round(pow((1+sqrt(5))/2, x)/sqrt(5));
}
unless you argue that loops are needed to compute sqrt() and pow().
The brain and DNA use redundancy and parallelism and don't use loops because
their operations are slow and unreliable. This is not necessarily the best
strategy for computers because computers are fast and reliable but don't have a
lot of parallelism.
-- Matt Mahoney, matmaho...@yahoo.com
- Original Message
From: Michael Swan ms...@voyagergaming.com
To: agi agi@v2.listbox.com
Sent: Wed, July 14, 2010 12:18:40 AM
Subject: Re: [agi] What is the smallest set of operations that can potentially
define everything and how do you combine them ?
Brain loops:
Premise:
Biological brain code does not contain looping constructs, or the
ability to creating looping code, (due to the fact they are extremely
dangerous on unreliable hardware) except for 1 global loop that fires
about 200 times a second.
Hypothesis:
Brains cannot calculate iterative problems quickly, where calculations
in the previous iteration are needed for the next iteration and, where
brute force operations are the only valid option.
Proof:
Take as an example, Fibonacci numbers
http://en.wikipedia.org/wiki/Fibonacci_number
What are the first 100 Fibonacci numbers?
int Fibonacci[102];
Fibonacci[0] = 0;
Fibonacci[1] = 1;
for(int i = 0; i 100; i++)
{
// Getting the next Fibonacci number relies on the previous values
Fibonacci[i+2] = Fibonacci[i] + Fibonacci[i+1];
}
My brain knows the process to solve this problem but it can't directly
write a looping construct into itself. And so it solves it very slowly
compared to a computer.
The brain probably consists of vast repeating look-up tables. Of course,
run in parallel these seem fast.
DNA has vast tracks of repeating data. Why would DNA contain repeating
data, instead of just having the data once and the number of times it's
repeated like in a loop? One explanation is that DNA can't do looping
construct either.
On Wed, 2010-07-14 at 02:43 +0100, Mike Tintner wrote:
Michael: We can't do operations that
require 1,000,000 loop iterations. I wish someone would give me a PHD
for discovering this ;) It far better describes our differences than any
other theory.
Michael,
This isn't a competitive point - but I think I've made that point several
times (and so of course has Hawkins). Quite obviously, (unless you think the
brain has fabulous hidden powers), it conducts searches and other operations
with extremely few limited steps, and nothing remotely like the routine
millions to billions of current computers. It must therefore work v.
fundamentally differently.
Are you saying anything significantly different to that?
--
From: Michael Swan ms...@voyagergaming.com
Sent: Wednesday, July 14, 2010 1:34 AM
To: agi agi@v2.listbox.com
Subject: Re: [agi] What is the smallest set of operations that can
potentially define everything and how do you combine them ?
On Tue, 2010-07-13 at 07:00 -0400, Ben Goertzel wrote:
Well, if you want a simple but complete operator set, you can go with
-- Schonfinkel combinator plus two parentheses
I'll check this out soon.
or
-- S and K combinator plus two parentheses
and I suppose you could add
-- input
-- output
-- forget
statements to this, but I'm not sure what this gets you...
Actually, adding other operators doesn't necessarily
increase the search space your AI faces -- rather, it
**decreases** the search space **if** you choose the right operators,
that
encapsulate regularities in the environment faced by the AI
Unfortunately, an AGI needs to be absolutely general. You are right that
higher level concepts reduce combinations, however, using them, will
increase combinations for simpler operator combinations, and if you
miss a necessary operator, then some concepts will be impossible to
achieve. The smallest set can define higher level concepts, these
concepts can be later integrated as single operations, which means
using operators than can be understood in terms of smaller operators
in the beginning, will definitely increase you combinations later on.
The smallest operator set is like absolute zero. It has a defined end. A
defined way of finding out what they are.
Exemplifying this, writing programs doing humanly simple things
using S and K is a pain and involves piling a lot of S and K and
parentheses
on top of each other, whereas if we introduce loops and conditionals and
such, these programs get shorter. Because loops and conditionals happen
to match the stuff that our human-written programs need to do...
Loops are evil in most situations.
Let me show you why:
Draw a square using put_pixel(x,y)
// loops are more scalable, but, damage this code
Re: [agi] Comments On My Skepticism of Solomonoff Induction
Jim Bromer wrote: Last week I came up with a sketch that I felt showed that Solomonoff Induction was incomputable in practice using a variation of Cantor's Diagonal Argument. Cantor proved that there are more sequences (infinite length strings) than there are (finite length) strings, even though both sets are infinite. This means that some, but not all, sequences have finite length descriptions or are the output of finite length programs (which is the same thing in a more formal sense). For example, the digits of pi or sqrt(2) are infinite length sequences that have finite descriptions (or finite programs that output them). There are many more sequences that don't have finite length descriptions, but unfortunately I can't describe any of them except to say they contain infinite amounts of random data. Cantor does not prove that Solomonoff induction is not computable. That was proved by Kolmogorov (and also by Solomonoff). Solomonoff induction says to use the shortest program that outputs the observed sequence to predict the next symbol. However, there is no procedure for finding the length of the shortest description. The proof sketch is that if there were, then I could describe the first string that cannot be described in less than a million bits even though I just did. The formal proof is http://en.wikipedia.org/wiki/Kolmogorov_complexity#Incomputability_of_Kolmogorov_complexity I think your confusion is using the uncomputability of Solomonoff induction to question its applicability. That is an experimental question, not one of mathematics. The validity of using the shortest or simplest explanation of the past to predict the future was first observed by William of Ockham in the 1400's. It is standard practice in all fields of science. The minimum description length principle is applicable to all branches of machine learning. However, in the conclusion of http://mattmahoney.net/dc/dce.html#Section_Conclusion I argue for Solomonoff induction on the basis of physics. Solomonoff induction supposes that all observable strings are finite prefixes of computable sequences. Occam's Razor might not hold if it were possible for the universe to produce uncomputable sequences, i.e. infinite sources of random data. I argue that is not possible because the observable universe if finitely computable according to the laws of physics as they are now understood. -- Matt Mahoney, matmaho...@yahoo.com From: Jim Bromer jimbro...@gmail.com To: agi agi@v2.listbox.com Sent: Wed, July 14, 2010 11:29:13 AM Subject: [agi] Comments On My Skepticism of Solomonoff Induction Last week I came up with a sketch that I felt showed that Solomonoff Induction was incomputable in practice using a variation of Cantor's Diagonal Argument. I wondered if my argument made sense or not. I will explain why I think it did. First of all, I should have started out by saying something like, Suppose Solomonoff Induction was computable, since there is some reason why people feel that it isn't. Secondly I don't think I needed to use Cantor's Diagonal Argument (for the in practice case), because it would be sufficient to point out that since it was impossible to say whether or not the probabilities ever approached any sustained (collared) limits due to the lack of adequate mathematical definition of the concept all programs, it would be impossible to make the claim that they were actual representations of the probabilities of all programs that could produce certain strings. But before I start to explain why I think my variation of the Diagonal Argument was valid, I would like to make another comment about what was being claimed. Take a look at the n-ary expansion of the square root of 2 (such as the decimal expansion or the binary expansion). The decimal expansion or the binary expansion of the square root of 2 is an infinite string. To say that the algorithm that produces the value is predicting the value is a torturous use of meaning of the word 'prediction'. Now I have less than perfect grammar, but the idea of prediction is so important in the field of intelligence that I do not feel that this kind of reduction of the concept of prediction is illuminating. Incidentally, There are infinite ways to produce the square root of 2 (sqrt 2 +1-1, sqrt2 +2-2, sqrt2 +3-3,...). So the idea that the square root of 2 is unlikely is another stretch of conventional thinking. But since there are an infinite ways for a program to produce any number (that can be produced by a program) we would imagine that the probability that one of the infinite ways to produce the square root of 2 approaches 0 but never reaches it. We can imagine it, but we cannot prove that this occurs in Solomonoff Induction because Solomonoff Induction is not limited to just this class of programs (which could be proven to approach a limit). For example, we could make
Re: [agi] What is the smallest set of operations that can potentially define everything and how do you combine them ?
Michael Swan wrote:
What 3456/6 ?
we don't know, at least not from the top of our head.
No, it took me about 10 or 20 seconds to get 576. Starting with the first
digit,
3/6 = 1/2 (from long term memory) and 3 is in the thousands place, so 1/2 of
1000 is 500 (1/2 = .5 from LTM). I write 500 into short term memory (STM),
which
only has enough space to hold about 7 digits. Then to divide 45/6 I get 42/6 =
7
with a remainder of 3, or 7.5, but since this is in the tens place I get 75. I
put 75 in STM, add to 500 to get 575, put the result back in STM replacing 500
and 75 for which there is no longer room. Finally, 6/6 = 1, which I add to 575
to get 576. I hold this number in STM long enough to check with a calculator.
One could argue that this calculation in my head uses a loop iterator (in STM)
to keep track of which digit I am working on. It definitely involves a sequence
of instructions with intermediate results being stored temporarily. The brain
can only execute 2 or 3 sequential instructions per second and has very limited
short term memory, so it needs to draw from a large database of rules to
perform
calculations like this. A calculator, being faster and having more RAM, is able
to use simpler but more tedious algorithms such as converting to binary,
division by shift and subtract, and converting back to decimal. Doing this with
a carbon based computer would require pencil and paper to make up for lack of
STM, and it would require enough steps to have a high probability of making a
mistake.
Intelligence = knowledge + computing power. The human brain has a lot of
knowledge. The calculator has less knowledge, but makes up for it in speed and
memory.
-- Matt Mahoney, matmaho...@yahoo.com
- Original Message
From: Michael Swan ms...@voyagergaming.com
To: agi agi@v2.listbox.com
Sent: Wed, July 14, 2010 7:53:33 PM
Subject: Re: [agi] What is the smallest set of operations that can potentially
define everything and how do you combine them ?
On Wed, 2010-07-14 at 07:48 -0700, Matt Mahoney wrote:
Actually, Fibonacci numbers can be computed without loops or recursion.
int fib(int x) {
return round(pow((1+sqrt(5))/2, x)/sqrt(5));
}
;) I know. I was wondering if someone would pick up on it. This won't
prove that brains have loops though, so I wasn't concerned about the
shortcuts.
unless you argue that loops are needed to compute sqrt() and pow().
I would find it extremely unlikely that brains have *, /, and even more
unlikely to have sqrt and pow inbuilt. Even more unlikely, even if it
did have them, to figure out how to combine them to round(pow((1
+sqrt(5))/2, x)/sqrt(5)).
Does this mean we should discount all maths that use any complex
operations ?
I suspect the brain is full of look-up tables mainly, with some fairly
primitive methods of combining the data.
eg What's 6 / 3 ?
ans = 2 most people would get that because it's been wrote learnt, a
common problem.
What 3456/6 ?
we don't know, at least not from the top of our head.
The brain and DNA use redundancy and parallelism and don't use loops because
their operations are slow and unreliable. This is not necessarily the best
strategy for computers because computers are fast and reliable but don't have
a
lot of parallelism.
The brains slow and unreliable methods I think are the price paid for
generality and innately unreliable hardware. Imagine writing a computer
program that runs for 120 years without crashing and surviving damage
like a brain can. I suspect the perfect AGI program is a rigorous
combination of the 2.
-- Matt Mahoney, matmaho...@yahoo.com
- Original Message
From: Michael Swan ms...@voyagergaming.com
To: agi agi@v2.listbox.com
Sent: Wed, July 14, 2010 12:18:40 AM
Subject: Re: [agi] What is the smallest set of operations that can
potentially
define everything and how do you combine them ?
Brain loops:
Premise:
Biological brain code does not contain looping constructs, or the
ability to creating looping code, (due to the fact they are extremely
dangerous on unreliable hardware) except for 1 global loop that fires
about 200 times a second.
Hypothesis:
Brains cannot calculate iterative problems quickly, where calculations
in the previous iteration are needed for the next iteration and, where
brute force operations are the only valid option.
Proof:
Take as an example, Fibonacci numbers
http://en.wikipedia.org/wiki/Fibonacci_number
What are the first 100 Fibonacci numbers?
int Fibonacci[102];
Fibonacci[0] = 0;
Fibonacci[1] = 1;
for(int i = 0; i 100; i++)
{
// Getting the next Fibonacci number relies on the previous values
Fibonacci[i+2] = Fibonacci[i] + Fibonacci[i+1];
}
My brain knows the process to solve this problem but it can't directly
write a looping construct into itself. And so it solves it very slowly
compared to a computer.
The brain probably consists of vast repeating look-up
Re: [agi] Mechanical Analogy for Neural Operation!
Steve Richfield wrote: No, I am NOT proposing building mechanical contraptions, just using the concept to compute neuronal characteristics (or AGI formulas for learning). Funny you should mention that. Ross Ashby actually built such a device in 1948 called a homeostat ( http://en.wikipedia.org/wiki/Homeostat ), a fully interconnected neural network with 4 neurons using mechanical components and vacuum tubes. Synaptic weights were implemented by motor driven water filled potentiometers in which electrodes moved through a tank to vary the electrical resistance. It implemented a type of learning algorithm in which weights were varied using a rotating switch wired randomly using the RAND book of a million random digits. He described the device in his 1960 book, Design for a Brain. -- Matt Mahoney, matmaho...@yahoo.com From: Steve Richfield steve.richfi...@gmail.com To: agi agi@v2.listbox.com Sent: Mon, July 12, 2010 2:02:20 AM Subject: [agi] Mechanical Analogy for Neural Operation! Everyone has heard about the water analogy for electrical operation. I have a mechanical analogy for neural operation that just might be solid enough to compute at least some characteristics optimally. No, I am NOT proposing building mechanical contraptions, just using the concept to compute neuronal characteristics (or AGI formulas for learning). Suppose neurons were mechanical contraptions, that receive inputs and communicate outputs via mechanical movements. If one or more of the neurons connected to an output of a neuron, can't make sense of a given input given its other inputs, then its mechanism would physically resist the several inputs that didn't make mutual sense because its mechanism would jam, with the resistance possibly coming from some downstream neuron. This would utilize position to resolve opposing forces, e.g. one force being the observed inputs, and the other force being that they don't make sense, suggest some painful outcome, etc. In short, this would enforce the sort of equation over the present formulaic view of neurons (and AGI coding) that I have suggested in past postings may be present, and show that the math may not be all that challenging. Uncertainty would be expressed in stiffness/flexibility, computed limitations would be handled with over-running clutches, etc. Propagation of forces would come close (perfect?) to being able to identify just where in a complex network something should change to learn as efficiently as possible. Once the force concentrates at some point, it then gives, something slips or bends, to unjam the mechanism. Thus, learning is effected. Note that this suggests little difference between forward propagation and backwards propagation, though real-world wet design considerations would clearly prefer fast mechanisms for forward propagation, and compact mechanisms for backwards propagation. Epiphany or mania? Any thoughts? Steve agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] Solomonoff Induction is Not Universal and Probability is not Prediction
Ben Goertzel wrote: Secondly, since it cannot be computed it is useless. Third, it is not the sort of thing that is useful for AGI in the first place. I agree with these two statements The principle of Solomonoff induction can be applied to computable subsets of the (infinite) hypothesis space. For example, if you are using a neural network to make predictions, the principle says to use the smallest network that computes the past training data. -- Matt Mahoney, matmaho...@yahoo.com From: Ben Goertzel b...@goertzel.org To: agi agi@v2.listbox.com Sent: Fri, July 9, 2010 7:56:53 AM Subject: Re: [agi] Solomonoff Induction is Not Universal and Probability is not Prediction On Fri, Jul 9, 2010 at 7:49 AM, Jim Bromer jimbro...@gmail.com wrote: Abram, Solomoff Induction would produce poor predictions if it could be used to compute them. Solomonoff induction is a mathematical, not verbal, construct. Based on the most obvious mapping from the verbal terms you've used above into mathematical definitions in terms of which Solomonoff induction is constructed, the above statement of yours is FALSE. If you're going to argue against a mathematical theorem, your argument must be mathematical not verbal. Please explain one of 1) which step in the proof about Solomonoff induction's effectiveness you believe is in error 2) which of the assumptions of this proof you think is inapplicable to real intelligence [apart from the assumption of infinite or massive compute resources] Otherwise, your statement is in the same category as the statement by the protagonist of Dostoesvky's Notes from the Underground -- I admit that two times two makes four is an excellent thing, but if we are to give everything its due, two times two makes five is sometimes a very charming thing too. ;-) Secondly, since it cannot be computed it is useless. Third, it is not the sort of thing that is useful for AGI in the first place. I agree with these two statements -- ben G agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] Solomonoff Induction is Not Universal and Probability is not Prediction
Jim Bromer wrote: But, a more interesting question is, given that the first digits are 000, what are the chances that the next digit will be 1? Dim Induction will report .5, which of course is nonsense and a whole less useful than making a rough guess. Wrong. The probability of a 1 is p(0001)/(p()+p(0001)) where the probabilities are computed using Solomonoff induction. A program that outputs will be shorter in most languages than a program that outputs 0001, so 0 is the most likely next bit. More generally, probability and prediction are equivalent by the chain rule. Given any 2 strings x followed by y, the prediction p(y|x) = p(xy)/p(x). -- Matt Mahoney, matmaho...@yahoo.com From: Jim Bromer jimbro...@gmail.com To: agi agi@v2.listbox.com Sent: Wed, July 7, 2010 10:10:37 AM Subject: [agi] Solomonoff Induction is Not Universal and Probability is not Prediction Suppose you have sets of programs that produce two strings. One set of outputs is 00 and the other is 11. Now suppose you used these sets of programs to chart the probabilities of the output of the strings. If the two strings were each output by the same number of programs then you'd have a .5 probability that either string would be output. That's ok. But, a more interesting question is, given that the first digits are 000, what are the chances that the next digit will be 1? Dim Induction will report .5, which of course is nonsense and a whole less useful than making a rough guess. But, of course, Solomonoff Induction purports to be able, if it was feasible, to compute the possibilities for all possible programs. Ok, but now, try thinking about this a little bit. If you have ever tried writing random program instructions what do you usually get? Well, I'll take a hazard and guess (a lot better than the bogus method of confusing shallow probability with prediction in my example) and say that you will get a lot of programs that crash. Well, most of my experiments with that have ended up with programs that go into an infinite loop or which crash. Now on a universal Turing machine, the results would probably look a little different. Some strings will output nothing and go into an infinite loop. Some programs will output something and then either stop outputting anything or start outputting an infinite loop of the same substring. Other programs will go on to infinity producing something that looks like random strings. But the idea that all possible programs would produce well distributed strings is complete hogwash. Since Solomonoff Induction does not define what kind of programs should be used, the assumption that the distribution would produce useful data is absurd. In particular, the use of the method to determine the probability based given an initial string (as in what follows given the first digits are 000) is wrong as in really wrong. The idea that this crude probability can be used as prediction is unsophisticated. Of course you could develop an infinite set of Solomonoff Induction values for each possible given initial sequence of digits. Hey when you're working with infeasible functions why not dream anything? I might be wrong of course. Maybe there is something you guys haven't been able to get across to me. Even if you can think for yourself you can still make mistakes. So if anyone has actually tried writing a program to output all possible programs (up to some feasible point) on a Turing Machine simulator, let me know how it went. Jim Bromer agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] Hutter - A fundamental misdirection?
Gorrell and Webb describe a neural implementation of LSA that seems more
biologically plausible than the usual matrix factoring implementation.
http://www.dcs.shef.ac.uk/~genevieve/gorrell_webb.pdf
In the usual implementation, a word-word matrix A is factored to A = USV where
S
is diagonal (containing eigenvalues), and then the smaller elements of S are
discarded. In the Gorrell model, U and V are the weights of a 3 layer neural
network mapping words to words, and the nonzero elements of S represent the
semantic space in the middle layer. As the network is trained, neurons are
added
to S. Thus, the network is trained online in a single pass, unlike factoring,
which is offline.
-- Matt Mahoney, matmaho...@yahoo.com
From: Gabriel Recchia grecc...@gmail.com
To: agi agi@v2.listbox.com
Sent: Wed, July 7, 2010 12:12:00 PM
Subject: Re: [agi] Hutter - A fundamental misdirection?
In short, instead of a pot of neurons, we might instead have a pot of
dozens
of types of
neurons that each have their own complex rules regarding what other types of
neurons they
can connect to, and how they process information...
...there is plenty of evidence (from the slowness of evolution, the large
number (~200)
of neuron types, etc.), that it is many-layered and quite complex...
