Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
Sorry about the late reply. snip some stuff sorted out 2008/6/30 Vladimir Nesov [EMAIL PROTECTED]: On Tue, Jul 1, 2008 at 2:02 AM, William Pearson [EMAIL PROTECTED] wrote: 2008/6/30 Vladimir Nesov [EMAIL PROTECTED]: If internals are programmed by humans, why do you need automatic system to assess them? It would be useful if you needed to construct and test some kind of combination/setting automatically, but not if you just test manually-programmed systems. How does the assessment platform help in improving/accelerating the research? Because to be interesting the human specified programs need to be autogenous, as in Josh Storr Hall's terminology, which means self-building. Capable of altering the stuff they are made of. In this case machine code equivalent. So you need the human to assess the improvements the system makes, for whatever purpose the human wants the system to perform. Altering the stuff they are made of is instrumental to achieving the goal, and should be performed where necessary, but it doesn't happen, for example, with individual brains. I think it happens at the level of neural structures. I.e. I think neural structures control the development of other neural structures. (I was planning to do the next blog post on this theme, maybe tomorrow.) Do you mean to create population of altered initial designs and somehow select from them (I hope not, it is orthogonal to what modification is for in the first place)? Otherwise, why do you still need automated testing? Could you present a more detailed use case? I'll try and give a fuller explanation later on. This means he needs to use a bunch more resources to get a singular useful system. Also the system might not do what he wants, but I don't think he minds about that. I'm allowing humans to design everything, just allowing the very low level to vary. Is this clearer? What do you mean by varying low level, especially in human-designed systems? The machine code the program is written in. Or in a java VM, the java bytecode. This still didn't make this point clearer. You can't vary the semantics of low-level elements from which software is built, and if you don't modify the semantics, any other modification is superficial and irrelevant. If it's not quite 'software' that you are running, and it is able to survive the modification of lower level, using the terms like 'machine code' and 'software' is misleading. And in any case, it's not clear what this modification of low level achieves. You can't extract work from obfuscation and tinkering, the optimization comes from the lawful and consistent pressure in the same direction. Okay let us clear things up. There are two things that need to be designed, a computer architecture or virtual machine and programs that form the initial set of programs within the system. Let us call the internal programs vmprograms to avoid confusion.The vmprograms should do all the heavy lifting (reasoning, creating new programs), this is where the lawlful and consistent pressure would come from. It is at source code of vmprograms that all needs to be changeable. However the pressure will have to be somewhat experimental to be powerful, you don't know what bugs a new program will have (if you are doing a non-tight proof search through the space of programs). So the point of the VM is to provide a safety net. If an experiment goes awry, then the VM should allow each program to limit the bugged vmprograms ability to affect it and eventually have it removed and the resources applied to it. Here is a toy scenario where the system needs this ability. *Note it is not anything that is like a full AI but illustrates a facet of something a full AI needs IMO*. Consider a system trying to solve a task, e.g. navigate a maze, that also has a number of different people out there giving helpful hints on how to solve the maze. These hints are in the form of patches to the vmprograms, e.g. changing the representation to 6-dimensional, giving another patch language that has better patches. So the system would make copies of the part of it to be patched and then patch it. Now you could give a patch evaluation module to see which patch works best, but what would happen if the module that implemented that vmprogram wanted to be patched? My solution to the problem is to allow the patch and non-patched version compete in the adhoc economic arena, and see which one wins. Does this clear things up? Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
Terren, This is going too far. We can reconstruct to a considerable extent how humans think about problems - their conscious thoughts. Artists have been doing this reasonably well for hundreds of years. Science has so far avoided this, just as it avoided studying first the mind, with behaviourism, then consciousness,. The main reason cognitive science and psychology have avoided stream-of-thought studies (apart from v. odd scientists like Jerome Singer) is that conscious thought about problems is v. different from the highly ordered, rational, thinking of programmed computers which cog. sci. uses as its basic paradigm. In fact, human thinking is fundamentally different - the conscious self has major difficulty concentrating on any problem for any length of time - controlling the mind for more than a relatively few seconds, (as religious and humanistic thinkers have been telling us for thousands of years). Computers of course have perfect concentration forever. But that's because computers haven't had to deal with the type of problems that we do - the problematic problems where you don't, basically, know the answer, or how to find the answer, before you start. For this kind of problem - which is actually what differentiates