The disconnect between the low-level neural hardware and the implementation of
algorithms that build conceptual spaces via dimensionality reduction--which
generally ignore facts such as the existence of different types of neurons, the
apparently hierarchical organization of neocortex, etc.--seems significant.
Have
there been attempts to develop computational models capable of LSA-style feats
(e.g., constructing a vector space in which words with similar meanings tend to
be relatively close to each other) that take into account basic facts about how
neurons actually operate (ideally in a more sophisticated way than the nodes of
early connectionist networks which, as we now know, are not particularly
neuron-like at all)? If so, I would love to know about them.
On Tue, Jun 29, 2010 at 3:02 PM, Ian Parker ianpark...@gmail.com wrote:
The paper seems very similar in principle to LSA. What you need for a concept
vector (or position) is the application of LSA followed by K-Means which will
give you your concept clusters.
I would not knock Hutter too much. After all LSA reduces {primavera,
mamanthal,
salsa, resorte} to one word giving 2 bits saving on Hutter.
- Ian Parker
On 29 June 2010 07:32, rob levy r.p.l...@gmail.com wrote:
Sorry, the link I included was invalid, this is what I meant:
http://www.geog.ucsb.edu/~raubal/Publications/RefConferences/ICSC_2009_AdamsRaubal_Camera-FINAL.pdf
On Tue, Jun 29, 2010 at 2:28 AM, rob levy r.p.l...@gmail.com wrote:
On Mon, Jun 28, 2010 at 5:23 PM, Steve Richfield steve.richfi...@gmail.com
wrote:
Rob,
I just LOVE opaque postings, because they identify people who see things
differently than I do. I'm not sure what you are saying here, so I'll make
some
random responses to exhibit my ignorance and elicit more explanation.
I think based on what you wrote, you understood (mostly) what I was trying
to
get across. So I'm glad it was at least quasi-intelligible. :)
It sounds like this is a finer measure than the dimensionality that I was
referencing. However, I don't see how to reduce anything as quantized as
dimensionality into finer measures. Can you say some more about this?
I was just referencing Gardenfors' research program of conceptual spaces
(I
was intentionally vague about committing to this fully though because I
don't
necessarily think this is the whole answer). Page 2 of this article
summarizes
it pretty succinctly:
http://www.geog.ucsb.edu/.../ICSC_2009_AdamsRaubal_Camera-FINAL.pdf
However, different people's brains, even the brains of identical twins, have
DIFFERENT mappings. This would seem to mandate experience-formed topology.
Yes definitely.
Since these conceptual spaces that structure sensorimotor
expectation/prediction
(including in higher order embodied exploration of concepts I think) are
multidimensional spaces, it seems likely that some kind of neural
computation
over these spaces must occur,
I agree.
though I wonder what it actually would be in terms of neurons, (and if that
matters).
I don't see any route to the answer except via neurons.
I agree this is true of natural intelligence, though maybe in modeling, the
neural level can be shortcut to the topo map level without recourse to
neural
computation (use some more straightforward computation like matrix algebra
instead).
Rob
agi | Archives | Modify Your Subscription
agi | Archives | Modify Your Subscription
agi | Archives | Modify Your Subscription
---
agi
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https
Re: [agi] Reward function vs utility
Perhaps we now have a better understanding of the risks of uploading to a form where we could modify our own software. We already do this to some extent using drugs. Evolution will eliminate such failures. -- Matt Mahoney, matmaho...@yahoo.com From: Abram Demski abramdem...@gmail.com To: agi agi@v2.listbox.com Sent: Sun, July 4, 2010 11:43:46 AM Subject: Re: [agi] Reward function vs utility Joshua, But couldn't it game the external utility function by taking actions which modify it? For example, if the suggestion is taken literally and you have a person deciding the reward at each moment, an AI would want to focus on making that person *think* the reward should be high, rather than focusing on actually doing well at whatever task it's set...and the two would tend to diverge greatly for more and more complex/difficult tasks, since these tend to be harder to judge. Furthermore, the AI would be very pleased to knock the human out of the loop and push its own buttons. Similar comments would apply to automated reward calculations. --Abram On Sun, Jul 4, 2010 at 4:40 AM, Joshua Fox joshuat...@gmail.com wrote: Another point. I'm probably repeating the obvious, but perhaps this will be useful to some. On the one hand, an agent could not game a Legg-like intelligence metric by altering the utility function, even an internal one,, since the metric is based on the function before any such change. On the other hand, since an internally-calculated utility function would necessarily be a function of observations, rather than of actual world state, it could be successfully gamed by altering observations. This latter objection does not apply to functions which are externally calculated, whether known or unknown. Joshua On Fri, Jul 2, 2010 at 7:23 PM, Joshua Fox joshuat...@gmail.com wrote: I found the answer as given by Legg, Machine Superintelligence, p. 72, copied below. A reward function is used to bypass potential difficulty in communicating a utility function to the agent. Joshua The existence of a goal raises the problem of how the agent knows what the goal is. One possibility would be for the goal to be known in advance and for this knowledge to be built into the agent. The problem with this is that it limits each agent to just one goal. We need to allow agents that are more flexible, specifically, we need to be able to inform the agent of what the goal is. For humans this is easily done using language. In general however, the possession of a suffciently high level of language is too strong an assumption to make about the agent. Indeed, even for something as intelligent as a dog or a cat, direct explanation is not very effective. Fortunately there is another possibility which is, in some sense, a blend of the above two. We define an additional communication channel with the sim- plest possible semantics: a signal that indicates how good the agent’s current situation is. We will call this signal the reward. The agent simply has to maximise the amount of reward it receives, which is a function of the goal. In a complex setting the agent might be rewarded for winning a game or solving a puzzle. If the agent is to succeed in its environment, that is, receive a lot of reward, it must learn about the structure of the environment and in particular what it needs to do in order to get reward. On Mon, Jun 28, 2010 at 1:32 AM, Ben Goertzel b...@goertzel.org wrote: You can always build the utility function into the assumed universal Turing machine underlying the definition of algorithmic information... I guess this will improve learning rate by some additive constant, in the long run ;) ben On Sun, Jun 27, 2010 at 4:22 PM, Joshua Fox joshuat...@gmail.com wrote: This has probably been discussed at length, so I will appreciate a reference on this: Why does Legg's definition of intelligence (following on Hutters' AIXI and related work) involve a reward function rather than a utility function? For this purpose, reward is a function of the word state/history which is unknown to the agent while a utility function is known to the agent. Even if we replace the former with the latter, we can still have a definition of intelligence that integrates optimization capacity over possible all utility functions. What is the real significance of the difference between the two types of functions here? Joshua agi | Archives | Modify Your Subscription -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC CTO, Genescient Corp Vice Chairman, Humanity+ Advisor, Singularity University and Singularity Institute External Research Professor, Xiamen University, China b...@goertzel.org “When nothing seems to help, I go look at a stonecutter hammering away at his rock, perhaps a hundred times without as much as a crack showing in it. Yet at the hundred and first blow it will split in two, and I know
Re: [agi] Re: Huge Progress on the Core of AGI
Jim Bromer wrote:
You can't assume a priori that the diagonal argument is not relevant.
When I say infinite in my proof of Solomonoff induction, I mean countably
infinite, as in aleph-null, as in there is a 1 to 1 mapping between the set and
N, the set of natural numbers. There are a countably infinite number of finite
strings, or of finite programs, or of finite length descriptions of any
particular string. For any finite length string or program or description x
with nonzero probability, there are a countably infinite number of finite
length strings or programs or descriptions that are longer and less likely than
x, and a finite number of finite length strings or programs or descriptions
that are either shorter or more likely or both than x.
Aleph-null is larger than any finite integer. This means that for any finite
set and any countably infinite set, there is not a 1 to 1 mapping between the
elements, and if you do map all of the elements of the finite set to elements
of the infinite set, then there are unmapped elements of the infinite set left
over.
Cantor's diagonalization argument proves that there are infinities larger than
aleph-null, such as the cardinality of the set of real numbers, which we call
uncountably infinite. But since I am not using any uncountably infinite sets, I
don't understand your objection.
-- Matt Mahoney, matmaho...@yahoo.com
From: Jim Bromer jimbro...@gmail.com
To: agi agi@v2.listbox.com
Sent: Sat, July 3, 2010 9:43:15 AM
Subject: Re: [agi] Re: Huge Progress on the Core of AGI
On Fri, Jul 2, 2010 at 6:08 PM, Matt Mahoney matmaho...@yahoo.com wrote:
Jim, to address all of your points,
Solomonoff induction claims that the probability of a string is proportional
to the number of programs that output the string, where each program M is
weighted by 2^-|M|. The probability is dominated by the shortest program
(Kolmogorov complexity), but it is not exactly the same. The difference is
small enough that we may neglect it, just as we neglect differences that
depend on choice of language.
The infinite number of programs that could output the infinite number of
strings that are to be considered (for example while using Solomonoff induction
to predict what string is being output) lays out the potential for the
diagonal argument. You can't assume a priori that the diagonal argument is not
relevant. I don't believe that you can prove that it isn't relevant since as
you say, Kolmogorov Complexity is not computable, and you cannot be sure that
you have listed all the programs that were able to output a particular string.
This creates a situation in which the underlying logic of using Solmonoff
induction is based on incomputable reasoning which can be shown using the
diagonal argument.
This kind of criticism cannot be answered with the kinds of presumptions that
you used to derive the conclusions that you did. It has to be answered
directly. I can think of other infinity to infinity relations in which the
potential mappings can be countably derived from the formulas or equations, but
I have yet to see any analysis which explains why this usage can be. Although
you may imagine that the summation of the probabilities can be used just like
it was an ordinary number, the unchecked usage is faulty. In other words the
criticism has to be considered more carefully by someone capable of dealing
with complex mathematical problems that involve the legitimacy of claims
between infinite to infinite mappings.
Jim Bromer
On Fri, Jul 2, 2010 at 6:08 PM, Matt Mahoney matmaho...@yahoo.com wrote:
Jim, to address all of your points,
Solomonoff induction claims that the probability of a string is proportional
to the number of programs that output the string, where each program M is
weighted by 2^-|M|. The probability is dominated by the shortest program
(Kolmogorov complexity), but it is not exactly the same. The difference is
small enough that we may neglect it, just as we neglect differences that
depend on choice of language.
Here is the proof that Kolmogorov complexity is not computable. Suppose it
were. Then I could test the Kolmogorov complexity of strings in increasing
order of length (breaking ties lexicographically) and describe the first
string that cannot be described in less than a million bits, contradicting
the fact that I just did. (Formally, I could write a program that outputs the
first string whose Kolmogorov complexity is at least n bits, choosing n to be
larger than my program).
Here is the argument that Occam's Razor and Solomonoff distribution must be
true. Consider all possible probability distributions p(x) over any infinite
set X of possible finite strings x, i.e. any X = {x: p(x) 0} that is
infinite. All such distributions must favor shorter strings over longer ones.
Consider any x in X. Then p(x) 0. There can be at most a finite number (less
than 1/p(x
Re: [agi] masterpiece on an iPad
It could be done a lot faster if the iPad had a camera. -- Matt Mahoney, matmaho...@yahoo.com From: Mike Tintner tint...@blueyonder.co.uk To: agi agi@v2.listbox.com Sent: Fri, July 2, 2010 6:28:58 AM Subject: [agi] masterpiece on an iPad http://www.telegraph.co.uk/culture/culturevideo/artvideo/7865736/Artist-creates-masterpiece-on-an-iPad.html McLuhan argues that touch is the central sense - the one that binds the others. He may be right. The i-devices integrate touch into intelligence. agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] masterpiece on an iPad
AGI is all about building machines that think, so you don't have to. -- Matt Mahoney, matmaho...@yahoo.com From: Mike Tintner tint...@blueyonder.co.uk To: agi agi@v2.listbox.com Sent: Fri, July 2, 2010 9:37:51 AM Subject: Re: [agi] masterpiece on an iPad that's like saying cartography or cartoons could be done a lot faster if they just used cameras - ask Michael to explain what the hand can draw that the camera can't From: Matt Mahoney Sent: Friday, July 02, 2010 2:21 PM To: agi Subject: Re: [agi] masterpiece on an iPad It could be done a lot faster if the iPad had a camera. -- Matt Mahoney, matmaho...@yahoo.com From: Mike Tintner tint...@blueyonder.co.uk To: agi agi@v2.listbox.com Sent: Fri, July 2, 2010 6:28:58 AM Subject: [agi] masterpiece on an iPad http://www.telegraph.co.uk/culture/culturevideo/artvideo/7865736/Artist-creates-masterpiece-on-an-iPad.html McLuhan argues that touch is the central sense - the one that binds the others. He may be right. The i-devices integrate touch into intelligence. agi | Archives | Modify Your Subscription agi | Archives | Modify Your Subscription agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] masterpiece on an iPad
An AGI only has to predict your behavior so that it can serve you better by giving you what you want without you asking for it. It is not a copy of your mind. It is a program that can call a function that simulates your mind for some arbitrary purpose determined by its programmer. -- Matt Mahoney, matmaho...@yahoo.com From: John G. Rose johnr...@polyplexic.com To: agi agi@v2.listbox.com Sent: Fri, July 2, 2010 11:39:23 AM Subject: RE: [agi] masterpiece on an iPad An AGI may not really think like we do, it may just execute code. Though I suppose you could program a lot of fuzzy loops and idle speculation, entertaining possibilities, having human think envy.. John From:Matt Mahoney [mailto:matmaho...@yahoo.com] Sent: Friday, July 02, 2010 8:21 AM To: agi Subject: Re: [agi] masterpiece on an iPad AGI is all about building machines that think, so you don't have to. -- Matt Mahoney, matmaho...@yahoo.com From:Mike Tintner tint...@blueyonder.co.uk To: agi agi@v2.listbox.com Sent: Fri, July 2, 2010 9:37:51 AM Subject: Re: [agi] masterpiece on an iPad that's like saying cartography or cartoons could be done a lot faster if they just used cameras - ask Michael to explain what the hand can draw that the camera can't From:Matt Mahoney Sent:Friday, July 02, 2010 2:21 PM To:agi Subject:Re: [agi] masterpiece on an iPad It could be done a lot faster if the iPad had a camera. -- Matt Mahoney, matmaho...@yahoo.com From:Mike Tintner tint...@blueyonder.co.uk To: agi agi@v2.listbox.com Sent: Fri, July 2, 2010 6:28:58 AM Subject: [agi] masterpiece on an iPad http://www.telegraph.co.uk/culture/culturevideo/artvideo/7865736/Artist-creates-masterpiece-on-an-iPad.html McLuhan argues that touch is the central sense - the one that binds the others. He may be right. The i-devices integrate touch into intelligence. agi| Archives | Modify Your Subscription agi| Archives| Modify Your Subscription agi| Archives| Modify Your Subscription agi| Archives| ModifyYour Subscription agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] Re: Huge Progress on the Core of AGI
Jim, to address all of your points,
Solomonoff induction claims that the probability of a string is proportional to
the number of programs that output the string, where each program M is weighted
by 2^-|M|. The probability is dominated by the shortest program (Kolmogorov
complexity), but it is not exactly the same. The difference is small enough
that we may neglect it, just as we neglect differences that depend on choice of
language.
Here is the proof that Kolmogorov complexity is not computable. Suppose it
were. Then I could test the Kolmogorov complexity of strings in increasing
order of length (breaking ties lexicographically) and describe the first
string that cannot be described in less than a million bits, contradicting the
fact that I just did. (Formally, I could write a program that outputs the first
string whose Kolmogorov complexity is at least n bits, choosing n to be larger
than my program).
Here is the argument that Occam's Razor and Solomonoff distribution must be
true. Consider all possible probability distributions p(x) over any infinite
set X of possible finite strings x, i.e. any X = {x: p(x) 0} that is
infinite. All such distributions must favor shorter strings over longer ones.
Consider any x in X. Then p(x) 0. There can be at most a finite number (less
than 1/p(x)) of strings that are more likely than x, and therefore an infinite
number of strings which are less likely than x. Of this infinite set, only a
finite number (less than 2^|x|) can be shorter than x, and therefore there must
be an infinite number that are longer than x. So for each x we can partition X
into 4 subsets as follows:
- shorter and more likely than x: finite
- shorter and less likely than x: finite
- longer and more likely than x: finite
- longer and less likely than x: infinite.
So in this sense, any distribution over the set of strings must favor shorter
strings over longer ones.
-- Matt Mahoney, matmaho...@yahoo.com
From: Jim Bromer jimbro...@gmail.com
To: agi agi@v2.listbox.com
Sent: Fri, July 2, 2010 4:09:38 PM
Subject: Re: [agi] Re: Huge Progress on the Core of AGI
On Fri, Jul 2, 2010 at 2:25 PM, Jim Bromer jimbro...@gmail.com wrote:
There cannot be a one to one correspondence to the representation of the
shortest program to produce a string and the strings that they produce. This
means that if the consideration of the hypotheses were to be put into general
mathematical form it must include the potential of many to one relations
between candidate programs (or subprograms) and output strings.
But, there is also no way to determine what the shortest program is, since
there may be different programs that are the same length. That means that
there is a many to one relation between programs and program length. So the
claim that you could just iterate through programs by length is false. This is
the goal of algorithmic information theory not a premise of a methodology that
can be used. So you have the diagonalization problem.
A counter argument is that there are only a finite number of Turing Machine
programs of a given length. However, since you guys have specifically
designated that this theorem applies to any construction of a Turing Machine it
is not clear that this counter argument can be used. And there is still the
specific problem that you might want to try a program that writes a longer
program to output a string (or many strings). Or you might want to write a
program that can be called to write longer programs on a dynamic basis. I
think these cases, where you might consider a program that outputs a longer
program, (or another instruction string for another Turing Machine) constitutes
a serious problem, that at the least, deserves to be answered with sound
analysis.
Part of my original intuitive argument, that I formed some years ago, was that
without a heavy constraint on the instructions for the program, it will be
practically impossible to test or declare that some program is indeed the
shortest program. However, I can't quite get to the point now that I can say
that there is definitely a diagonalization problem.
Jim Bromer
agi | Archives | Modify Your Subscription
---
agi
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/
Modify Your Subscription:
https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c
Powered by Listbox: http://www.listbox.com
Re: [agi] Re: Huge Progress on the Core of AGI
Jim, what evidence do you have that Occam's Razor or algorithmic information theory is wrong, besides your own opinions? It is well established that elegant (short) theories are preferred in all branches of science because they have greater predictive power. Also, what does this have to do with Cantor's diagonalization argument? AIT considers only the countably infinite set of hypotheses. -- Matt Mahoney, matmaho...@yahoo.com From: Jim Bromer jimbro...@gmail.com To: agi agi@v2.listbox.com Sent: Wed, June 30, 2010 9:13:44 AM Subject: Re: [agi] Re: Huge Progress on the Core of AGI On Tue, Jun 29, 2010 at 11:46 PM, Abram Demski abramdem...@gmail.com wrote: In brief, the answer to your question is: we formalize the description length heuristic by assigning lower probabilities to longer hypotheses, and we apply Bayes law to update these probabilities given the data we observe. This updating captures the idea that we should reward theories which explain/expect more of the observations; it also provides a natural way to balance simplicity vs explanatory power, so that we can compare any two theories with a single scoring mechanism. Bayes Law automatically places the right amount of pressure to avoid overly elegant explanations which don't get much right, and to avoid overly complex explanations which fit the observations perfectly but which probably won't generalize to new data. ... If you go down this path, you will eventually come to understand (and, probably, accept) algorithmic information theory. Matt may be tring to force it on you too soon. :) --Abram David was asking about theories of explanation, and here you are suggesting that following a certain path of reasoning will lead to accepting AIT. What nonsense. Even assuming that Baye's law can be used to update probabilities of idealized utility, the connection between description length and explanatory power in general AI is tenuous. And when you realize that AIT is an unattainable idealism that lacks mathematical power (I do not believe that it is a valid mathematical method because it is incomputable and therefore innumerable and cannot be used to derive probability distributions even as ideals) you have to accept that the connection between explanatory theories and AIT is not established except as a special case based on the imagination that a similarities between a subclass of practical examples is the same as a powerful generalization of those examples. The problem is that while compression seems to be related to intelligence, it is not equivalent to intelligence. A much stronger but similarly false argument is that memory is intelligence. Of course memory is a major part of intelligence, but it is not everything. The argument that AIT is a reasonable substitute for developing more sophisticated theories about conceptual explanation is not well founded, it lacks any experimental evidence other than a spattering of results on simplistic cases, and it is just wrong to suggest that there is no reason to consider other theories of explanation. Yes compression has something to do with intelligence and, in some special cases it can be shown to act as an idealism for numerical rationality. And yes unattainable theories that examine the boundaries of productive mathematical systems is a legitimate subject for mathematics. But there is so much more to theories of explanatory reasoning that I genuinely feel sorry for those of you, who originally motivated to develop better AGI programs, would get caught in the obvious traps of AIT and AIXI. Jim Bromer On Tue, Jun 29, 2010 at 11:46 PM, Abram Demski abramdem...@gmail.com wrote: David, What Matt is trying to explain is all right, but I think a better way of answering your question would be to invoke the mighty mysterious Bayes' Law. I had an epiphany similar to yours (the one that started this thread) about 5 years ago now. At the time I did not know that it had all been done before. I think many people feel this way about MDL. Looking into the MDL (minimum description length) literature would be a good starting point. In brief, the answer to your question is: we formalize the description length heuristic by assigning lower probabilities to longer hypotheses, and we apply Bayes law to update these probabilities given the data we observe. This updating captures the idea that we should reward theories which explain/expect more of the observations; it also provides a natural way to balance simplicity vs explanatory power, so that we can compare any two theories with a single scoring mechanism. Bayes Law automatically places the right amount of pressure to avoid overly elegant explanations which don't get much right, and to avoid overly complex explanations which fit the observations perfectly but which probably won't generalize to new data. Bayes' Law and MDL have strong connections, though
Re: [agi] Re: Huge Progress on the Core of AGI
David Jones wrote: If anyone has any knowledge of or references to the state of the art in explanation-based reasoning, can you send me keywords or links? The simplest explanation of the past is the best predictor of the future. http://en.wikipedia.org/wiki/Occam's_razor http://www.scholarpedia.org/article/Algorithmic_probability -- Matt Mahoney, matmaho...@yahoo.com From: David Jones davidher...@gmail.com To: agi agi@v2.listbox.com Sent: Tue, June 29, 2010 9:05:45 AM Subject: [agi] Re: Huge Progress on the Core of AGI If anyone has any knowledge of or references to the state of the art in explanation-based reasoning, can you send me keywords or links? I've read some through google, but I'm not really satisfied with anything I've found. Thanks, Dave On Sun, Jun 27, 2010 at 1:31 AM, David Jones davidher...@gmail.com wrote: A method for comparing hypotheses in explanatory-based reasoning: We prefer the hypothesis or explanation that *expects* more observations. If both explanations expect the same observations, then the simpler of the two is preferred (because the unnecessary terms of the more complicated explanation do not add to the predictive power). Why are expected events so important? They are a measure of 1) explanatory power and 2) predictive power. The more predictive and the more explanatory a hypothesis is, the more likely the hypothesis is when compared to a competing hypothesis. Here are two case studies I've been analyzing from sensory perception of simplified visual input: The goal of the case studies is to answer the following: How do you generate the most likely motion hypothesis in a way that is general and applicable to AGI? Case Study 1) Here is a link to an example: animated gif of two black squares move from left to right. Description: Two black squares are moving in unison from left to right across a white screen. In each frame the black squares shift to the right so that square 1 steals square 2's original position and square two moves an equal distance to the right. Case Study 2) Here is a link to an example: the interrupted square. Description: A single square is moving from left to right. Suddenly in the third frame, a single black square is added in the middle of the expected path of the original black square. This second square just stays there. So, what happened? Did the square moving from left to right keep moving? Or did it stop and then another square suddenly appeared and moved from left to right? Here is a simplified version of how we solve case study 1: The important hypotheses to consider are: 1) the square from frame 1 of the video that has a very close position to the square from frame 2 should be matched (we hypothesize that they are the same square and that any difference in position is motion). So, what happens is that in each two frames of the video, we only match one square. The other square goes unmatched. 2) We do the same thing as in hypothesis #1, but this time we also match the remaining squares and hypothesize motion as follows: the first square jumps over the second square from left to right. We hypothesize that this happens over and over in each frame of the video. Square 2 stops and square 1 jumps over it over and over again. 3) We hypothesize that both squares move to the right in unison. This is the correct hypothesis. So, why should we prefer the correct hypothesis, #3 over the other two? Well, first of all, #3 is correct because it has the most explanatory power of the three and is the simplest of the three. Simpler is better because, with the given evidence and information, there is no reason to desire a more complicated hypothesis such as #2. So, the answer to the question is because explanation #3 expects the most observations, such as: 1) the consistent relative positions of the squares in each frame are expected. 2) It also expects their new positions in each from based on velocity calculations. 3) It expects both squares to occur in each frame. Explanation 1 ignores 1 square from each frame of the video, because it can't match it. Hypothesis #1 doesn't have a reason for why the a new square appears in each frame and why one disappears. It doesn't expect these observations. In fact, explanation 1 doesn't expect anything that happens because something new happens in each frame, which doesn't give it a chance to confirm its hypotheses in subsequent frames. The power of this method is immediately clear. It is general and it solves the problem very cleanly. Here is a simplified version of how we solve case study 2: We expect the original square to move at a similar velocity from left to right because we hypothesized that it did move from left to right and we calculated its velocity. If this expectation is confirmed, then it is more likely than saying that the square suddenly stopped and another started moving
Re: [agi] Re: Huge Progress on the Core of AGI
Right. But Occam's Razor is not complete. It says simpler is better, but 1) this only applies when two hypotheses have the same explanatory power and 2) what defines simpler? A hypothesis is a program that outputs the observed data. It explains the data if its output matches what is observed. The simpler hypothesis is the shorter program, measured in bits. The language used to describe the data can be any Turing complete programming language (C, Lisp, etc) or any natural language such as English. It does not matter much which language you use, because for any two languages there is a fixed length procedure, described in either of the languages, independent of the data, that translates descriptions in one language to the other. For example, the simplest hypothesis for all visual interpretation is that everything in the first image is gone in the second image, and everything in the second image is a new object. Simple. Done. Solved :) right? The hypothesis is not the simplest. The program that outputs the two frames as if independent cannot be smaller than the two frames compressed independently. The program could be made smaller if it only described how the second frame is different than the first. It would be more likely to correctly predict the third frame if it continued to run and described how it would be different than the second frame. I don't think much progress has been made in this area, but I'd like to know what other people have done and any successes they've had. Kolmogorov proved that the solution is not computable. Given a hypothesis (a description of the observed data, or a program that outputs the observed data), there is no general procedure or test to determine whether a shorter (simpler, better) hypothesis exists. Proof: suppose there were. Then I could describe the first data set that cannot be described in less than a million bits even though I just did. (By first I mean the first data set encoded by a string from shortest to longest, breaking ties lexicographically). That said, I believe the state of the art in both language and vision are based on hierarchical neural models, i.e. pattern recognition using learned weighted combinations of simpler patterns. I am more familiar with language. The top ranked programs can be found at http://mattmahoney.net/dc/text.html -- Matt Mahoney, matmaho...@yahoo.com From: David Jones davidher...@gmail.com To: agi agi@v2.listbox.com Sent: Tue, June 29, 2010 10:44:41 AM Subject: Re: [agi] Re: Huge Progress on the Core of AGI Thanks Matt, Right. But Occam's Razor is not complete. It says simpler is better, but 1) this only applies when two hypotheses have the same explanatory power and 2) what defines simpler? So, maybe what I want to know from the state of the art in research is: 1) how precisely do other people define simpler and 2) More importantly, how do you compare competing explanations/hypotheses that have more or less explanatory power. Simpler does not apply unless you are comparing equally explanatory hypotheses. For example, the simplest hypothesis for all visual interpretation is that everything in the first image is gone in the second image, and everything in the second image is a new object. Simple. Done. Solved :) right? Well, clearly a more complicated explanation is warranted because a more complicated explanation is more *explanatory* and a better explanation. So, why is it better? Can it be defined as better in a precise way so that you can compare arbitrary hypotheses or explanations? That is what I'm trying to learn about. I don't think much progress has been made in this area, but I'd like to know what other people have done and any successes they've had. Dave On Tue, Jun 29, 2010 at 10:29 AM, Matt Mahoney matmaho...@yahoo.com wrote: David Jones wrote: If anyone has any knowledge of or references to the state of the art in explanation-based reasoning, can you send me keywords or links? The simplest explanation of the past is the best predictor of the future. http://en.wikipedia.org/wiki/Occam's_razor http://www.scholarpedia.org/article/Algorithmic_probability -- Matt Mahoney, matmaho...@yahoo.com From: David Jones davidher...@gmail.com To: agi agi@v2.listbox.com Sent: Tue, June 29, 2010 9:05:45 AM Subject: [agi] Re: Huge Progress on the Core of AGI If anyone has any knowledge of or references to the state of the art in explanation-based reasoning, can you send me keywords or links? I've read some through google, but I'm not really satisfied with anything I've found. Thanks, Dave On Sun, Jun 27, 2010 at 1:31 AM, David Jones davidher...@gmail.com wrote: A method for comparing hypotheses in explanatory-based reasoning: We prefer the hypothesis or explanation that *expects* more observations. If both explanations expect the same observations, then the simpler of the two
Re: [agi] A Primary Distinction for an AGI
David Jones wrote: I wish people understood this better. For example, animals can be intelligent even though they lack language because they can see. True, but an AGI with language skills is more useful than one without. And yes, I realize that language, vision, motor skills, hearing, and all the other senses and outputs are tied together. Skills in any area make learning the others easier. -- Matt Mahoney, matmaho...@yahoo.com From: David Jones davidher...@gmail.com To: agi agi@v2.listbox.com Sent: Tue, June 29, 2010 1:42:51 PM Subject: Re: [agi] A Primary Distinction for an AGI Mike, THIS is the flawed reasoning that causes people to ignore vision as the right way to create AGI. And I've finally come up with a great way to show you how wrong this reasoning is. I'll give you an extremely obvious argument that proves that vision requires much less knowledge to interpret than language does. Let's say that you have never been to egypt, you have never seen some particular movie before. But if you see the movie, an alien landscape, an alien world, a new place or any such new visual experience, you can immediately interpret it in terms of spacial, temporal, compositional and other relationships. Now, go to egypt and listen to them speak. Can you interpret it? Nope. Why?! Because you don't have enough information. The language itself does not contain any information to help you interpret it. We do not learn language simply by listening. We learn based on evidence from how the language is used and how it occurs in our daily lives. Without that experience, you cannot interpret it. But with vision, you do not need extra knowledge to interpret a new situation. You can recognize completely new objects without any training except for simply observing them in their natural state. I wish people understood this better. Dave On Tue, Jun 29, 2010 at 12:51 PM, Mike Tintner tint...@blueyonder.co.uk wrote: Just off the cuff here - isn't the same true for vision? You can't learn vision from vision. Just as all NLP has no connection with the real world, and totally relies on the human programmer's knowledge of that world. Your visual program actually relies totally on your visual vocabulary - not its own. That is the inevitable penalty of processing unreal signals on a computer screen which are not in fact connected to the real world any more than the verbal/letter signals involved in NLP are. What you need to do - what anyone in your situation with anything like your asprations needs to do - is to hook up with a roboticist. Everyone here should be doing that. From: David Jones Sent: Tuesday, June 29, 2010 5:27 PM To: agi Subject: Re: [agi] A Primary Distinction for an AGI You can't learn language from language without embedding way more knowledge than is reasonable. Language does not contain the information required for its interpretation. There is no *reason* to interpret the language into any of the infinite possible interpretaions. There is nothing to explain but it requires explanatory reasoning to determine the correct real world interpretation On Jun 29, 2010 10:58 AM, Matt Mahoney matmaho...@yahoo.com wrote: David Jones wrote: Natural language requires more than the words on the page in the real world. Of... Any knowledge that can be demonstrated over a text-only channel (as in the Turing test) can also be learned over a text-only channel. Cyc also is trying to store knowledge about a super complicated world in simplistic forms and al... Cyc failed because it lacks natural language. The vast knowledge store of the internet is unintelligible to Cyc. The average person can't use it because they don't speak Cycl and because they have neither the ability nor the patience to translate their implicit thoughts into augmented first order logic. Cyc's approach was understandable when they started in 1984 when they had neither the internet nor the vast computing power that is required to learn natural language from unlabeled examples like children do. Vision and other sensory interpretaion, on the other hand, do not require more info because that... Without natural language, your system will fail too. You don't have enough computing power to learn language, much less the million times more computing power you need to learn to see. -- Matt Mahoney, matmaho...@yahoo.com From: David Jones davidher...@gmail.com To: agi a...@v2.listbox.c...sent: Mon, June 28, 2010 9:28:57 PM Subject: Re: [agi] A Primary Distinction for an AGI Natural language requires more than the words on the page in the real world. Of course that didn't ... agi | Archives | Modify Your Subscription agi | Archives | Modify Your Subscription agi | Archives | Modify Your Subscription agi | Archives | Modify Your Subscription
Re: [agi] Re: Huge Progress on the Core of AGI
David Jones wrote:
I really don't think this is the right way to calculate simplicity.
I will give you an example, because examples are more convincing than proofs.
Suppose you perform a sequence of experiments whose outcome can either be 0 or
1. In the first 10 trials you observe 00. What do you expect to observe
in the next trial?
Hypothesis 1: the outcome is always 0.
Hypothesis 2: the outcome is 0 for the first 10 trials and 1 thereafter.
Hypothesis 1 is shorter than 2, so it is more likely to be correct.
If I describe the two hypotheses in French or Chinese, then 1 is still shorter
than 2.
If I describe the two hypotheses in C, then 1 is shorter than 2.
void hypothesis_1() {
while (1) printf(0);
}
void hypothesis_2() {
int i;
for (i=0; i10; ++i) printf(0);
while (1) printf(1);
}
If I translate these programs into Perl or Lisp or x86 assembler, then 1 will
still be shorter than 2.
I realize there might be smaller equivalent programs. But I think you could
find a smaller program equivalent to hypothesis_1 than hypothesis_2.
I realize there are other hypotheses than 1 or 2. But I think that the smallest
one you can find that outputs eleven bits of which the first ten are zeros will
be a program that outputs another zero.
I realize that you could rewrite 1 so that it is longer than 2. But it is the
shortest version that counts. More specifically consider all programs in which
the first 10 outputs are 0. Then weight each program by 2^-length. So the
shortest programs dominate.
I realize you could make up a language where the shortest encoding of
hypothesis 2 is shorter than 1. You could do this for any pair of hypotheses.
However, I think if you stick to simple languages (and I realize this is a
circular definition), then 1 will usually be shorter than 2.
-- Matt Mahoney, matmaho...@yahoo.com
From: David Jones davidher...@gmail.com
To: agi agi@v2.listbox.com
Sent: Tue, June 29, 2010 1:31:01 PM
Subject: Re: [agi] Re: Huge Progress on the Core of AGI
On Tue, Jun 29, 2010 at 11:26 AM, Matt Mahoney matmaho...@yahoo.com wrote:
Right. But Occam's Razor is not complete. It says simpler is better, but 1)
this only applies when two hypotheses have the same explanatory power and 2)
what defines simpler?
A hypothesis is a program that outputs the observed data. It explains the
data if its output matches what is observed. The simpler hypothesis is the
shorter program, measured in bits.
I can't be confident that bits is the right way to do it. I suspect bits is an
approximation of a more accurate method. I also suspect that you can write a
more complex explanation program with the same number of bits. So, there are
some flaws with this approach. It is an interesting idea to consider though.
The language used to describe the data can be any Turing complete programming
language (C, Lisp, etc) or any natural language such as English. It does not
matter much which language you use, because for any two languages there is a
fixed length procedure, described in either of the languages, independent of
the data, that translates descriptions in
one language to the other.
Hypotheses don't have to be written in actual computer code and probably
shouldn't be because hypotheses are not really meant to be run per say. And
outputs are not necessarily the right way to put it either. Outputs imply
prediction. And as mike has often pointed out, things cannot be precisely
predicted. We can, however, determine whether a particular observation fits
expectations, rather than equals some prediction. There may be multiple
possible outcomes that we expect and which would be consistent with a
hypothesis, which is why actual prediction should not be used.
For example, the simplest hypothesis for all visual interpretation is that
everything in the first image is gone in the second image, and everything in
the second image is a new object. Simple. Done. Solved :) right?
The hypothesis is not the simplest. The program that outputs the two frames as
if independent cannot be smaller than the two frames compressed independently.
The program could be made smaller if it only described how the second frame is
different than the first. It would be more likely to correctly predict the
third frame if it continued to run and described how it would be different
than the second frame.
I really don't think this is the right way to calculate simplicity.
I don't think much progress has been made in this area, but I'd like to know
what other people have done and any successes they've had.
Kolmogorov proved that the solution is not
computable. Given a hypothesis (a description of the observed data, or a
program that outputs the observed data), there is no general procedure or
test to determine whether a shorter (simpler, better) hypothesis exists.