AGI from narrow AI - human thinking, creative as opposed to rational, stumbling, scatty, and freely associative, is actually IDEAL, for all its imperfections. Yes, even if we extend our model of intelligence to include creative as well as rational thinking, it will still be an impoverished model, which may not include embodied thinking and perhaps other dimensions. But hey, we'll get there bit by bit, (just not, as we both agree, all at once in one five-year leap). Terren: My points about the pitfalls of theorizing about intelligence apply to any and all humans who would attempt it - meaning, it's not necessary to characterize AI folks in one way or another. There are any number of aspects of intelligence we could highlight that pose a challenge to orthodox models of intelligence, but the bigger point is that there are fundamental limits to the ability of an intelligence to observe itself, in exactly the same way that an eye cannot see itself. Consciousness and intelligence are present in every possible act of contemplation, so it is impossible to gain a vantage point of intelligence from outside of it. And that's exactly what we pretend to do when we conceptualize it within an artificial construct. This is the principle conceit of AI, that we can understand intelligence in an objective way, and model it well enough to reproduce by design. Terren --- On Tue, 7/1/08, Mike Tintner [EMAIL PROTECTED] wrote: Terren:It's to make the larger point that we may be so immersed in our own conceptualizations of intelligence - particularly because we live in our models and draw on our own experience and introspection to elaborate them - that we may have tunnel vision about the possibilities for better or different models. Or, we may take for granted huge swaths of what makes us so smart, because it's so familiar, or below the radar of our conscious awareness, that it doesn't even occur to us to reflect on it. No 2 is more relevant - AI-ers don't seem to introspect much. It's an irony that the way AI-ers think when creating a program bears v. little resemblance to the way programmed computers think. (Matt started to broach this when he talked a while back of computer programming as an art). But AI-ers seem to have no interest in the discrepancy - which again is ironic, because analysing it would surely help them with their programming as well as the small matter of understanding how general intelligence actually works. In fact - I just looked - there is a longstanding field on psychology of programming. But it seems to share the deficiency of psychology and cognitive science generally which is : no study of the stream-of-conscious-thought, especially conscious problemsolving. The only AI figure I know who did take some interest here was Herbert Simon who helped establish the use of verbal protocols. --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
On Wed, Jul 2, 2008 at 2:48 PM, William Pearson [EMAIL PROTECTED] wrote: Okay let us clear things up. There are two things that need to be designed, a computer architecture or virtual machine and programs that form the initial set of programs within the system. Let us call the internal programs vmprograms to avoid confusion.The vmprograms should do all the heavy lifting (reasoning, creating new programs), this is where the lawlful and consistent pressure would come from. It is at source code of vmprograms that all needs to be changeable. However the pressure will have to be somewhat experimental to be powerful, you don't know what bugs a new program will have (if you are doing a non-tight proof search through the space of programs). So the point of the VM is to provide a safety net. If an experiment goes awry, then the VM should allow each program to limit the bugged vmprograms ability to affect it and eventually have it removed and the resources applied to it. Here is a toy scenario where the system needs this ability. *Note it is not anything that is like a full AI but illustrates a facet of something a full AI needs IMO*. Consider a system trying to solve a task, e.g. navigate a maze, that also has a number of different people out there giving helpful hints on how to solve the maze. These hints are in the form of patches to the vmprograms, e.g. changing the representation to 6-dimensional, giving another patch language that has better patches. So the system would make copies of the part of it to be patched and then patch it. Now you could give a patch evaluation module to see which patch works best, but what would happen if the module that implemented that vmprogram wanted to be patched? My solution to the problem is to allow the patch and non-patched version compete in the adhoc economic arena, and see which one wins. What are the criteria that VM applies to vmprograms? If VM just shortcircuits the economic pressure of agents to one another, it in itself doesn't specify the direction of the search. The human economy works to efficiently satisfy the goals of human beings who already have their moral complexity. It propagates the decisions that customers make, and fuels the allocation of resources based on these decisions. Efficiency of economy is in efficiency of responding to information about human goals. If your VM just feeds the decisions on themselves, what stops the economy from focusing on efficiently doing nothing? -- Vladimir Nesov [EMAIL PROTECTED] http://causalityrelay.wordpress.com/ --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] Simple example of the complex systems problem, for those in a hurry