Proof: suppose there were. Then I could describe
Re: [agi] A Primary Distinction for an AGI
Experiments in text compression show that text alone is sufficient for learning to predict text. I realize that for a machine to pass the Turing test, it needs a visual model of the world. Otherwise it would have a hard time with questions like what word in this ernai1 did I spell wrong? Obviously the easiest way to build a visual model is with vision, but it is not the only way. -- Matt Mahoney, matmaho...@yahoo.com From: David Jones davidher...@gmail.com To: agi agi@v2.listbox.com Sent: Tue, June 29, 2010 3:22:33 PM Subject: Re: [agi] A Primary Distinction for an AGI I certainly agree that the techniques and explanation generating algorithms for learning language are hard coded into our brain. But, those techniques alone are not sufficient to learn language in the absence of sensory perception or some other way of getting the data required. Dave On Tue, Jun 29, 2010 at 3:19 PM, Matt Mahoney matmaho...@yahoo.com wrote: David Jones wrote: The knowledge for interpreting language though should not be pre-programmed. I think that human brains are wired differently than other animals to make language learning easier. We have not been successful in training other primates to speak, even though they have all the right anatomy such as vocal chords, tongue, lips, etc. When primates have been taught sign language, they have not successfully mastered forming sentences. -- Matt Mahoney, matmaho...@yahoo.com From: David Jones davidher...@gmail.com To: agi agi@v2.listbox.com Sent: Tue, June 29, 2010 3:00:09 PM Subject: Re: [agi] A Primary Distinction for an AGI The point I was trying to make is that an approach that tries to interpret language just using language itself and without sufficient information or the means to realistically acquire that information, *should* fail. On the other hand, an approach that tries to interpret vision with minimal upfront knowledge needs *should* succeed because the knowledge required to automatically learn to interpret images is amenable to preprogramming. In addition, such knowledge must be pre-programmed. The knowledge for interpreting language though should not be pre-programmed. Dave On Tue, Jun 29, 2010 at 2:51 PM, Matt Mahoney matmaho...@yahoo.com wrote: David Jones wrote: I wish people understood this better. For example, animals can be intelligent even though they lack language because they can see. True, but an AGI with language skills is more useful than one without. And yes, I realize that language, vision, motor skills, hearing, and all the other senses and outputs are tied together. Skills in any area make learning the others easier. -- Matt Mahoney, matmaho...@yahoo.com From: David Jones davidher...@gmail.com To: agi agi@v2.listbox.com Sent: Tue, June 29, 2010 1:42:51 PM Subject: Re: [agi] A Primary Distinction for an AGI Mike, THIS is the flawed reasoning that causes people to ignore vision as the right way to create AGI. And I've finally come up with a great way to show you how wrong this reasoning is. I'll give you an extremely obvious argument that proves that vision requires much less knowledge to interpret than language does. Let's say that you have never been to egypt, you have never seen some particular movie before. But if you see the movie, an alien landscape, an alien world, a new place or any such new visual experience, you can immediately interpret it in terms of spacial, temporal, compositional and other relationships. Now, go to egypt and listen to them speak. Can you interpret it? Nope. Why?! Because you don't have enough information. The language itself does not contain any information to help you interpret it. We do not learn language simply by listening. We learn based on evidence from how the language is used and how it occurs in our daily lives. Without that experience, you cannot interpret it. But with vision, you do not need extra knowledge to interpret a new situation. You can recognize completely new objects without any training except for simply observing them in their natural state. I wish people understood this better. Dave On Tue, Jun 29, 2010 at 12:51 PM, Mike Tintner tint...@blueyonder.co.uk wrote: Just off the cuff here - isn't the same true for vision? You can't learn vision from vision. Just as all NLP has no connection with the real world, and totally relies on the human programmer's knowledge of that world. Your visual program actually relies totally on your visual vocabulary - not its own. That is the inevitable penalty of processing unreal signals on a computer screen which are not in fact connected to the real world any more than the verbal/letter signals involved in NLP are. What you need to do - what anyone in your situation with anything like your asprations needs to do - is to hook up with a roboticist
Re: [agi] A Primary Distinction for an AGI
Answering questions is the same problem as predicting the answers. If you can compute p(A|Q) where Q is the question (and previous context of the conversation) and A is the answer, then you can also choose an answer A from the same distribution. If p() correctly models human communication, then the response would be indistinguishable from a human in a Turing test. -- Matt Mahoney, matmaho...@yahoo.com From: David Jones davidher...@gmail.com To: agi agi@v2.listbox.com Sent: Tue, June 29, 2010 3:43:53 PM Subject: Re: [agi] A Primary Distinction for an AGI the purpose of text is to convey something. It has to be interpreted. who cares about predicting the next word if you can't interpret a single bit of it. On Tue, Jun 29, 2010 at 3:43 PM, David Jones davidher...@gmail.com wrote: People do not predict the next words of text. We anticipate it, but when something different shows up, we accept it if it is *explanatory*. Using compression like algorithms though will never be able to do this type of explanatory reasoning, which is required to disambiguate text. It is certainly not sufficient for learning language, which is not at all about predicting text. On Tue, Jun 29, 2010 at 3:38 PM, Matt Mahoney matmaho...@yahoo.com wrote: Experiments in text compression show that text alone is sufficient for learning to predict text. I realize that for a machine to pass the Turing test, it needs a visual model of the world. Otherwise it would have a hard time with questions like what word in this ernai1 did I spell wrong? Obviously the easiest way to build a visual model is with vision, but it is not the only way. -- Matt Mahoney, matmaho...@yahoo.com From: David Jones davidher...@gmail.com To: agi agi@v2.listbox.com Sent: Tue, June 29, 2010 3:22:33 PM Subject: Re: [agi] A Primary Distinction for an AGI I certainly agree that the techniques and explanation generating algorithms for learning language are hard coded into our brain. But, those techniques alone are not sufficient to learn language in the absence of sensory perception or some other way of getting the data required. Dave On Tue, Jun 29, 2010 at 3:19 PM, Matt Mahoney matmaho...@yahoo.com wrote: David Jones wrote: The knowledge for interpreting language though should not be pre-programmed. I think that human brains are wired differently than other animals to make language learning easier. We have not been successful in training other primates to speak, even though they have all the right anatomy such as vocal chords, tongue, lips, etc. When primates have been taught sign language, they have not successfully mastered forming sentences. -- Matt Mahoney, matmaho...@yahoo.com From: David Jones davidher...@gmail.com To: agi agi@v2.listbox.com Sent: Tue, June 29, 2010 3:00:09 PM Subject: Re: [agi] A Primary Distinction for an AGI The point I was trying to make is that an approach that tries to interpret language just using language itself and without sufficient information or the means to realistically acquire that information, *should* fail. On the other hand, an approach that tries to interpret vision with minimal upfront knowledge needs *should* succeed because the knowledge required to automatically learn to interpret images is amenable to preprogramming. In addition, such knowledge must be pre-programmed. The knowledge for interpreting language though should not be pre-programmed. Dave On Tue, Jun 29, 2010 at 2:51 PM, Matt Mahoney matmaho...@yahoo.com wrote: David Jones wrote: I wish people understood this better. For example, animals can be intelligent even though they lack language because they can see. True, but an AGI with language skills is more useful than one without. And yes, I realize that language, vision, motor skills, hearing, and all the other senses and outputs are tied together. Skills in any area make learning the others easier. -- Matt Mahoney, matmaho...@yahoo.com From: David Jones davidher...@gmail.com To: agi agi@v2.listbox.com Sent: Tue, June 29, 2010 1:42:51 PM Subject: Re: [agi] A Primary Distinction for an AGI Mike, THIS is the flawed reasoning that causes people to ignore vision as the right way to create AGI. And I've finally come up with a great way to show you how wrong this reasoning is. I'll give you an extremely obvious argument that proves that vision requires much less knowledge to interpret than language does. Let's say that you have never been to egypt, you have never seen some particular movie before. But if you see the movie, an alien landscape, an alien world, a new place or any such new visual experience, you can immediately interpret it in terms of spacial, temporal, compositional and other relationships. Now, go to egypt and listen to them speak. Can you interpret it? Nope
Re: [agi] Re: Huge Progress on the Core of AGI
You can always find languages that favor either hypothesis. Suppose that you
want to predict the sequence 10, 21, 32, ? and we write our hypothesis as a
function that takes the trial number (0, 1, 2, 3...) and returns the outcome.
The sequence 10, 21, 32, 43, 54... would be coded:
int hypothesis_1(int trial) {
return trial*11+10;
}
The sequence 10, 21, 32, 10, 21, 32... would be coded
int hypothesis_2(int trial) {
return trial%3*11+10;
}
which is longer and therefore less likely.
Here is another example: predict the sequence 0, 1, 4, 9, 16, 25, 36, 49, ?
Can you find a program shorter than this that doesn't predict 64?
int hypothesis_1(int trial) {
return trial*trial;
}
-- Matt Mahoney, matmaho...@yahoo.com
From: David Jones davidher...@gmail.com
To: agi agi@v2.listbox.com
Sent: Tue, June 29, 2010 3:48:01 PM
Subject: Re: [agi] Re: Huge Progress on the Core of AGI
Such an example is no where near sufficient to accept the assertion that
program size is the right way to define simplicity of a hypothesis.
Here is a counter example. It requires a slightly more complex example because
all zeros doesn't leave any room for alternative hypotheses.
Here is the sequence: 10, 21, 32
void hypothesis_1() {
int ten = 10;
int counter = 0;
while (1)
{
print(ten+counter)
ten = ten + 10;
counter = counter + 1;
}
}
void hypothesis_2() {
while (1)
print(10 21 32)
}
Hypothesis 2 is simpler, yet clearly wrong. These examples don't really show
anything.
Dave
On Tue, Jun 29, 2010 at 3:15 PM, Matt Mahoney matmaho...@yahoo.com wrote:
David Jones wrote:
I really don't think this is the right way to calculate simplicity.
I will give you an example, because examples are more convincing than proofs.
Suppose you perform a sequence of experiments whose outcome can either be 0 or
1. In the first 10 trials you observe 00. What do you expect to
observe in the next trial?
Hypothesis 1: the outcome is always 0.
Hypothesis 2: the outcome is 0 for the first 10 trials and 1 thereafter.
Hypothesis 1 is shorter than 2, so it is more likely to be correct.
If I describe
the two hypotheses in French or Chinese, then 1 is still shorter than 2.
If I describe the two hypotheses in C, then 1 is shorter than 2.
void hypothesis_1() {
while (1) printf(0);
}
void hypothesis_2() {
int i;
for (i=0; i10; ++i) printf(0);
while (1) printf(1);
}
If I translate these programs into Perl or Lisp or x86 assembler, then 1 will
still be shorter than 2.
I realize there might be smaller equivalent programs. But I think you could
find a smaller program equivalent to hypothesis_1 than hypothesis_2.
I realize there are other hypotheses than 1 or 2. But I think that the
smallest one you can find that outputs
eleven bits of which the first ten are zeros will be a program that outputs
another zero.
I realize that you could rewrite 1 so that it is longer than 2. But it is the
shortest version that counts. More specifically consider all programs in which
the first 10 outputs are 0. Then weight each program by 2^-length. So the
shortest programs dominate.
I realize you could make up a language where the shortest encoding of
hypothesis 2 is shorter than 1. You could do this for any pair of hypotheses.
However, I think if you stick to simple languages (and I realize this is a
circular definition), then 1 will usually be shorter than 2.
-- Matt Mahoney, matmaho...@yahoo.com
From: David Jones davidher...@gmail.com
To: agi agi@v2.listbox.com
Sent: Tue, June 29, 2010 1:31:01 PM
Subject: Re: [agi] Re: Huge Progress on the Core of AGI
On Tue, Jun 29, 2010 at 11:26 AM, Matt Mahoney matmaho...@yahoo.com wrote:
Right. But Occam's Razor is not complete. It says simpler is better, but 1)
this only applies when two hypotheses have the same explanatory power and
2) what defines simpler?
A hypothesis is a program that outputs the observed data. It explains the
data if its output matches what is observed. The simpler hypothesis is the
shorter program, measured in bits.
I can't be confident that bits is the right way to do it. I suspect bits is an
approximation of a more accurate method. I also suspect that you can write a
more complex explanation program with the same number of bits. So, there are
some flaws with this approach. It is an interesting idea to consider though.
The language used to describe the data can be any Turing complete programming
language (C, Lisp, etc) or any natural language such as English. It does not
matter much which language you use, because for any two languages there is a
fixed length procedure, described in either of the languages, independent of
the data, that translates descriptions in
one language to the other.
Hypotheses don't have to be written in actual computer code and probably
shouldn't
Re: [agi] A Primary Distinction for an AGI
David Jones wrote: I also want to mention that I develop solutions to the toy problems with the real problems in mind. I also fully intend to work my way up to the real thing by incrementally adding complexity and exploring the problem well at each level of complexity. A little research will show you the folly of this approach. For example, the toy approach to language modeling is to write a simplified grammar that approximates English, then write a parser, then some code to analyze the parse tree and take some action. The classic example is SHRDLU (blocks world, http://en.wikipedia.org/wiki/SHRDLU ). Efforts like that have always stalled. That is not how people learn language. People learn from lots of examples, not explicit rules, and they learn semantics before grammar. For a second example, the toy approach to modeling logical reasoning is to design a knowledge representation based on augmented first order logic, then write code to implement deduction, forward chaining, backward chaining, etc. The classic example is Cyc. Efforts like that have always stalled. That is not how people reason. People learn to associate events that occur in quick succession, and then reason by chaining associations. This model is built in. People might later learn math, programming, and formal logic as rules for manipulating symbols within the framework of natural language learning. For a third example, the toy approach to modeling vision is to segment the image into regions and try to interpret the meaning of each region. Efforts like that have always stalled. That is not how people see. People learn to recognize visual features that they have seen before. Features are made up of weighted sums of lots of simpler features with learned weights. Features range from dots, edges, color, and motion at the lowest levels, to complex objects like faces at the higher levels. Vision is integrated with lots of other knowledge sources. You see what you expect to see. The common theme is that real AGI consists of a learning algorithm, an opaque knowledge representation, and a vast amount of training data and computing power. It is not an extension of a toy system where you code all the knowledge yourself. That doesn't scale. You can't know more than an AGI that knows more than you. So I suggest you do a little research instead of continuing to repeat all the mistakes that were made 50 years ago. You aren't the first person to do these kinds of experiments. -- Matt Mahoney, matmaho...@yahoo.com From: David Jones davidher...@gmail.com To: agi agi@v2.listbox.com Sent: Mon, June 28, 2010 4:00:24 PM Subject: Re: [agi] A Primary Distinction for an AGI I also want to mention that I develop solutions to the toy problems with the real problems in mind. I also fully intend to work my way up to the real thing by incrementally adding complexity and exploring the problem well at each level of complexity. As you do this, the flaws in the design will be clear and I can retrace my steps to create a different solution. The benefit to this strategy is that we fully understand the problems at each level of complexity. When you run into something that is not accounted, you are much more likely to know how to solve it. Despite its difficulties, I prefer my strategy to the alternatives. Dave On Mon, Jun 28, 2010 at 3:56 PM, David Jones davidher...@gmail.com wrote: That does not have to be the case. Yes, you need to know what problems you might have in more complicated domains to avoid developing completely useless theories on toy problems. But, as you develop for full complexity problems, you are confronted with several sub problems. Because you have no previous experience, what tends to happen is you hack together a solution that barely works and simply isn't right or scalable because we don't have a full understanding of the individual sub problems. Having experience with the full problem is important, but forcing yourself to solve every sub problem at once is not a better strategy at all. You may think my strategies has flaws, but I know that and still chose it because the alternative strategies are worse. Dave On Mon, Jun 28, 2010 at 3:41 PM, Russell Wallace russell.wall...@gmail.com wrote: On Mon, Jun 28, 2010 at 4:54 PM, David Jones davidher...@gmail.com wrote: But, that's why it is important to force oneself to solve them in such a way that it IS applicable to AGI. It doesn't mean that you have to choose a problem that is so hard you can't cheat. It's unnecessary to do that unless you can't control your desire to cheat. I can. That would be relevant if it was entirely a problem of willpower and self-discipline, but it isn't. It's also a problem of guidance. A real problem gives you feedback at every step of the way, it keeps blowing your ideas out of the water until you come up with one that will actually work, that you would never have
Re: [agi] Questions for an AGI
This is wishful thinking. Wishful thinking is dangerous. How about instead of hoping that AGI won't destroy the world, you study the problem and come up with a safe design. -- Matt Mahoney, matmaho...@yahoo.com From: rob levy r.p.l...@gmail.com To: agi agi@v2.listbox.com Sent: Sat, June 26, 2010 1:14:22 PM Subject: Re: [agi] Questions for an AGI why should AGIs give a damn about us? I like to think that they will give a damn because humans have a unique way of experiencing reality and there is no reason to not take advantage of that precious opportunity to create astonishment or bliss. If anything is important in the universe, its insuring positive experiences for all areas in which it is conscious, I think it will realize that. And with the resources available in the solar system alone, I don't think we will be much of a burden. I like that idea. Another reason might be that we won't crack the problem of autonomous general intelligence, but the singularity will proceed regardless as a symbiotic relationship between life and AI. That would be beneficial to us as a form of intelligence expansion, and beneficial to the artificial entity a way of being alive and having an experience of the world. agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] Reward function vs utility
The definition of universal intelligence being over all utility functions implies that the utility function is unknown. Otherwise there is a fixed solution. -- Matt Mahoney, matmaho...@yahoo.com From: Joshua Fox joshuat...@gmail.com To: agi agi@v2.listbox.com Sent: Sun, June 27, 2010 4:22:19 PM Subject: [agi] Reward function vs utility This has probably been discussed at length, so I will appreciate a reference on this: Why does Legg's definition of intelligence (following on Hutters' AIXI and related work) involve a reward function rather than a utility function? For this purpose, reward is a function of the word state/history which is unknown to the agent while a utility function is known to the agent. Even if we replace the former with the latter, we can still have a definition of intelligence that integrates optimization capacity over possible all utility functions. What is the real significance of the difference between the two types of functions here? Joshua agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] Questions for an AGI
rob levy wrote: This is wishful thinking. I definitely agree, however we lack a convincing model or plan of any sort for the construction of systems demonstrating subjectivity, Define subjectivity. An objective decision might appear subjective to you only because you aren't intelligent enough to understand the decision process. Therefore it is reasonable to consider symbiosis How does that follow? as both a safe design How do you know that a self replicating organism that we create won't evolve to kill us instead? Do we control evolution? and potentially the only possible design It is not the only possible design. It is possible to create systems that are more intelligent than a single human but less intelligent than all of humanity, without the capability to modify itself or reproduce without the collective permission of the billions of humans that own and maintain control over it. An example would be the internet. -- Matt Mahoney, matmaho...@yahoo.com From: rob levy r.p.l...@gmail.com To: agi agi@v2.listbox.com Sent: Sun, June 27, 2010 2:37:15 PM Subject: Re: [agi] Questions for an AGI I definitely agree, however we lack a convincing model or plan of any sort for the construction of systems demonstrating subjectivity, and it seems plausible that subjectivity is functionally necessary for general intelligence. Therefore it is reasonable to consider symbiosis as both a safe design and potentially the only possible design (at least at first), depending on how creative and resourceful we get in cog sci/ AGI in coming years. On Sun, Jun 27, 2010 at 1:13 PM, Matt Mahoney matmaho...@yahoo.com wrote: This is wishful thinking. Wishful thinking is dangerous. How about instead of hoping that AGI won't destroy the world, you study the problem and come up with a safe design. -- Matt Mahoney, matmaho...@yahoo.com From: rob levy r.p.l...@gmail.com To: agi agi@v2.listbox.com Sent: Sat, June 26, 2010 1:14:22 PM Subject: Re: [agi] Questions for an AGI why should AGIs give a damn about us? I like to think that they will give a damn because humans have a unique way of experiencing reality and there is no reason to not take advantage of that precious opportunity to create astonishment or bliss. If anything is important in the universe, its insuring positive experiences for all areas in which it is conscious, I think it will realize that. And with the resources available in the solar system alone, I don't think we will be much of a burden. I like that idea. Another reason might be that we won't crack the problem of autonomous general intelligence, but the singularity will proceed regardless as a symbiotic relationship between life and AI. That would be beneficial to us as a form of intelligence expansion, and beneficial to the artificial entity a way of being alive and having an experience of the world. agi | Archives | Modify Your Subscription agi | Archives | Modify Your Subscription agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] Questions for an AGI
Travis Lenting wrote: I don't like the idea of enhancing human intelligence before the singularity. The singularity is a point of infinite collective knowledge, and therefore infinite unpredictability. Everything has to happen before the singularity because there is no after. I think crime has to be made impossible even for an enhanced humans first. That is easy. Eliminate all laws. I would like to see the singularity enabling AI to be as least like a reproduction machine as possible. Is there a difference between enhancing our intelligence by uploading and creating killer robots? Think about it. Does it really need to be a general AI to cause a singularity? Can it not just stick to scientific data and quantify human uncertainty? It seems like it would be less likely to ever care about killing all humans so it can rule the galaxy or that its an omnipotent servant. Assume we succeed. People want to be happy. Depending on how our minds are implemented, it's either a matter of rewiring our neurons or rewriting our software. Is that better than a gray goo accident? -- Matt Mahoney, matmaho...@yahoo.com From: Travis Lenting travlent...@gmail.com To: agi agi@v2.listbox.com Sent: Sun, June 27, 2010 5:21:24 PM Subject: Re: [agi] Questions for an AGI I don't like the idea of enhancing human intelligence before the singularity. I think crime has to be made impossible even for an enhanced humans first. I think life is too adapt to abusing opportunities if possible. I would like to see the singularity enabling AI to be as least like a reproduction machine as possible. Does it really need to be a general AI to cause a singularity? Can it not just stick to scientific data and quantify human uncertainty? It seems like it would be less likely to ever care about killing all humans so it can rule the galaxy or that its an omnipotent servant. On Sun, Jun 27, 2010 at 11:39 AM, The Wizard key.unive...@gmail.com wrote: This is wishful thinking. Wishful thinking is dangerous. How about instead of hoping that AGI won't destroy the world, you study the problem and come up with a safe design. Agreed on this dangerous thought! On Sun, Jun 27, 2010 at 1:13 PM, Matt Mahoney matmaho...@yahoo.com wrote: This is wishful thinking. Wishful thinking is dangerous. How about instead of hoping that AGI won't destroy the world, you study the problem and come up with a safe design. -- Matt Mahoney, matmaho...@yahoo.com From: rob levy r.p.l...@gmail.com To: agi agi@v2.listbox.com Sent: Sat, June 26, 2010 1:14:22 PM Subject: Re: [agi] Questions for an AGI why should AGIs give a damn about us? I like to think that they will give a damn because humans have a unique way of experiencing reality and there is no reason to not take advantage of that precious opportunity to create astonishment or bliss. If anything is important in the universe, its insuring positive experiences for all areas in which it is conscious, I think it will realize that. And with the resources available in the solar system alone, I don't think we will be much of a burden. I like that idea. Another reason might be that we won't crack the problem of autonomous general intelligence, but the singularity will proceed regardless as a symbiotic relationship between life and AI. That would be beneficial to us as a form of intelligence expansion, and beneficial to the artificial entity a way of being alive and having an experience of the world. agi | Archives | Modify Your Subscription agi | Archives | Modify Your Subscription -- Carlos A Mejia Taking life one singularity at a time. www.Transalchemy.com agi | Archives | Modify Your Subscription agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] Theory of Hardcoded Intelligence
Correct. Intelligence = log(knowledge) + log(computing power). At the extreme left of your graph is AIXI, which has no knowledge but infinite computing power. At the extreme right you have a giant lookup table. -- Matt Mahoney, matmaho...@yahoo.com From: M E botag...@hotmail.com To: agi agi@v2.listbox.com Sent: Sun, June 27, 2010 5:36:38 PM Subject: [agi] Theory of Hardcoded Intelligence I sketched a graph the other day which represented my thoughts on the usefulness of hardcoding knowledge into an AI. (Graph attached) Basically, the more hardcoded knowledge you include in an AI, of AGI, the lower the overall intelligence it will have, but that faster you will reach that value. I would include any real AGI to be toward the left of the graph with systems like CYC to be toward the right. Matt The New Busy is not the old busy. Search, chat and e-mail from your inbox. Get started. agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] Questions for an AGI
Travis Lenting wrote: Is there a difference between enhancing our intelligence by uploading and creating killer robots? Think about it. Well yes, we're not all bad but I think you read me wrong because thats basically my worry. What I mean is that one way to look at uploading is to create a robot that behaves like you and then dying. The question is whether you become the robot. But it is a nonsense question. Nothing changes whichever way you answer it. Assume we succeed. People want to be happy. Depending on how our minds are implemented, it's either a matter of rewiring our neurons or rewriting our software. Is that better than a gray goo accident? Are you asking if changing your hardware or software ends your true existence like a grey goo accident would? A state of maximum happiness or maximum utility is a degenerate mental state where any thought or perception would be unpleasant because it would result in a different mental state. In a competition with machines that can't have everything they want (for example, they fear death and later die), the other machines would win because you would have no interest in self preservation and they would. Assuming the goo is unconscious, What do you mean by unconscious? it would be worse because there is the potential for a peaceful experience free from the power struggle for limited resources even if humans don't truly exist or not. That result could be reached by a dead planet, which BTW, is the only stable attractor in the chaotic process of evolution. Does anyone else worry about how we're going to keep this machine's unprecedented resourcefulness from being abused by an elite few to further protect and advance their social superiority? If the elite few kill off all their competition, then theirs is the only ethical model that matters. From their point of view, it would be a good thing. How do you feel about humans currently being at the top of the food chain? To me it seems like if we can't create a democratic society where people have real choices concerning the issues that affect them most and it just ends up being a continuation of the class war we have today, then maybe grey goo would be the better option before we start promoting democracy throughout the universe. Freedom and fairness are important to us because they were programmed into our ethical models, not because they are actually important. As a counterexample, they are irrelevant to evolution. Gray goo might be collectively vastly more intelligent than humanity, if that makes you feel any better. -- Matt Mahoney, matmaho...@yahoo.com From: Travis Lenting travlent...@gmail.com To: agi agi@v2.listbox.com Sent: Sun, June 27, 2010 6:53:14 PM Subject: Re: [agi] Questions for an AGI Everything has to happen before the singularity because there is no after. I meant when machines take over technological evolution. That is easy. Eliminate all laws. I would prefer a surveillance state. I should say impossible to get away with if conducted in public. Is there a difference between enhancing our intelligence by uploading and creating killer robots? Think about it. Well yes, we're not all bad but I think you read me wrong because thats basically my worry. Assume we succeed. People want to be happy. Depending on how our minds are implemented, it's either a matter of rewiring our neurons or rewriting our software. Is that better than a gray goo accident? Are you asking if changing your hardware or software ends your true existence like a grey goo accident would? Assuming the goo is unconscious, it would be worse because there is the potential for a peaceful experience free from the power struggle for limited resources even if humans don't truly exist or not. Does anyone else worry about how we're going to keep this machine's unprecedented resourcefulness from being abused by an elite few to further protect and advance their social superiority? To me it seems like if we can't create a democratic society where people have real choices concerning the issues that affect them most and it just ends up being a continuation of the class war we have today, then maybe grey goo would be the better option before we start promoting democracy throughout the universe. On Sun, Jun 27, 2010 at 2:43 PM, Matt Mahoney matmaho...@yahoo.com wrote: Travis Lenting wrote: I don't like the idea of enhancing human intelligence before the singularity. The singularity is a point of infinite collective knowledge, and therefore infinite unpredictability. Everything has to happen before the singularity because there is no after. I think crime has to be made impossible even for an enhanced humans first. That is easy. Eliminate all laws. I would like to see the singularity enabling AI to be as least like a reproduction machine as possible. Is there a difference between enhancing our intelligence
Re: [agi] Questions for an AGI
Am I a human or am I an AGI? Dana Ream wrote: How do you work? Just like you designed me to. deepakjnath wrote: What should I ask if I could ask AGI anything? The Wizard wrote: What should I ask an agi You don't need to ask me anything. I will do all of your thinking for you. Florent Bethert wrote: Tell me what I need to know, by order of importance. Nothing. I will do all of your thinking for you. A. T. Murray wrote: Who killed Donald Young, a gay sex partner of U.S. President Barak Obama It must have been that other AGI, Mentifex. I never did trust it ;-) -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] An alternative plan to discover self-organization theory
Mike Tintner wrote: Matt:It is like the way evolution works, except that there is a human in the loop to make the process a little more intelligent. IOW this is like AGI, except that it's narrow AI. That's the whole point - you have to remove the human from the loop. In fact, it also sounds like a misconceived and rather literal idea of evolution as opposed to the reality. You're right. It is narrow AI. You keep pointing out that we haven't solved the general problem. You are absolutely correct. So, do you have any constructive ideas on how to solve it? Preferably something that takes less than 3 billion years on a planet sized molecular computer. -- Matt Mahoney, matmaho...@yahoo.com From: Mike Tintner tint...@blueyonder.co.uk To: agi agi@v2.listbox.com Sent: Mon, June 21, 2010 7:59:29 AM Subject: Re: [agi] An alternative plan to discover self-organization theory Matt:It is like the way evolution works, except that there is a human in the loop to make the process a little more intelligent. IOW this is like AGI, except that it's narrow AI. That's the whole point - you have to remove the human from the loop. In fact, it also sounds like a misconceived and rather literal idea of evolution as opposed to the reality. From: Matt Mahoney Sent: Monday, June 21, 2010 3:01 AM To: agi Subject: Re: [agi] An alternative plan to discover self-organization theory Steve Richfield wrote: He suggested that I construct a simple NN that couldn't work without self organizing, and make dozens/hundreds of different neuron and synapse operational characteristics selectable ala genetic programming, put it on the fastest computer I could get my hands on, turn it loose trying arbitrary combinations of characteristics, and see what the winning combination turns out to be. Then, armed with that knowledge, refine the genetic characteristics and do it again, and iterate until it efficiently self organizes. This might go on for months, but self-organization theory might just emerge from such an effort. Well, that is the process that created human intelligence, no? But months? It actually took 3 billion years on a planet sized molecular computer. That doesn't mean it won't work. It just means you have to narrow your search space and lower your goals. I can give you an example of a similar process. Look at the code for PAQ8HP12ANY and LPAQ9M data compressors by Alexander Ratushnyak, which are the basis of winning Hutter prize submissions. The basic principle is that you have a model that receives a stream of bits from an unknown source and it uses a complex hierarchy of models to predict the next bit. It is sort of like a neural network because it averages together the results of lots of adaptive pattern recognizers by processes that are themselves adaptive. But I would describe the code as inscrutable, kind of like your DNA. There are lots of parameters to tweak, such as how to preprocess the data, arrange the dictionary, compute various contexts, arrange the order of prediction flows, adjust various learning rates and storage capacities, and make various tradeoffs sacrificing compression to meet memory and speed requirements. It is simple to describe the process of writing the code. You make random changes and keep the ones that work. It is like the way evolution works, except that there is a human in the loop to make the process a little more intelligent. There are also fully automated optimizers for compression algorithms, but they are more limited in their search space. For example, the experimental PPM based EPM by Serge Osnach includes a program EPMOPT that adjusts 20 numeric parameters up or down using a hill climbing search to find the best compression. It can be very slow. Another program, M1X2 by Christopher Mattern, uses a context mixing (PAQ like) algorithm in which the contexts are selected by using a hill climbing genetic algorithm to select a set of 64-bit masks. One version was run for 3 days to find the best options to compress a file that normally takes 45 seconds. -- Matt Mahoney, matmaho...@yahoo.com From: Steve Richfield steve.richfi...@gmail.com To: agi agi@v2.listbox.com Sent: Sun, June 20, 2010 2:06:55 AM Subject: [agi] An alternative plan to discover self-organization theory No, I haven't been smokin' any wacky tobacy. Instead, I was having a long talk with my son Eddie, about self-organization theory. This is his proposal: He suggested that I construct a simple NN that couldn't work without self organizing, and make dozens/hundreds of different neuron and synapse operational characteristics selectable ala genetic programming, put it on the fastest computer I could get my hands on, turn it loose trying arbitrary combinations of characteristics, and see what the winning combination turns out to be. Then, armed with that knowledge, refine
Re: [agi] Re: High Frame Rates Reduce Uncertainty
Your computer monitor flashes 75 frames per second, but you don't notice any flicker because light sensing neurons have a response delay of about 100 ms. Motion detection begins in the retina by cells that respond to contrast between light and dark moving in specific directions computed by simple, fixed weight circuits. Higher up in the processing chain, you detect motion when your eyes and head smoothly track moving objects using kinesthetic feedback from your eye and neck muscles and input from your built in accelerometer in the semicircular canals in your ears. This is all very complicated of course. You are more likely to detect motion in objects that you recognize and expect to move, like people, animals, cars, etc. -- Matt Mahoney, matmaho...@yahoo.com From: David Jones davidher...@gmail.com To: agi agi@v2.listbox.com Sent: Mon, June 21, 2010 9:39:30 AM Subject: [agi] Re: High Frame Rates Reduce Uncertainty Ignoring Steve because we are simply going to have to agree to disagree... And I don't see enough value in trying to understand his paper. I said the math was overly complex, but what I really meant is that the approach is overly complex and so filled with research specific jargon, I don't care to try understand it. It is overly converned with copying the way that the brain does things. I don't care how the brain does it. I care about why the brain does it. Its the same as the analogy of giving a man a fish or teaching him to fish. You may figure out how the brain works, but it does you little good if you don't understand why it works that way. You would have to create a synthetic brain to take advantage of the knowledge, which is not a approach to AGI for many reasons. There are a million other ways, even better ways, to do it than the way the brain does it. Just because the brain accidentally found 1 way out of a million to do it doesn't make it the right way for us to develop AGI. So, moving on I can't find references online, but I've read that the Air Force studied the ability of the human eye to identify aircraft in images that were flashed on a screen at 1/220th of a second. So, clearly, the human eye can at least distinguish 220 fps if it operated that way. Of course, it may not operate on fps second, but that is besides the point. I've also heard other people say that a study has shown that the human eye takes 1000 exposures per second. They had no references though, so it is hearsay. The point was that the brain takes advantage of the fact that with such a high exposure rate, the changes between each image are very small if the objects are moving. This allows it to distinguish movement and visual changes with extremely low uncertainty. If it detects that the changes required to match two parts of an image are too high or the distance between matches is too far, it can reject a match. This allows it to distinguish only very low uncertainty changes and reject changes that have high uncertainty. I think this is a very significant discovery regarding how the brain is able to learn in such an ambiguous world with so many variables that are difficult to disambiguate, interpret and understand. Dave On Fri, Jun 18, 2010 at 2:19 PM, David Jones davidher...@gmail.com wrote: I just came up with an awesome idea. I just realized that the brain takes advantage of high frame rates to reduce uncertainty when it is estimating motion. The slower the frame rate, the more uncertainty there is because objects may have traveled too far between images to match with high certainty using simple techniques. So, this made me think, what if the secret to the brain's ability to learn generally stems from this high frame rate trick. What if we made a system that could process even high frame rates than the brain can. By doing this you can reduce the uncertainty of matches very very low (well in my theory so far). If you can do that, then you can learn about the objects in a video, how they move together or separately with very high certainty. You see, matching is the main barrier when learning about objects. But with a very high frame rate, we can use a fast algorithm and could potentially reduce the uncertainty to almost nothing. Once we learn about objects, matching gets easier because now we have training data and experience to take advantage of. In addition, you can also gain knowledge about lighting, color variation, noise, etc. With that knowledge, you can then automatically create a model of the object with extremely high confidence. You will also be able to determine the effects of light and noise on the object's appearance, which will help match the object invariantly in the future. It allows you to determine what is expected and unexpected for the object's appearance with much higher confidence. Pretty cool idea huh? Dave agi | Archives | Modify Your Subscription
Re: [agi] Fwd: AGI question
rob levy wrote: I am secondarily motivated by the fact that (considerations of morality or amorality aside) AGI is inevitable, though it is far from being a forgone conclusion that powerful general thinking machines will have a first-hand subjective relationship to a world, as living creatures do-- and therefore it is vital that we do as well as possible in understanding what makes systems conscious. A zombie machine intelligence singularity is something I would refer to rather as a holocaust, even if no one were directly killed, assuming these entities could ultimately prevail over the previous forms of life on our planet. What do you mean by conscious? If your brain were removed and replaced by a functionally equivalent computer that simulated your behavior (presumably a zombie), how would you be any different? Why would it matter? -- Matt Mahoney, matmaho...@yahoo.com From: rob levy r.p.l...@gmail.com To: agi agi@v2.listbox.com Sent: Mon, June 21, 2010 11:53:29 AM Subject: [agi] Fwd: AGI question Hi I'm new to this list, but I've been thinking about consciousness, cognition and AI for about half of my life (I'm 32 years old). As is probably the case for many of us here, my interests began with direct recognition of the depth and wonder of varieties of phenomenological experiences-- and attempting to comprehend how these constellations of significance fit in with a larger picture of what we can reliably know about the natural world. I am secondarily motivated by the fact that (considerations of morality or amorality aside) AGI is inevitable, though it is far from being a forgone conclusion that powerful general thinking machines will have a first-hand subjective relationship to a world, as living creatures do-- and therefore it is vital that we do as well as possible in understanding what makes systems conscious. A zombie machine intelligence singularity is something I would refer to rather as a holocaust, even if no one were directly killed, assuming these entities could ultimately prevail over the previous forms of life on our planet. I'm sure I'm not the only one on this list who sees a behavioral/ecological level of analysis as the most likely correct level at which to study perception and cognition, and perception as being a kind of active relationship between an organism and an environment. Having thoroughly convinced my self of a non-dualist, embodied, externalist perspective on cognition, I turn to the nature of life itself (and possibly even physics but maybe that level will not be necessary) to make sense of the nature of subjectivity. I like Bohm's or Bateson's panpsychism about systems as wholes, and significance as informational distinctions (which it would be natural to understand as being the basis of subjective experience), but this is descriptive rather than explanatory. I am not a biologist, but I am increasingly interested in finding answers to what it is about living organisms that gives them a unity such that something is something to the system as a whole. The line of investigation that theoretical biologists like Robert Rosen and other NLDS/chaos people have pursued is interesting, but I am unfamiliar with related work that might have made more progress on the system-level properties that give life its characteristic unity and system-level responsiveness. To me, this seems the most likely candidate for a paradigm shift that would produce AGI. In contrast I'm not particularly convinced that modeling a brain is a good way to get AGI, although I'd guess we could learn a few more things about the coordination of complex behavior if we could really understand them. Another way to put this is that obviously evolutionary computation would be more than just boring hill-climbing if we knew what an organism even IS (perhaps in a more precise computational sense). If we can know what an organism is then it should be (maybe) trivial to model concepts, consciousness, and high level semantics to the umpteenth degree, or at least this would be a major hurtle I think. Even assuming a solution to the problem posed above, there is still plenty of room for other minds skepticism in non-living entities implemented on questionably foreign mediums but there would be a lot more reason to sleep well that the science/technology is leading in a direction in which questions about subjectivity could be meaningfully investigated. Rob agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=8660244-6e7fb59c Powered by Listbox: http://www.listbox.com
Re: [agi] An alternative plan to discover self-organization theory
rob levy wrote: On a related note, what is everyone's opinion on why evolutionary algorithms are such a miserable failure as creative machines, despite their successes in narrow optimization problems? Lack of computing power. How much computation would you need to simulate the 3 billion years of evolution that created human intelligence? -- Matt Mahoney, matmaho...@yahoo.com From: rob levy r.p.l...@gmail.com To: agi agi@v2.listbox.com Sent: Mon, June 21, 2010 11:56:53 AM Subject: Re: [agi] An alternative plan to discover self-organization theory (I'm a little late in this conversation. I tried to send this message the other day but I had my list membership configured wrong. -Rob) -- Forwarded message -- From: rob levy r.p.l...@gmail.com Date: Sun, Jun 20, 2010 at 5:48 PM Subject: Re: [agi] An alternative plan to discover self-organization theory To: agi@v2.listbox.com On a related note, what is everyone's opinion on why evolutionary algorithms are such a miserable failure as creative machines, despite their successes in narrow optimization problems? I don't want to conflate the possibly separable problems of biological development and evolution, though they are interrelated. There are various approaches to evolutionary theory such as Lima de Faria's evolution without selection ideas and Reid's evolution by natural experiment that suggest natural selection is not all it's cracked up to be, and that the step of generating, (mutating, combining, ) is where the more interesting stuff happens. Most of the alternatives to Neodarwinian Synthesis I have seen are based in dynamic models of emergence in complex systems. The upshot is, you don't get creativity for free, you actually still need to solve a problem that is as hard as AGI in order to get creativity for free. So, you would need to solve the AGI-hard problem of evolution and development of life, in order to then solve AGI itself (reminds me of the old SNL sketch: first, get a million dollars...). Also, my hunch is that there is quite a bit of overlap between the solutions to the two problems. Rob Disclaimer: I'm discussing things above that I'm not and don't claim to be an expert in, but from what I have seen so far on this list, that should be alright. AGI is by its nature very multidisciplinary which necessitates often being breadth-first, and therefore shallow in some areas. On Sun, Jun 20, 2010 at 2:06 AM, Steve Richfield steve.richfi...@gmail.com wrote: No, I haven't been smokin' any wacky tobacy. Instead, I was having a long talk with my son Eddie, about self-organization theory. This is his proposal: He suggested that I construct a simple NN that couldn't work without self organizing, and make dozens/hundreds of different neuron and synapse operational characteristics selectable ala genetic programming, put it on the fastest computer I could get my hands on, turn it loose trying arbitrary combinations of characteristics, and see what the winning combination turns out to be. Then, armed with that knowledge, refine the genetic characteristics and do it again, and iterate until it efficiently self organizes. This might go on for months, but self-organization theory might just emerge from such an effort. I had a bunch of objections to his approach, e.g. Q. What if it needs something REALLY strange to work? A. Who better than you to come up with a long list of really strange functionality? Q. There are at least hundreds of bits in the genome. A. Try combinations in pseudo-random order, with each bit getting asserted in ~half of the tests. If/when you stumble onto a combination that sort of works, switch to varying the bits one-at-a-time, and iterate in this way until the best combination is found. Q. Where are we if this just burns electricity for a few months and finds nothing? A. Print out the best combination, break out the wacky tobacy, and come up with even better/crazier parameters to test. I have never written a line of genetic programming, but I know that others here have. Perhaps you could bring some rationality to this discussion? What would be a simple NN that needs self-organization? Maybe a small pot of neurons that could only work if they were organized into layers, e.g. a simple 64-neuron system that would work as a 4x4x4-layer visual recognition system, given the input that I fed it? Any thoughts on how to score partial successes? Has anyone tried anything like this in the past? Is anyone here crazy enough to want to help with such an effort? This Monte Carlo approach might just be simple enough to work, and simple enough that it just HAS to be tried. All thoughts, stones, and rotten fruit will be gratefully appreciated. Thanks in advance. Steve agi | Archives | Modify Your Subscription agi | Archives | Modify Your Subscription --- agi
Re: [agi] An alternative plan to discover self-organization theory
Steve Richfield wrote: He suggested that I construct a simple NN that couldn't work without self organizing, and make dozens/hundreds of different neuron and synapse operational characteristics selectable ala genetic programming, put it on the fastest computer I could get my hands on, turn it loose trying arbitrary combinations of characteristics, and see what the winning combination turns out to be. Then, armed with that knowledge, refine the genetic characteristics and do it again, and iterate until it efficiently self organizes. This might go on for months, but self-organization theory might just emerge from such an effort. Well, that is the process that created human intelligence, no? But months? It actually took 3 billion years on a planet sized molecular computer. That doesn't mean it won't work. It just means you have to narrow your search space and lower your goals. I can give you an example of a similar process. Look at the code for PAQ8HP12ANY and LPAQ9M data compressors by Alexander Ratushnyak, which are the basis of winning Hutter prize submissions. The basic principle is that you have a model that receives a stream of bits from an unknown source and it uses a complex hierarchy of models to predict the next bit. It is sort of like a neural network because it averages together the results of lots of adaptive pattern recognizers by processes that are themselves adaptive. But I would describe the code as inscrutable, kind of like your DNA. There are lots of parameters to tweak, such as how to preprocess the data, arrange the dictionary, compute various contexts, arrange the order of prediction flows, adjust various learning rates and storage capacities, and make various tradeoffs sacrificing compression to meet memory and speed requirements. It is simple to describe the process of writing the code. You make random changes and keep the ones that work. It is like the way evolution works, except that there is a human in the loop to make the process a little more intelligent. There are also fully automated optimizers for compression algorithms, but they are more limited in their search space. For example, the experimental PPM based EPM by Serge Osnach includes a program EPMOPT that adjusts 20 numeric parameters up or down using a hill climbing search to find the best compression. It can be very slow. Another program, M1X2 by Christopher Mattern, uses a context mixing (PAQ like) algorithm in which the contexts are selected by using a hill climbing genetic algorithm to select a set of 64-bit masks. One version was run for 3 days to find the best options to compress a file that normally takes 45 seconds. -- Matt Mahoney, matmaho...@yahoo.com From: Steve Richfield steve.richfi...@gmail.com To: agi agi@v2.listbox.com Sent: Sun, June 20, 2010 2:06:55 AM Subject: [agi] An alternative plan to discover self-organization theory No, I haven't been smokin' any wacky tobacy. Instead, I was having a long talk with my son Eddie, about self-organization theory. This is his proposal: He suggested that I construct a simple NN that couldn't work without self organizing, and make dozens/hundreds of different neuron and synapse operational characteristics selectable ala genetic programming, put it on the fastest computer I could get my hands on, turn it loose trying arbitrary combinations of characteristics, and see what the winning combination turns out to be. Then, armed with that knowledge, refine the genetic characteristics and do it again, and iterate until it efficiently self organizes. This might go on for months, but self-organization theory might just emerge from such an effort. I had a bunch of objections to his approach, e.g. Q. What if it needs something REALLY strange to work? A. Who better than you to come up with a long list of really strange functionality? Q. There are at least hundreds of bits in the genome. A. Try combinations in pseudo-random order, with each bit getting asserted in ~half of the tests. If/when you stumble onto a combination that sort of works, switch to varying the bits one-at-a-time, and iterate in this way until the best combination is found. Q. Where are we if this just burns electricity for a few months and finds nothing? A. Print out the best combination, break out the wacky tobacy, and come up with even better/crazier parameters to test. I have never written a line of genetic programming, but I know that others here have. Perhaps you could bring some rationality to this discussion? What would be a simple NN that needs self-organization? Maybe a small pot of neurons that could only work if they were organized into layers, e.g. a simple 64-neuron system that would work as a 4x4x4-layer visual recognition system, given the input that I fed it? Any thoughts on how to score partial successes? Has anyone tried anything like this in the past? Is anyone here crazy
Re: [agi] just a thought
--- On Wed, 1/14/09, Christopher Carr cac...@pdx.edu wrote: Problems with IQ notwithstanding, I'm confident that, were my silly IQ of 145 merely doubled, I could convince Dr. Goertzel to give me the majority of his assets, including control of his businesses. And if he were to really meet someone that bright, he would be a fool or super-human not to do so, which he isn't (a fool, that is). First, if you knew what you would do if you were twice as smart, you would already be that smart. Therefore you don't know. Second, you have never even met anyone with an IQ of 290. How do you know what they would do? How do you measure an IQ of 100n? - Ability to remember n times as much? - Ability to learn n times faster? - Ability to solve problems n times faster? - Ability to do the work of n people? - Ability to make n times as much money? - Ability to communicate with n people at once? Please give me an IQ test that measures something that can't be done by n log n people (allowing for some organizational overhead). -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=126863270-d7b0b0 Powered by Listbox: http://www.listbox.com
Re: [agi] What Must a World Be That a Humanlike Intelligence May Develop In It?