John G. Rose wrote: [snip] Building a complex based intelligence much different from the human brain design but still basically dependant on complexity is not impossible just formidable. Working with software systems that have designed complexity and getting predicted emergence and in this case cognition, well that is something that takes special talent. We have tools now that nature and evolution didn't have. We understand things through collective knowledge accumulated over time. It can be more than trial and error. And the existing trial and error can be narrowed down. Ah, but now you are stating the Standard Reply, and what you have to understand is that the Standard Reply boils down to this: We are so smart that we will figure a way around this limitation, without having to do any so crass as just copying the human design. The problem is that if you apply that logic to well-known cases of complex systems, it amounts to nothing more than baseless, stubborn optimism in the face of any intractable problem. It is this baseless stubborn optimism that I am trying to bring to everyone's attention. In all my efforts to get this issue onto people's mental agenda, my goal is to make them realize that they would NEVER say such a silly thing about the vast majority of complex systems (nobody has any idea how to build an analytical theory of the relationship between the patterns that emerge in Game Of Life, for example, and that is one of the most trivial examples of a complex system that I can think of!). But whereas most mathematicians would refuse to waste any time at all trying to make a global-to-local theory for complex systems in which there is really vicious self-organisation at work, AI researchers blithely walk in and say We reckon we can just use our smarts and figure out some heuristics to get around it. I'm just trying to get people to do a reality check. Oh, and meanwhile (when I am not firing off occasional broadsides on this list) I *am* working on a solution. Richard Loosemore --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] the uncomputable
So yes, I think there are perfectly fine, rather simple definitions for computing machines that can (it seems like) perform calculations that turing machines cannot. It should really be noted that quantum computers fall into this class. This is very interesting. Previously, I had heard (but not from a definitive source) that quantum computers could compute in principle only what a Turing machine could compute, but could do it much more efficiently (something like the square root of the effort a Turing machine would need, at least for some tasks). Can you cite any source on this? But I should emphasize that what I am really interested in is computable approximation of uncomputable things. My stance is that an AGI should be able to reason about uncomputable concepts in a coherent manner (like we can), not that it needs to be able to actually compute them (which we can't). On Tue, Jul 1, 2008 at 2:35 AM, Linas Vepstas [EMAIL PROTECTED] wrote: 2008/6/16 Abram Demski [EMAIL PROTECTED]: I previously posted here claiming that the human mind (and therefore an ideal AGI) entertains uncomputable models, counter to the AIXI/Solomonoff model. There was little enthusiasm about this idea. :) I missed your earlier posts. However, I believe that there are models of computation can compute things that turing machines cannot, and that this is not arcane, just not widely known or studied. Here is a quick sketch: Topological finite automata, or geometric finite automata, (of which the quantum finite automata is a special case) generalize the notion of non-deterministic finite automata by replacing its powerset of states with a general topological or geometric space (complex projective space in the quantum case). It is important to note that these general spaces are in general uncountable (have the cardinality of the continuum). It is well known that the languages accepted by quantum finite automata are not regular languages, they are bigger and more complex in some ways. I am not sure what is known about the languages accepted by quantum push-down automata, but intuitively these are clearly different (and bigger) than the class of context-free languages. I believe the concepts of topological finite automata extend just fine to a generalization of turing machines, but I also believe this is a poorly-explored area of mathematics. I beleive such machines can compute things that turing machiens can't .. this should not be a surprise, since, after all, these systems have, in general, an uncountably infinite number of internal states (cardinality of the continuum!), and (as a side effect of the definition), perform infinite-precision addition and multiplication in finite time. So yes, I think there are perfectly fine, rather simple definitions for computing machines that can (it seems like) perform calculations that turing machines cannot. It should really be noted that quantum computers fall into this class. Considerably more confusing is the relationship of such machines (and the languages they accept) to lambda calculus, or first-order (or higher-order) logic. This is where the rubber hits the road, and even for the simplest examples, the systems are poorly understood, or not even studied. So, yeah, I think there's plenty of room for the uncomputable in some rather simple mathematical models of generalized computation. --linas --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