My response to Ben's paper is to be cautious about drawing conclusions from simulated environments. Human level AGI has an algorithmic complexity of 10^9 bits (as estimated by Landauer). It is not possible to learn this much information from an environment that is less complex. If a baby AI did perform well in a simplified simulation of the world, it would not imply that the same system would work in the real world. It would be like training a language model on a simple, artificial language and then concluding that the system could be scaled up to learn English. This is a lesson from my dissertation work in network intrusion anomaly detection. This was a machine learning task in which the system was trained on attack-free network traffic, and then identified anything out of the ordinary as malicious. For development and testing, we used the 1999 MIT-DARPA Lincoln Labs data set consisting of 5 weeks of synthetic network traffic with hundreds of labeled attacks. The test set developers took great care to make the data as realistic as possible. They collected statistics from real networks, built an isolated network of 4 real computers running different operating systems, and thousands of simulated computers that generated HTTP requests to public websites and mailing lists, and generated synthetic email using English word bigram frequencies, and other kinds of traffic. In my work I discovered a simple algorithm that beat the best intrusion detection systems available at the time. I parsed network packets into individual 1-4 byte fields, recorded all the values that ever occurred at least once in training, and flagged any new value in the test data as suspicious, with a score inversely proportional to the size of the set of values observed in training and proportional to the time since the previous anomaly. Not surprisingly, the simple algorithm failed on real network traffic. There were too many false alarms for it to be even remotely useful. The reason it worked on the synthetic traffic was that it was algorithmically simple compared to real traffic. For example, one of the most effective tests was the TTL value, a counter that decrements with each IP routing hop, intended to prevent routing loops. It turned out that most of the attacks were simulated from a machine that was one hop further away than the machines simulating normal traffic. A problem like that could have been fixed, but there were a dozen others that I found, and probably many that I didn't find. It's not that the test set developers weren't careful. They spent probably $1 million developing it (several people over 2 years). It's that you can't simulate the high complexity of thousands of computers and human users with anything less than that. Simple problems have simple solutions, but that's not AGI. -- Matt Mahoney, matmaho...@yahoo.com --- On Fri, 1/9/09, Ben Goertzel b...@goertzel.org wrote: From: Ben Goertzel b...@goertzel.org Subject: [agi] What Must a World Be That a Humanlike Intelligence May Develop In It? To: agi@v2.listbox.com Date: Friday, January 9, 2009, 5:58 PM Hi all, I intend to submit the following paper to JAGI shortly, but I figured I'd run it past you folks on this list first, and incorporate any useful feedback into the draft I submit This is an attempt to articulate a virtual world infrastructure that will be adequate for the development of human-level AGI http://www.goertzel.org/papers/BlocksNBeadsWorld.pdf Most of the paper is taken up by conceptual and requirements issues, but at the end specific world-design proposals are made. This complements my earlier paper on AGI Preschool. It attempts to define what kind of underlying virtual world infrastructure an effective AGI preschool would minimally require. thx Ben G -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI b...@goertzel.org I intend to live forever, or die trying. -- Groucho Marx --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=126863270-d7b0b0 Powered by Listbox: http://www.listbox.com
Re: [agi] What Must a World Be That a Humanlike Intelligence May Develop In It?
--- On Tue, 1/13/09, Ben Goertzel b...@goertzel.org wrote: The complexity of a simulated environment is tricky to estimate, if the environment contains complex self-organizing dynamics, random number generation, and complex human interactions ... In fact it's not computable. But if you write 10^6 bits of code for your simulator, you know it's less than 10^6 bits. But I wonder which is a better test of AI. http://cs.fit.edu/~mmahoney/compression/text.html is based on natural language prediction, equivalent to the Turing test. The data has 10^9 bits of complexity, just enough to train a human adult language model. http://cs.fit.edu/~mmahoney/compression/uiq/ is based on Legg and Hutter's universal intelligence. It probably has a few hundred bits of complexity, designed to be just beyond the reach of tractability for universal algorithms like AIXI^tl. -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=126863270-d7b0b0 Powered by Listbox: http://www.listbox.com
Re: [agi] [WAS The Smushaby] The Logic of Creativity
I think what Mike is saying is that I could draw what I think a flying house would look like, and you could look at my picture and say it was a flying house, even though neither of us has ever seen one. Therefore, AGI should be able to solve the same kind of problems, and why aren't we designing and testing AGI this way? But don't worry about it. Mike doesn't know how to solve the problem either. -- Matt Mahoney, matmaho...@yahoo.com --- On Tue, 1/13/09, Jim Bromer jimbro...@gmail.com wrote: From: Jim Bromer jimbro...@gmail.com Subject: Re: [agi] [WAS The Smushaby] The Logic of Creativity To: agi@v2.listbox.com Date: Tuesday, January 13, 2009, 3:02 PM I am reluctant to say this, but I am not sure if I actually understand what Mike is getting at. He described a number of logical (in the greater sense of being reasonable and structured) methods by which one could achieve some procedural goal, and then he declares that logic (in this greater sense that I believe acknowledged) was incapable of achieving it. Let's take a flying house. I have to say that there was a very great chance that I misunderstood what Mike was saying, since I believe that he effectively said that a computer program, using logically derived systems could not come to the point where it could creatively draw a picture of a flying house like a child might. If that was what he was saying then it is very strange. Obviously, one could program a computer to draw a flying house. So right away, his point must have been under stated, because that means that a computer program using computer logic (somewhere within this greater sense of the term) could follow a program designed to get it to draw a flying house. So right away, Mike's challenge can't be taken seriously. If we can use logical design to get the computer program to draw a flying house, we can find more creative ways to get it to the same point. Do you understand what I am saying? You aren't actually going to challenge me to write a rather insipid program that will draw a flying house for you are you? You accept the statement that I could do that if I wanted to right? If you do accept that statement, then you should be able to accept the fact that I could also write a more elaborate computer program to do the same thing, only it might, for example, do so only after the words house and flying were input. I think you understand that I could write a slightly more elaborate computer program to do the something like that. Ok, now I could keep making it more complicated and eventually I could get to the point where where it could take parts of pictures that it was exposed to and draw them in more creative combinations. If it was exposed to pictures of airplanes flying, and if it was exposed to pictures of houses, it might,. through quasi random experimentation try drawing a picture of the airplane flying past the house as if the house was an immense mountain, and then it might try some clouds as landscaping for the house and then it might try a cloud with a driveway, garbage can and a chimney, and eventually it might even draw a picture of a house with wings. All I need to do that is to use some shape detecting algorithms that have been developed for graphics programs and are used all the time by graphic artists that can approximately determine the shape of the house and airplane in the different pictures and then it would just be a matter of time before it could (and would) try to draw a flying house. Which step do you doubt, or did I completely misunderstand you? 1. I could (I hope I don't have to) write a program that could draw a flying house. 2. I could make it slightly more elaborate so, for example, that it would only draw the flying house if the words 'house' and 'flying' were input. 3. I could vary the program in many other ways. Now suppose that I showed you one of these programs. After that I could make it more complicated so that it went through a slightly more creative process than the program you saw the previous time. 4. I could continue to make the program more and more complicated. I could, (with a lot of graphics techniques that I know about but haven't actually mastered) write the program so that if it was exposed to pictures of houses and to pictures of flying, would have the ability to eventually draw a picture of a flying house (along with a lot of other creative efforts that you have not) even thought of. But the thing is, that I can do this without using advanced AGI techniques! So, I must retain the recognition that I may not have been able to understand you because what you are saying is not totally reasonable to me. Jim Bromer --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your
Re: [agi] [WAS The Smushaby] The Logic of Creativity
--- On Tue, 1/13/09, Mike Tintner tint...@blueyonder.co.uk wrote: Oh and just to answer Matt - if you want to keep doing narrow AI, like everyone else, then he's right - don't worry about it. Pretend it doesn't exist. Compress things :). Now, Mike, it is actually a simple problem. 1. Collect about 10^8 random photos (about what we see in a lifetime). 2. Label all the ones of houses, and all the ones of things flying. 3. Train an image recognition system (a hierarchical neural network, probably 3-5 layers, 10^7 neurons, 10^11 connections) to detect these two features. You'll need about 10^19 CPU operations, or about a month on a 1000 CPU cluster. 4. Invert the network by iteratively drawing images that activate these two features and work down the hierarchy. (Should be faster than step 3). When you are done, you will have a picture of a flying house. Let me know if you have any trouble implementing this. And BTW the first 2 steps are done. http://images.google.com/images?q=flying+houseum=1ie=UTF-8sa=Xoi=image_result_groupresnum=5ct=title -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=126863270-d7b0b0 Powered by Listbox: http://www.listbox.com
Re: [agi] [WAS The Smushaby] The Logic of Creativity
Mike, it's not cheating. It's called research :-) -- Matt Mahoney, matmaho...@yahoo.com --- On Tue, 1/13/09, Mike Tintner tint...@blueyonder.co.uk wrote: From: Mike Tintner tint...@blueyonder.co.uk Subject: Re: [agi] [WAS The Smushaby] The Logic of Creativity To: agi@v2.listbox.com Date: Tuesday, January 13, 2009, 7:38 PM Matt, Well little Matt, as your class teacher, in one sense this is quite clever of you. But you see, little Matt, when I gave you and the class that exercise, the idea was for you to show me what *you* could do - what you could produce from your own brain. I didn't mean you to copy someone else's flying house from a textbook. That's cheating Matt, - getting someone else to do the work for you - and we don't like cheats do we? So perhaps you can go away and draw a flying house all by yourself - a superduper one with lots of fabbo new bits that no one has ever drawn before, and all kinds of wonderful bells and whistles, that will be ten times better than that silly old foto. I know you can Matt, I have faith in you. And I know if you really, really try, you can understand the difference between creating your own drawing, and copying someone else's. Because, well frankly, Matt, every time I give you an exercise - ask you to write an essay, or tell me a story in your own words - you always, always copy from other people, even if you try to disguise it by copying from several people. Now that's not fair, is it Matt? That's not the American way. You have to get over this lack of confidence in yourself. Matt/Mike Tintner tint...@blueyonder.co.uk wrote: Oh and just to answer Matt - if you want to keep doing narrow AI, like everyone else, then he's right - don't worry about it. Pretend it doesn't exist. Compress things :). Now, Mike, it is actually a simple problem. 1. Collect about 10^8 random photos (about what we see in a lifetime). 2. Label all the ones of houses, and all the ones of things flying. 3. Train an image recognition system (a hierarchical neural network, probably 3-5 layers, 10^7 neurons, 10^11 connections) to detect these two features. You'll need about 10^19 CPU operations, or about a month on a 1000 CPU cluster. 4. Invert the network by iteratively drawing images that activate these two features and work down the hierarchy. (Should be faster than step 3). When you are done, you will have a picture of a flying house. Let me know if you have any trouble implementing this. And BTW the first 2 steps are done. http://images.google.com/images?q=flying+houseum=1ie=UTF-8sa=Xoi=image_result_groupresnum=5ct=title -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=126863270-d7b0b0 Powered by Listbox: http://www.listbox.com
Re: [agi] just a thought
--- On Tue, 1/13/09, Valentina Poletti jamwa...@gmail.com wrote: Anyways my point is, the reason why we have achieved so much technology, so much knowledge in this time is precisely the we, it's the union of several individuals together with their ability to communicate with one-other that has made us advance so much. I agree. A machine that is 10 times as smart as a human in every way could not achieve much more than hiring 10 more people. In order to automate the economy, we have to replicate the capabilities of not one human mind, but a system of 10^10 minds. That is why my AGI proposal is so hideously expensive. http://www.mattmahoney.net/agi2.html -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=126863270-d7b0b0 Powered by Listbox: http://www.listbox.com
Re: [agi] initial reaction to A2I2's call center product
--- On Mon, 1/12/09, Ben Goertzel b...@goertzel.org wrote: AGI company A2I2 has released a product for automating call center functionality, see... http://www.smartaction.com/index.html It would be nice to see some transcripts of actual conversation between the system and customers to get some idea of how well the system actually works. You'll notice that the contact for more information is a live human... -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=126863270-d7b0b0 Powered by Listbox: http://www.listbox.com
Re: [agi] fuzzy-probabilistic logic again
--- On Mon, 1/12/09, YKY (Yan King Yin) generic.intellige...@gmail.com wrote: I have refined my P(Z) logic a bit. Now the truth values are all unified to one type, probability distribution over Z, which has a pretty nice interpretation. The new stuff are at sections 4.4.2 and 4.4.3. http://www.geocities.com/genericai/P-Z-logic-excerpt-12-Jan-2009.pdf Do you have any experimental results supporting your proposed probabilistic fuzzy logic implementation? How would you devise such an experiment (for example, a prediction task) to test alternative interpretations of logical operators like AND, OR, NOT, IF-THEN, etc? Maybe you could manually encoding knowledge in your system (like you did with Goldilocks) and test whether it can make inferences? I'd be more interested to see results on real data, however. (Also, instead of a disclaimer about political correctness, couldn't you just find examples that don't reveal your obsession with sex?) -- Matt Mahoney, matmaho...@yahoo.com I'm wondering if anyone is interested in helping me implement the logic or develop an AGI basing on it? I have already written part of the inference engine in Lisp. Also, is anyone here working on fuzzy or probabilistic logics, other than Ben and Pei and me? YKY --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=126863270-d7b0b0 Powered by Listbox: http://www.listbox.com
Re: [agi] The Smushaby of Flatway.
Mike, after a sequence of free associations, you drift from the original
domain. How is that incompatible with the model I described? I use A, B, C, as
variables to represent arbitrary thoughts.
-- Matt Mahoney, matmaho...@yahoo.com
--- On Fri, 1/9/09, Mike Tintner tint...@blueyonder.co.uk wrote:
From: Mike Tintner tint...@blueyonder.co.uk
Subject: Re: [agi] The Smushaby of Flatway.
To: agi@v2.listbox.com
Date: Friday, January 9, 2009, 10:08 AM
I
_filtered #yiv455060292 {
font-family:Courier;}
_filtered #yiv455060292 {
font-family:Tms Rmn;}
_filtered #yiv455060292 {margin:1.0in 77.95pt 1.0in 77.95pt;}
#yiv455060292 P.MsoNormal {
FONT-SIZE:12pt;MARGIN:0in 0in 0pt;FONT-FAMILY:Courier;}
#yiv455060292 LI.MsoNormal {
FONT-SIZE:12pt;MARGIN:0in 0in 0pt;FONT-FAMILY:Courier;}
#yiv455060292 DIV.MsoNormal {
FONT-SIZE:12pt;MARGIN:0in 0in 0pt;FONT-FAMILY:Courier;}
#yiv455060292 H1 {
FONT-WEIGHT:normal;FONT-SIZE:12pt;MARGIN:12pt 0in 0pt;FONT-FAMILY:Courier;}
#yiv455060292 H2 {
FONT-WEIGHT:normal;FONT-SIZE:12pt;MARGIN:6pt 0in 0pt;FONT-FAMILY:Courier;}
#yiv455060292 H3 {
FONT-WEIGHT:normal;FONT-SIZE:12pt;MARGIN:0in 0in 0pt;FONT-FAMILY:Courier;}
#yiv455060292 H4 {
FONT-WEIGHT:normal;FONT-SIZE:12pt;MARGIN:0in 0in 0pt;FONT-FAMILY:Courier;}
#yiv455060292 H5 {
FONT-WEIGHT:normal;FONT-SIZE:12pt;MARGIN:0in 0in 0pt;FONT-FAMILY:Courier;}
#yiv455060292 H6 {
FONT-WEIGHT:normal;FONT-SIZE:12pt;MARGIN:0in 0in 0pt;FONT-FAMILY:Courier;}
#yiv455060292 P.MsoHeading7 {
FONT-SIZE:12pt;MARGIN:0in 0in 0pt;FONT-FAMILY:Courier;}
#yiv455060292 LI.MsoHeading7 {
FONT-SIZE:12pt;MARGIN:0in 0in 0pt;FONT-FAMILY:Courier;}
#yiv455060292 DIV.MsoHeading7 {
FONT-SIZE:12pt;MARGIN:0in 0in 0pt;FONT-FAMILY:Courier;}
#yiv455060292 P.MsoHeading8 {
FONT-SIZE:12pt;MARGIN:0in 0in 0pt;FONT-FAMILY:Courier;}
#yiv455060292 LI.MsoHeading8 {
FONT-SIZE:12pt;MARGIN:0in 0in 0pt;FONT-FAMILY:Courier;}
#yiv455060292 DIV.MsoHeading8 {
FONT-SIZE:12pt;MARGIN:0in 0in 0pt;FONT-FAMILY:Courier;}
#yiv455060292 P.MsoHeading9 {
FONT-SIZE:12pt;MARGIN:0in 0in 0pt;FONT-FAMILY:Courier;}
#yiv455060292 LI.MsoHeading9 {
FONT-SIZE:12pt;MARGIN:0in 0in 0pt;FONT-FAMILY:Courier;}
#yiv455060292 DIV.MsoHeading9 {
FONT-SIZE:12pt;MARGIN:0in 0in 0pt;FONT-FAMILY:Courier;}
#yiv455060292 P.MsoNormalIndent {
FONT-SIZE:12pt;MARGIN:0in 0in 0pt 0.5in;FONT-FAMILY:Courier;}
#yiv455060292 LI.MsoNormalIndent {
FONT-SIZE:12pt;MARGIN:0in 0in 0pt 0.5in;FONT-FAMILY:Courier;}
#yiv455060292 DIV.MsoNormalIndent {
FONT-SIZE:12pt;MARGIN:0in 0in 0pt 0.5in;FONT-FAMILY:Courier;}
#yiv455060292 A:link {
COLOR:blue;TEXT-DECORATION:underline;}
#yiv455060292 SPAN.MsoHyperlink {
COLOR:blue;TEXT-DECORATION:underline;}
#yiv455060292 A:visited {
COLOR:purple;TEXT-DECORATION:underline;}
#yiv455060292 SPAN.MsoHyperlinkFollowed {
COLOR:purple;TEXT-DECORATION:underline;}
#yiv455060292 P.MsoPlainText {
FONT-SIZE:10pt;MARGIN:0in 0in 0pt;FONT-FAMILY:Courier New;}
#yiv455060292 LI.MsoPlainText {
FONT-SIZE:10pt;MARGIN:0in 0in 0pt;FONT-FAMILY:Courier New;}
#yiv455060292 DIV.MsoPlainText {
FONT-SIZE:10pt;MARGIN:0in 0in 0pt;FONT-FAMILY:Courier New;}
#yiv455060292 DIV.Section1 {
}
Matt,
I mainly want to lay down a marker here for a
future discussion.
What you have done is what all AGI-ers/AI-ers
do. Faced with the problem of domain-switching - (I pointed out that the human
brain and human thought are * freely domain-switching*), - you
have simply ignored it - and, I imagine, are completely unaware that you have
done so. And this, remember, is *the* problem of AGI - what should be
the central focus of all discussion here.
If you look at your examples, you will find that
they are all *intra-domain* and do not address domain-switching at all
-
a. if you learned the associations A-B and B-C, then A will predict C.
That is called reasoning
b) a word-word
matrix M from a large text corpus, ..gives you something similar to
your free association chain like rain-wet-water-...
No domain-switching there.
Compare these
with my
b)
domain-switching chain -COW - DOG - TAIL - CURRENT CRISIS - LOCAL
VS
GLOBAL
THINKING - WHAT A NICE DAY - MUST GET ON- CANT
SPEND MUCH
MORE TIME ON
THIS
(switching between the domains of - Animals - Politics/Economics -
Weather - Personal Timetable)
a) your
(extremely limited) idea of (logical) reasoning is also entirely
intra-domain - the domain of the Alphabet,
(A-B-C).
But my creative
and similar creative chains are analogous to switching from say
an Alphabet domain (A-B-C) to a Foreign Languages domain (alpha - omega)
to a Semiotics one (symbol - sign - representation) to a Fonts one (Courier -
Times Roman) etc. etc. - i.e. we could all easily and spontaneously form
such a domain-switching chain.
Your programs
and all the programs ever written are still incapable of doing this - switching
domains. This, it bears repeating, is the problem of
AGI.
Because you're
ignoring it, you don't see that you're in effect maintaining an absurdity
Re: [agi] The Smushaby of Flatway.
--- On Thu, 1/8/09, Vladimir Nesov robot...@gmail.com wrote: I claim that K(P) K(Q) because any description of P must include a description of Q plus a description of what P does for at least one other input. Even if you somehow must represent P as concatenation of Q and something else (you don't need to), it's not true that always K(P)K(Q). It's only true that length(P)length(Q), and longer strings can easily have smaller programs that output them. If P is 10^(10^10) symbols X, and Q is some random number of X smaller than 10^(10^10), it's probably K(P)K(Q), even though Q is a substring of P. Well, it is true that you can find |P| |Q| for some cases of P nontrivially simulating Q depending on the choice of language. However, it is not true on average. It is also not possible for P to nontrivially simulate itself because it is a contradiction to say that P does everything that Q does and at least one thing that Q doesn't do if P = Q. -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
Re: [agi] The Smushaby of Flatway.
--- On Wed, 1/7/09, Ben Goertzel b...@goertzel.org wrote: if proving Fermat's Last theorem was just a matter of doing math, it would have been done 150 years ago ;-p obviously, all hard problems that can be solved have already been solved... ??? In theory, FLT could be solved by brute force enumeration of proofs until a match to Wiles' is found. In theory, AGI could be solved by coding all the knowledge in LISP. The difference is that 50 years ago people actually expected the latter to work. Some people still believe so. AGI is an engineering and policy problem. We already have small scale neural models of learning, language, vision, and motor control. We currently lack the computing power (10^16 OPS, 10^15 bits) to implement these at human levels, but Moore's law will take care of that. But that is not the hard part of the problem. AGI is a system that eliminates our need to work, to think, and to function in the real world. Its value is USD 10^15, the value of the global economy. Once we have the hardware, we still need to extract 10^18 bits of knowledge from human brains. That is the complexity of the global economy (assuming 10^10 people x 10^9 bits per person x 0.1 fraction consisting of unique job skills). This is far bigger than the internet. The only way to extract this knowledge without new technology like brain scanning is by communication at the rate of 2 bits per second per person. The cheapest option is a system of pervasive surveillance where everything you say and do is public knowledge. AGI is too expensive for any person or group to build or own. It is a vastly improved internet, a communication system so efficient that the world's population starts to look like a single entity, and nobody notices or cares as silicon gradually replaces carbon. -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
Re: [agi] The Smushaby of Flatway.
--- On Thu, 1/8/09, Mike Tintner tint...@blueyonder.co.uk wrote: What then do you see as the way people *do* think? You surprise me, Matt, because both the details of your answer here and your thinking generally strike me as *very* logicomathematical - with lots of emphasis on numbers and compression - yet you seem to be acknowledging here, like Jim, the fundamental deficiencies of the logicomathematical form - and it is indeed only one form - of thinking. Pattern recognition in parallel, and hierarchical learning of increasingly complex patterns by classical conditioning (association), clustering in context space (feature creation), and reinforcement learning to meet evolved goals. You can't write a first order logic expression that inputs a picture and tells you whether it is a cat or a dog. Yet any child can do it. Logic is great for abstract mathematics. We regard it as the highest form of thought, the hardest thing that humans can learn, yet it is the easiest problem to solve on a computer. -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
Re: [agi] The Smushaby of Flatway.
--- On Thu, 1/8/09, Mike Tintner tint...@blueyonder.co.uk wrote: Matt, Thanks. But how do you see these: Pattern recognition in parallel, and hierarchical learning of increasingly complex patterns by classical conditioning (association), clustering in context space (feature creation), and reinforcement learning to meet evolved goals. as fundamentally different from logicomathematical thinking? (Reinforcement learning strikes me as literally extraneous and not a mode of thinking). Perhaps you need to explain why conditioned association is different. Free association is the basic way of recalling memories. If you experience A followed by B, then the next time you experience A you will think of (or predict) B. Pavlov demonstrated this type of learning in animals in 1927. Hebb proposed a neural model in 1949 which has since been widely accepted. The model is unrelated to first order logic. It is a strengthening of the connections from neuron A to neuron B. -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
Re: [agi] The Smushaby of Flatway.
Mike, Your own thought processes only seem mysterious because you can't predict what you will think without actually thinking it. It's not just a property of the human brain, but of all Turing machines. No program can non-trivially model itself. (By model, I mean that P models Q if for any input x, P can compute the output Q(x). By non-trivial, I mean that P does something else besides just model Q. (Every program trivially models itself). The proof is that for P to non-trivially model Q requires K(P) K(Q), where K is Kolmogorov complexity, because P needs a description of Q plus whatever else it does to make it non-trivial. It is obviously not possible for K(P) K(P)). So if you learned the associations A-B and B-C, then A will predict C. That is called reasoning. Also, each concept is associated with thousands of other concepts, not just A-B. If you pick the strongest associated concept not previously activated, you get the semi-random thought chain you describe. You can demonstrate this with a word-word matrix M from a large text corpus, where M[i,j] is the degree to which the i'th word in the vocabulary is associated with the j'th word, as measured by the probability of finding both words near each other in the corpus. Thus, M[rain,wet] and M[wet,water] have high values because the words often appear in the same paragraph. Traversing related words in M gives you something similar to your free association chain like rain-wet-water-... -- Matt Mahoney, matmaho...@yahoo.com --- On Thu, 1/8/09, Mike Tintner tint...@blueyonder.co.uk wrote: From: Mike Tintner tint...@blueyonder.co.uk Subject: Re: [agi] The Smushaby of Flatway. To: agi@v2.listbox.com Date: Thursday, January 8, 2009, 3:54 PM Matt:Free association is the basic way of recalling memories. If you experience A followed by B, then the next time you experience A you will think of (or predict) B. Pavlov demonstrated this type of learning in animals in 1927. Matt, You're not thinking your argument through. Look carefully at my spontaneous COW - DOG - TAIL - CURRENT CRISIS - LOCAL VS GLOBAL THINKING - WHAT A NICE DAY - MUST GET ON- CANT SPEND MUCH MORE TIME ON THIS etc. etc that's not A-B association. That's 1. A-B-C then 2. Gamma-Delta then 3. Languages then 4. Number of Lines in Letters. IOW the brain is typically not only freely associating *ideas* but switching freely across, and connecting, radically different *domains* in any given chain of association. [e.g above from Animals to Economics/Politics to Weather to Personal Timetable] It can do this partly because a) single ideas have multiple, often massively mutiple, idea/domain connections in the human brain, and allow one to go off in any of multiple tangents/directions b) humans have many things - and therefore multiple domains - on their mind at the same time concurrently - and can switch as above from the immediate subject to some other pressing subject domain (e.g. from economics/politics (local vs global) to the weather (what a nice day). If your A-B, everything-is-memory-recall thesis were true, our chains-of-thought-association would be largely repetitive, and the domain switches inevitable.. In fact, our chains (or networks) of free association and domain-switching are highly creative, and each one is typically, from a purely technical POV, novel and surprising. (I have never connected TAIL and CURRENT CRISIS before - though Animals and Politics yes. Nor have I connected LOCAL VS GLOBAL THINKING before with WHAT A NICE DAY and the weather). IOW I'm suggesting, the natural mode of human thought - and our continuous streams of association - are creative. And achieving such creativity is the principal problem/goal of AGI. So maybe it's worth taking 20 secs. of time - producing your own chain-of-free-association starting say with MAHONEY and going on for another 10 or so items - and trying to figure out how the result could.possibly be the narrow kind of memory-recall you're arguing for. It's an awful lot to ask for, but could you possibly try it, analyse it and report back? [Ben claims to have heard every type of argument I make before, (somewhat like your A-B memory claim), so perhaps he can tell me where he's read before about the Freely Associative, Freely Domain Switching nature of human thought - I'd be interested to follow up on it]. --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
Re: [agi] The Smushaby of Flatway.
--- On Thu, 1/8/09, Vladimir Nesov robot...@gmail.com wrote: On Fri, Jan 9, 2009 at 12:19 AM, Matt Mahoney matmaho...@yahoo.com wrote: Mike, Your own thought processes only seem mysterious because you can't predict what you will think without actually thinking it. It's not just a property of the human brain, but of all Turing machines. No program can non-trivially model itself. (By model, I mean that P models Q if for any input x, P can compute the output Q(x). By non-trivial, I mean that P does something else besides just model Q. (Every program trivially models itself). The proof is that for P to non-trivially model Q requires K(P) K(Q), where K is Kolmogorov complexity, because P needs a description of Q plus whatever else it does to make it non-trivial. It is obviously not possible for K(P) K(P)). Matt, please stop. I even constructed an explicit counterexample to this pseudomathematical assertion of yours once. You don't pay enough attention to formal definitions: what this has a description means, and which reference TMs specific Kolmogorov complexities are measured in. Your earlier counterexample was a trivial simulation. It simulated itself but did nothing else. If P did something that Q didn't, then Q would not be simulating P. This applies regardless of your choice of universal TM. I suppose I need to be more precise. I say P simulates Q if for all x, P(what is Q(x)?) = Q(x)=y iff Q(x)=y (where x and y are arbitrary strings). When I say that P does something else, I mean that it accepts at least one input not of the form what is Q(x)?. I claim that K(P) K(Q) because any description of P must include a description of Q plus a description of what P does for at least one other input. -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
Re: [agi] The Smushaby of Flatway.
Logic has not solved AGI because logic is a poor model of the way people think. Neural networks have not solved AGI because you would need about 10^15 bits of memory and 10^16 OPS to simulate a human brain sized network. Genetic algorithms have not solved AGI because the computational requirements are even worse. You would need 10^36 bits just to model all the world's DNA, and even if you could simulate it in real time, it took 3 billion years to produce human intelligence the first time. Probabilistic reasoning addresses only one of the many flaws of first order logic as a model of AGI. Reasoning under uncertainty is fine, but you haven't solved learning by induction, reinforcement learning, complex pattern recognition (e.g. vision), and language. If it was just a matter of writing the code, then it would have been done 50 years ago. -- Matt Mahoney, matmaho...@yahoo.com --- On Wed, 1/7/09, Jim Bromer jimbro...@gmail.com wrote: From: Jim Bromer jimbro...@gmail.com Subject: [agi] The Smushaby of Flatway. To: agi@v2.listbox.com Date: Wednesday, January 7, 2009, 8:23 PM All of the major AI paradigms, including those that are capable of learning, are flat according to my definition. What makes them flat is that the method of decision making is minimally-structured and they funnel all reasoning through a single narrowly focused process that smushes different inputs to produce output that can appear reasonable in some cases but is really flat and lacks any structure for complex reasoning. The classic example is of course logic. Every proposition can be described as being either True or False and any collection of propositions can be used in the derivation of a conclusion regardless of whether the input propositions had any significant relational structure that would actually have made it reasonable to draw the definitive conclusion that was drawn from them. But logic didn't do the trick, so along came neural networks and although the decision making is superficially distributed and can be thought of as being comprised of a structure of layer-like stages in some variations, the methodology of the system is really just as flat. Again anything can be dumped into the neural network and a single decision making process works on the input through a minimally-structured reasoning system and output is produced regardless of the lack of appropriate relative structure in it. In fact, this lack of discernment was seen as a major breakthrough! Surprise, neural networks did not work just like the mind works in spite of the years and years of hype-work that went into repeating this slogan in the 1980's. Then came Genetic Algorithms and finally we had a system that could truly learn to improve on its previous learning and how did it do this? It used another flat reasoning method whereby combinations of data components were processed according to one simple untiring method that was used over and over again regardless of any potential to see input as being structured in more ways than one. Is anyone else starting to discern a pattern here? Finally we reach the next century to find that the future of AI has already arrived and that future is probabilistic reasoning! And how is probabilistic reasoning different? Well, it can solve problems that logic, neural networks, genetic algorithms couldn't! And how does probabilistic reasoning do this? It uses a funnel minimally-structured method of reasoning whereby any input can be smushed together with other disparate input to produce a conclusion which is only limited by the human beings who strive to program it! The very allure of minimally-structured reasoning is that it works even in some cases where it shouldn't. It's the hip hooray and bally hoo of the smushababies of Flatway. Jim Bromer --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
RE: [agi] Universal intelligence test benchmark
John, So if consciousness is important for compression, then I suggest you write two compression programs, one conscious and one not, and see which one compresses better. Otherwise, this is nonsense. -- Matt Mahoney, matmaho...@yahoo.com --- On Tue, 12/30/08, John G. Rose johnr...@polyplexic.com wrote: From: John G. Rose johnr...@polyplexic.com Subject: RE: [agi] Universal intelligence test benchmark To: agi@v2.listbox.com Date: Tuesday, December 30, 2008, 9:46 AM If the agents were p-zombies or just not conscious they would have different motivations. Consciousness has properties of communication protocol and effects inter-agent communication. The idea being it enhances agents' existence and survival. I assume it facilitates collective intelligence, generally. For a multi-agent system with a goal of compression or prediction the agent consciousness would have to be catered. So introducing - Consciousness of X is: the idea or feeling that X is correlated with Consciousness of X to the agents would give them more glue if they expended that consciousness on one another. The communications dynamics of the system would change verses a similar non-conscious multi-agent system. John From: Ben Goertzel [mailto:b...@goertzel.org] Sent: Monday, December 29, 2008 2:30 PM To: agi@v2.listbox.com Subject: Re: [agi] Universal intelligence test benchmark Consciousness of X is: the idea or feeling that X is correlated with Consciousness of X ;-) ben g On Mon, Dec 29, 2008 at 4:23 PM, Matt Mahoney matmaho...@yahoo.com wrote: --- On Mon, 12/29/08, John G. Rose johnr...@polyplexic.com wrote: What does consciousness have to do with the rest of your argument? Multi-agent systems should need individual consciousness to achieve advanced levels of collective intelligence. So if you are programming a multi-agent system, potentially a compressor, having consciousness in the agents could have an intelligence amplifying effect instead of having non-conscious agents. Or some sort of primitive consciousness component since higher level consciousness has not really been programmed yet. Agree? No. What do you mean by consciousness? Some people use consciousness and intelligence interchangeably. If that is the case, then you are just using a circular argument. If not, then what is the difference? -- Matt Mahoney, matmaho...@yahoo.com agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
Re: [agi] Universal intelligence test benchmark
--- On Sun, 12/28/08, Philip Hunt cabala...@googlemail.com wrote: Please remember that I am not proposing compression as a solution to the AGI problem. I am proposing it as a measure of progress in an important component (prediction). Then why not cut out the middleman and measure prediction directly? Because a compressor proves the correctness of the measurement software at no additional cost in either space or time complexity or software complexity. The hard part of compression is modeling. Arithmetic coding is essentially a solved problem. A decompressor uses exactly the same model as a compressor. In high end compressors like PAQ, the arithmetic coder takes up about 1% of the software, 1% of the CPU time, and less than 1% of memory. In speech recognition research it is common to use word perplexity as a measure of the quality of a language model. Experimentally, it correlates well with word error rate. Perplexity is defined as 2^H where H is the average number of bits needed to encode a word. Unfortunately this is sometimes done in nonstandard ways, such as with restricted vocabularies and different methods of handling words outside the vocabulary, parsing, stemming, capitalization, punctuation, spacing, and numbers. Without accounting for this additional data, it makes published results difficult to compare. Compression removes the possibility of such ambiguities. -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
Re: [agi] Universal intelligence test benchmark
--- On Mon, 12/29/08, Philip Hunt cabala...@googlemail.com wrote: Incidently, reading Matt's posts got me interested in writing a compression program using Markov-chain prediction. The prediction bit was a piece of piss to write; the compression code is proving considerably more difficult. Well, there is plenty of open source software. http://cs.fit.edu/~mmahoney/compression/ If you want to write your own model and just need a simple arithmetic coder, you probably want fpaq0. Most of the other programs on this page use the same coder or some minor variation of it. -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
RE: [agi] Universal intelligence test benchmark
--- On Mon, 12/29/08, John G. Rose johnr...@polyplexic.com wrote: Well that's a question. Does death somehow enhance a lifeforms' collective intelligence? Yes, by weeding out the weak and stupid. Agents competing over finite resources.. I'm wondering if there were multi-agent evolutionary genetics going on would there be a finite resource of which there would be a relation to the collective goal of predicting the next symbol. No, prediction is a secondary goal. The primary goal is to have a lot of descendants. Agent knowledge is not only passed on in their genes, it is also passed around to other agents Does agent death hinder advances in intelligence or enhance it? And then would the intelligence collected thus be applicable to the goal. And if so, consciousness may be valuable. What does consciousness have to do with the rest of your argument? -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
RE: [agi] Universal intelligence test benchmark
--- On Mon, 12/29/08, John G. Rose johnr...@polyplexic.com wrote: What does consciousness have to do with the rest of your argument? Multi-agent systems should need individual consciousness to achieve advanced levels of collective intelligence. So if you are programming a multi-agent system, potentially a compressor, having consciousness in the agents could have an intelligence amplifying effect instead of having non-conscious agents. Or some sort of primitive consciousness component since higher level consciousness has not really been programmed yet. Agree? No. What do you mean by consciousness? Some people use consciousness and intelligence interchangeably. If that is the case, then you are just using a circular argument. If not, then what is the difference? -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
Re: [agi] [Science Daily] Our Unconscious Brain Makes The Best Decisions Possible
--- On Mon, 12/29/08, Richard Loosemore r...@lightlink.com wrote: 8-) Don't say that too loudly, Yudkowsky might hear you. :-) ... When I suggested that someone go check some of his ravings with an outside authority, he banned me from his discussion list. Yudkowsky's side of the story might be of interest... http://www.sl4.org/archive/0608/15895.html http://www.sl4.org/archive/0608/15928.html -- Matt Mahoney, matmaho...@yahoo.com From: Richard Loosemore r...@lightlink.com Subject: Re: [agi] [Science Daily] Our Unconscious Brain Makes The Best Decisions Possible To: agi@v2.listbox.com Date: Monday, December 29, 2008, 4:02 PM Lukasz Stafiniak wrote: http://www.sciencedaily.com/releases/2008/12/081224215542.htm Nothing surprising ;-) Nothing surprising?!! 8-) Don't say that too loudly, Yudkowsky might hear you. :-) The article is a bit naughty when it says, of Tversky and Kahnemann, that ...this has become conventional wisdom among cognition researchers. Actually, the original facts were interpreted in a variety of ways, some of which strongly disagreed with T K's original intepretation, just like this one you reference above. The only thing that is conventional wisdom is that the topic exists, and is the subject of dispute. And, as many people know, I made the mistake of challenging Yudkowsky on precisely this subject back in 2006, when he wrote an essay strongly advocating TK's original intepretation. Yudkowsky went completely berserk, accused me of being an idiot, having no brain, not reading any of the literature, never answering questions, and generally being something unspeakably worse than a slime-oozing crank. He literally wrote an essay denouncing me as equivalent to a flat-earth believing crackpot. When I suggested that someone go check some of his ravings with an outside authority, he banned me from his discussion list. Ah, such are the joys of being speaking truth to power(ful idiots). ;-) As far as this research goes, it sits somewhere down at the lower end of the available theories. My friend Mike Oaksford in the UK has written several papers giving a higher level cognitive theory that says that people are, in fact, doing something like bayesian estimation when then make judgments. In fact, people are very good at being bayesians, contra the loud protests of the I Am A Bayesian Rationalist crowd, who think they were the first to do it. Richard Loosemore --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
Re: [agi] Universal intelligence test benchmark
--- On Mon, 12/29/08, Philip Hunt cabala...@googlemail.com wrote: Am I right in understanding that the coder from fpaq0 could be used with any other predictor? Yes. It has a simple interface. You have a class called Predictor which is your bit sequence predictor. It has 2 member functions that you have to write. p() should return your estimated probability that the next bit will be a 1, as a 12 bit number (0 to 4095). update(y) then tells you what that bit actually was, a 0 or 1. The encoder will alternately call these 2 functions for each bit of the sequence. The predictor doesn't know whether it is compressing or decompressing because it sees exactly the same sequence either way. So the easy part is done :) -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
RE: [agi] Universal intelligence test benchmark
--- On Sun, 12/28/08, John G. Rose johnr...@polyplexic.com wrote: So maybe for improved genetic algorithms used for obtaining max compression there needs to be a consciousness component in the agents? Just an idea I think there is potential for distributed consciousness inside of command line compressors :) No, consciousness (as the term is commonly used) is the large set of properties of human mental processes that distinguish life from death, such as ability to think, learn, experience, make decisions, take actions, communicate, etc. It is only relevant as an independent concept to agents that have a concept of death and the goal of avoiding it. The only goal of a compressor is to predict the next input symbol. -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
RE: [agi] Universal intelligence test benchmark
--- On Sat, 12/27/08, John G. Rose johnr...@polyplexic.com wrote: How does consciousness fit into your compression intelligence modeling? It doesn't. Why is consciousness important? I was just prodding you on this. Many people on this list talk about the requirements of consciousness for AGI and I was imagining some sort of consciousness in one of your command line compressors :) I've yet to grasp the relationship between intelligence and consciousness though lately I think consciousness may be more of an evolutionary social thing. Home grown digital intelligence, since it is a loner, may not require much consciousness IMO.. What we commonly call consciousness is a large collection of features that distinguish living human brains from dead human brains: ability to think, communicate, perceive, make decisions, learn, move, talk, see, etc. We only attach significance to it because we evolved, like all animals, to fear a large set of things that can kill us. Max compression implies hacks, kludges and a large decompressor. As I discovered with the large text benchmark. Yep and the behavior of the metrics near max theoretical compression is erratic I think? It shouldn't be. There is a well defined (but possibly not computable) limit for each of the well defined universal Turing machines that the benchmark accepts (x86, C, C++, etc). I was hoping to discover an elegant theory for AI. It didn't quite work that way. It seems to be a kind of genetic algorithm: make random changes to the code and keep the ones that improve compression. -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
RE: [agi] Universal intelligence test benchmark
--- On Sat, 12/27/08, John G. Rose johnr...@polyplexic.com wrote: Well I think consciousness must be some sort of out of band intelligence that bolsters an entity in terms of survival. Intelligence probably stratifies or optimizes in zonal regions of similar environmental complexity, consciousness being one or an overriding out-of-band one... No, consciousness only seems mysterious because human brains are programmed that way. For example, I should logically be able to convince you that pain is just a signal that reduces the probability of you repeating whatever actions immediately preceded it. I can't do that because emotionally you are convinced that pain is real. Emotions can't be learned the way logical facts can, so emotions always win. If you could accept the logical consequences of your brain being just a computer, then you would not pass on your DNA. That's why you can't. BTW the best I can do is believe both that consciousness exists and consciousness does not exist. I realize these positions are inconsistent, and I leave it at that. I was hoping to discover an elegant theory for AI. It didn't quite work that way. It seems to be a kind of genetic algorithm: make random changes to the code and keep the ones that improve compression. Is this true for most data? For example would PI digit compression attempts result in genetic emergences the same as say compressing environmental noise? I'm just speculating that genetically originated data would require compression avenues of similar algorithmic complexity descriptors, for example PI digit data does not originate genetically so compression attempts would not show genetic emergences as chained as say environmental noise basically I'm asking if you can tell the difference from data that has a genetic origination ingredient verses all non-genetic... No, pi can be compressed to a simple program whose size is dominated by the log of the number of digits you want. For text, I suppose I should be satisfied that a genetic algorithm compresses it, except for the fact that so far the algorithm requires a human in the loop, so it doesn't solve the AI problem. -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
Re: [agi] Universal intelligence test benchmark
1 FC4A5294A52000 1 FC4A473E25239F1291C000 1 FC4A0941282504A09400 1 FC4A00 The best compressors will compress this data to just under 3 MB, which implies an average algorithmic complexity of less than 24 bits per string. However, the language allows the construction of arbitrary 128 bit strings in a fairly straightforward manner. -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
RE: [agi] Universal intelligence test benchmark
--- On Fri, 12/26/08, John G. Rose johnr...@polyplexic.com wrote: Human memory storage may be lossy compression and recall may be decompression. Some very rare individuals remember every day of their life in vivid detail, not sure what that means in terms of memory storage. Human perception is a form of lossy compression which has nothing to do with the lossless compression that I use to measure prediction accuracy. Many lossless compressors use lossy filters too. A simple example is an order-n context where we discard everything except the last n symbols. How does consciousness fit into your compression intelligence modeling? It doesn't. Why is consciousness important? Max compression implies hacks, kludges and a large decompressor. As I discovered with the large text benchmark. -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
Spatial indexing (was Re: [agi] Universal intelligence test benchmark)
--- On Fri, 12/26/08, J. Andrew Rogers and...@ceruleansystems.com wrote: For example, there is no general indexing algorithm described in computer science. Which was my thesis topic and is the basis of my AGI design. http://www.mattmahoney.net/agi2.html (I wanted to do my dissertation on AI/compression, but funding issues got in the way). Distributed indexing is critical to an AGI design consisting of a huge number of relatively dumb specialists and an infrastructure for getting messages to the right ones. In my thesis, I proposed a vector space model where messages are routed in O(n) time over n nodes. The problem is that the number of connections per node has to be on the order of the number of dimensions in the search space. For text, that is about 10^5. There are many other issues, of course, such as fault tolerance, security and ownership issues. There has to be an economic incentive to contribute knowledge and computing resources, because it is too expensive for anyone to own it. The human genome size has no meaningful relationship to the complexity of coding AGI. Yes it does. It is an upper bound on the complexity of a baby. -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
Re: [agi] Universal intelligence test benchmark
--- On Fri, 12/26/08, Ben Goertzel b...@goertzel.org wrote: IMO the test is *too* generic ... Hopefully this work will lead to general principles of learning and prediction that could be combined with more specific techniques. For example, a common way to compress text is to encode it with one symbol per word and feed the result to a general purpose compressor. Generic compression should improve the back end. My concern is the data is not generic enough. A string has an algorithmic complexity that is independent of language up to a small constant, but in practice that constant (the algorithmic complexity of the compiler) can be much larger than the string. I have not been able to find a good solution to this problem. I realize there are some very simple, Turing-complete systems, such as a 2 state machine with a 3 symbol alphabet, and a 6 state binary machine, as well as various cellular automata (like rule 110). The problem is that programming simple machines often requires long programs to do simple things. For example, it is difficult to find a simple language where the smallest program to output 100 zero bits is shorter than 100 bits. Existing languages and instruction sets tend to be complex and ad-hoc in order to allow programmers to be expressive. -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
Re: [agi] Universal intelligence test benchmark
--- On Fri, 12/26/08, Philip Hunt cabala...@googlemail.com wrote: Humans are very good at predicting sequences of symbols, e.g. the next word in a text stream. Why not have that as your problem domain, instead of text compression? That's the same thing, isn't it? While you're at it you may want to change the size of the chunks in each item of prediction, from characters to either strings or s-expressions. Though doing so doesn't fundamentally alter the problem. In the generic test, the fundamental units are bits. It's not entirely suitable for most existing compressors, which tend to be byte oriented. But they are only byte oriented because a lot of data is structured that way. In general, it doesn't need to be. -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
Re: Spatial indexing (was Re: [agi] Universal intelligence test benchmark)
--- On Sat, 12/27/08, Matt Mahoney matmaho...@yahoo.com wrote: In my thesis, I proposed a vector space model where messages are routed in O(n) time over n nodes. Oops, O(log n). -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
[agi] Universal intelligence test benchmark
I have been developing an experimental test set along the lines of Legg and Hutter's universal intelligence ( http://www.idsia.ch/idsiareport/IDSIA-04-05.pdf ). They define general intelligence as the expected reward of an AIXI agent in a Solomonoff distribution of environments (simulated by random Turing machines). AIXI is essentially a compression problem (find the shortest program consistent with the interaction so far). Thus, my benchmark is a large number (10^6) of small strings (1-32 bytes) generated by random Turing machines. The benchmark is here: http://cs.fit.edu/~mmahoney/compression/uiq/ I believe I have solved the technical issues related to experimental uncertainty and ensuring the source is cryptographically random. My goal was to make it an open benchmark with verifiable results while making it impossible to hard-code any knowledge of the test data into the agent. Other benchmarks solve this problem by including the decompressor size in the measurement, but my approach makes this unnecessary. However, I would appreciate any comments. A couple of issues arose in designing the benchmark. One is that compression results are highly dependent on the choice of universal Turing machine, even though all machines are theoretically equivalent. The problem is that even though any machine can simulate any other by appending a compiler or interpreter, this small constant is significant in practice where the complexity of the programs is already small. I tried to create a simple but expressive language based on a 2 tape machine (working plus output, both one sided and binary) and an instruction set that outputs a bit with each instruction. There are, of course, many options. I suppose I could use an experimental approach of finding languages that rank compressors in the same order as other benchmarks. But there doesn't seem to be a guiding principle. Also, it does not seem even possible to sample a Solomonoff distribution. Legg proved in http://arxiv.org/abs/cs.AI/0606070 that there are strings that are hard to learn, but that the time to create them grows as fast as the busy beaver problem. Of course I can't create such strings in my benchmark. I can create algorithmically complex sources, but they are necessarily easy to learn (for example, 100 random bits followed by all zero bits). Is it possible to test the intelligence of an agent without having at least as much computing power? Legg's paper seems to say no. -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
Re: FW: [agi] Lamarck Lives!(?)
--- On Thu, 12/11/08, Eric Burton brila...@gmail.com wrote: You can see though how genetic memory encoding opens the door to acquired phenotype changes over an organism's life, though, and those could become communicable. I think Lysenko was onto something like this. Let us hope all those Soviet farmers wouldn't have just starved! ;3 No, apparently you didn't understand anything I wrote. Please explain how the memory encoded separately as one bit each in 10^11 neurons through DNA methylation (the mechanism for cell differentiation, not genetic changes) is all collected together and encoded into genetic changes in a single egg or sperm cell, and back again to the brain when the organism matures. And please explain why you think that Lysenko's work should not have been discredited. http://en.wikipedia.org/wiki/Trofim_Lysenko -- Matt Mahoney, matmaho...@yahoo.com On 12/11/08, Matt Mahoney matmaho...@yahoo.com wrote: --- On Thu, 12/11/08, Eric Burton brila...@gmail.com wrote: It's all a big vindication for genetic memory, that's for certain. I was comfortable with the notion of certain templates, archetypes, being handed down as aspects of brain design via natural selection, but this really clears the way for organisms' life experiences to simply be copied in some form to their offspring. DNA form! No it's not. 1. There is no experimental evidence that learned memories are passed to offspring in humans or any other species. 2. If memory is encoded by DNA methylation as proposed in http://www.newscientist.com/article/mg20026845.000-memories-may-be-stored-on-your-dna.html then how is the memory encoded in 10^11 separate neurons (not to mention connectivity information) transferred to a single egg or sperm cell with less than 10^5 genes? The proposed mechanism is to activate one gene and turn off another -- 1 or 2 bits. 3. The article at http://www.technologyreview.com/biomedicine/21801/ says nothing about where memory is encoded, only that memory might be enhanced by manipulating neuron chemistry. There is nothing controversial here. It is well known that certain drugs affect learning. 4. The memory mechanism proposed in http://www.ncbi.nlm.nih.gov/pubmed/16822969?ordinalpos=14itool=EntrezSystem2.PEntrez.Pubmed.Pubmed_ResultsPanel.Pubmed_DefaultReportPanel.Pubmed_RVDocSum is distinct from (2). It proposes protein regulation at the mRNA level near synapses (consistent with the Hebbian model) rather than DNA in the nucleus. Such changes could not make their way back to the nucleus unless there was a mechanism to chemically distinguish the tens of thousands of synapses and encode this information, along with the connectivity information (about 10^6 bits per neuron) back to the nuclear DNA. Last week I showed how learning could occur in neurons rather than synapses in randomly and sparsely connected neural networks where all of the outputs of a neuron are constrained to have identical weights. The network is trained by tuning neurons toward excitation or inhibition to reduce the output error. In general an arbitrary X to Y bit binary function with N = Y 2^X bits of complexity can be learned using about 1.5N to 2N neurons with ~ N^1/2 synapses each and ~N log N training cycles. As an example I posted a program that learns a 3 by 3 bit multiplier in about 20 minutes on a PC using 640 neurons with 36 connections each. This is slower than Hebbian learning by a factor of O(N^1/2) on sequential computers, as well as being inefficient because sparse networks cannot be simulated efficiently using typical vector processing parallel hardware or memory optimized for sequential access. However this architecture is what we actually observe in neural tissue, which nevertheless does everything in parallel. The presence of neuron-centered learning does not preclude Hebbian learning occurring at the same time (perhaps at a different rate). However, the number of neurons (10^11) is much closer to Landauer's estimate of human long term memory capacity (10^9 bits) than the number of synapses (10^15). However, I don't mean to suggest that memory in either form can be inherited. There is no biological evidence for such a thing. -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
Re: FW: [agi] Lamarck Lives!(?)
--- On Thu, 12/11/08, Eric Burton brila...@gmail.com wrote: I don't think that each inheritor receives a full set of the original's memories. But there may have *evolved* in spite of the obvious barriers, a means of transferring primary or significant experience from one organism to another in genetic form... we can imagine such a thing given this news! Well, we could, if there was any evidence whatsoever for Lamarckian evolution, and if we thought with our reproductive organs. To me, it suggests that AGI could be implemented with a 10^4 speedup over whole brain emulation -- maybe. Is it possible to emulate a sparse neural network with 10^11 adjustable neurons and 10^15 fixed, random connections using a non-sparse neural network with 10^11 adjustable connections? -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
Re: FW: [agi] Lamarck Lives!(?)
--- On Thu, 12/11/08, Eric Burton brila...@gmail.com wrote: I don't know how you derived the value 10^4, Matt, but that seems reasonable to me. Terren, let me go back to the article and try to understand what exactly it says is happening. Certainly that's my editorial's crux A simulation of a neural network with 10^15 synapses requires 10^15 operations to update the activation levels of the neurons. If we assume 100 ms resolution, that is 10^16 operations per second. If memory is stored in neurons rather than synapses, as suggested in the original paper (see http://www.cell.com/neuron/retrieve/pii/S0896627307001420 ) then the brain has a memory capacity of at most 10^11 bits, which could be simulated by a neural network with 10^11 connections (or 10^12 operations per second). This assumes that (1) the networks are equivalent and (2) that there isn't any secondary storage in synapses in addition to neurons. The program I posted last week was intended to show (1). However (2) has not been shown. The fact that DNA methylation occurs in the cortex does not exclude the possibility of more than one memory mechanism. As a counter argument, the cortex has about 10^4 times as much storage as the hippocampus (10^4 days vs. 1 day), but is not 10^4 times larger. -- Matt Mahoney, matmaho...@yahoo.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=123753653-47f84b Powered by Listbox: http://www.listbox.com
Re: [agi] Machine Knowledge and Inverse Machine Knowledge...
Steve, the difference between Cyc and Dr. Eliza is that Cyc has much more knowledge. Cyc has millions of rules. The OpenCyc download is hundreds of MB compressed. Several months ago you posted the database file for Dr. Eliza. I recall it was a few hundred rules and I think under 1 MB. Both of these databases are far too small for AGI because neither has solved the learning problem. -- Matt Mahoney, [EMAIL PROTECTED] From: Steve Richfield [EMAIL PROTECTED] To: agi@v2.listbox.com Sent: Tuesday, December 9, 2008 3:06:08 AM Subject: [agi] Machine Knowledge and Inverse Machine Knowledge... Larry Lefkowitz, Stephen Reed, et al, First, Thanks Steve for your pointer to Larry Lefkowitz, and thanks Larry for so much time and effort in trying to relate our two approaches.. After discussions with Larry Lefkowitz of Cycorp, I have had a bit of an epiphany regarding machine knowledge that I would like to share for all to comment on... First, it wasn't as though there were points of incompatibility between Cycorp's idea of machine knowledge and that used in DrEliza.com, but rather, there were no apparent points of connection. How could two related things be so completely different, especially when both are driven by the real world? Then it struck me. Cycorp and others here on this forum seek to represent the structures of real world domains in a machine, whereas Dr. Eliza seeks only to represent the structure of the malfunctions within structures, while making no attempt whatever to represent the structures in which those malfunctions occur, as though those malfunctions have their very own structure, as they truly do. This seems a bit like simulating the holes in a semiconductor. OF COURSE there were no points of connection. Larry pointed out the limitations in my approach - which I already knew, namely, Dr. Eliza will NEVER EVER understand normal operation when all it has to go on are ABnormalities. Similarly, I pointed out that Cycorp's approach had the inverse problem, in that it would probably take the quadrillion dollars that Matt Mahoney keeps talking about to ever understand malfunctions starting from the wrong side (as seen from Dr. Eliza's viewpoint) of things. In short, I see both of these as being quite valid but completely incompatible approaches, that accomplish very different things via very different methods. Each could move toward the other's capabilities given infinite resources, but only a madman (like Matt Mahoney?) would ever throw money at such folly. Back to my reason for contacting Cycorp - to see if some sort of web standard to represent metadata could be hammered out. Neither Larry nor I could see how Dr. Eliza's approach could be adapted to Cycorp, and further, this is aside from Cycorp's present interests. Hence, I am on my own here. Hence, it is my present viewpoint that I should proceed with my present standard to accompany the only semi-commercial program that models malfunctions rather than the real world, somewhat akin to the original Eliza program. However, I should prominently label the standard and appropriate fields therein appropriately so that there is no future confusion between machine knowledge and Dr. Eliza's sort of inverse machine knowledge. Any thoughts? Steve Richfield agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=120640061-aded06 Powered by Listbox: http://www.listbox.com