Mike, This is going too far. We can reconstruct to a considerable extent how humans think about problems - their conscious thoughts. Why is it going too far? I agree with you that we can reconstruct thinking, to a point. I notice you didn't say we can completely reconstruct how humans think about problems. Why not? We have two primary means for understanding thought, and both are deeply flawed: 1. Introspection. Introspection allows us to analyze our mental life in a reflective way. This is possible because we are able to construct mental models of our mental models. There are three flaws with introspection. The first, least serious flaw is that we only have access to that which is present in our conscious awareness. We cannot introspect about unconscious processes, by definition. This is a less serious objection because it's possible in practice to become conscious of phenomena there were previously unconscious, by developing our meta-mental-models. The question here becomes, is there any reason in principle that we cannot become conscious of *all* mental processes? The second flaw is that, because introspection relies on the meta-models we need to make sense of our internal, mental life, the possibility is always present that our meta-models themselves are flawed. Worse, we have no way of knowing if they are wrong, because we often unconsciously, unwittingly deny evidence contrary to our conception of our own cognition, particularly when it runs counter to a positive account of our self-image. Harvard's Project Implicit experiment (https://implicit.harvard.edu/implicit/) is a great way to demonstrate how we remain ignorant of deep, unconscious biases. Another example is how little we understand the contribution of emotion to our decision-making. Joseph Ledoux and others have shown fairly convincingly that emotion is a crucial part of human cognition, but most of us (particularly us men) deny the influence of emotion on our decision making. The final flaw is the most serious. It says there is a fundamental limit to what introspection has access to. This is the an eye cannot see itself objection. But I can see my eyes in the mirror, says the devil's advocate. Of course, a mirror lets us observe a reflected version of our eye, and this is what introspection is. But we cannot see inside our own eye, directly - it's a fundamental limitation of any observational apparatus. Likewise, we cannot see inside the very act of model-simulation that enables introspection. Introspection relies on meta-models, or models about models, which are activated/simulated *after the fact*. We might observe ourselves in the act of introspection, but that is nothing but a meta-meta-model. Each introspectional act by necessity is one step (at least) removed from the direct, in-the-present flow of cognition. This means that we can never observe the cognitive machinery that enables the act of introspection itself. And if you don't believe that introspection relies on cognitive machinery (maybe you're a dualist, but then why are you on an AI list? :-), ask yourself why we can't introspect about ourselves before a certain point in our young lives. It relies on a sufficiently sophisticated toolset that requires a certain amount of development before it is even possible. 2. Theory. Our theories of cognition are another path to understanding, and much of theory is directly or indirectly informed by introspection. When introspection fails (as in language acquisition), we rely completely on theory. The flaw with theory should be obvious. We have no direct way of testing theories of cognition, since we don't understand the connection between the mental and the physical. At best, we can use clever indirect means for generating evidence, and we usually have to accept the limits of reliability of subjective reports. Terren --- On Wed, 7/2/08, Mike Tintner [EMAIL PROTECTED] wrote: Terren, This is going too far. We can reconstruct to a considerable extent how humans think about problems - their conscious thoughts. Artists have been doing this reasonably well for hundreds of years. Science has so far avoided this, just as it avoided studying first the mind, with behaviourism, then consciousness,. The main reason cognitive science and psychology have avoided stream-of-thought studies (apart from v. odd scientists like Jerome Singer) is that conscious thought about problems is v. different from the highly ordered, rational, thinking of programmed computers which cog. sci. uses as its basic paradigm. In fact, human thinking is fundamentally different - the conscious self has major difficulty concentrating on any problem for any length of time - controlling the mind for more than a relatively few seconds, (as religious and humanistic thinkers have been telling us for thousands of years). Computers of course have perfect concentration forever.
RE: [agi] Simple example of the complex systems problem, for those in a hurry
From: Richard Loosemore [mailto:[EMAIL PROTECTED] Ah, but now you are stating the Standard Reply, and what you have to understand is that the Standard Reply boils down to this: We are so smart that we will figure a way around this limitation, without having to do any so crass as just copying the human design. Well another reply could be - OK everyone AGI is impossible so you can go home now. That would work real well. Into the future more and more bodies(and brains) will be thrown at this no matter what. Satellite technologies make it all more attractive and worthwhile and make it appear that progress is being made, and it is. If everything else is figured out and engineered and the last thing is a CSP that is still progress EVEN if some of the components need to be totally redesigned. Remember even basic stuff like say a primitive distributed graph software library is still in early stages of being built for AGI amongst many other things. There are protocols, standards, all kinds of stuff needed yet not there, especially experience. The problem is that if you apply that logic to well-known cases of complex systems, it amounts to nothing more than baseless, stubborn optimism in the face of any intractable problem. It is this baseless stubborn optimism that I am trying to bring to everyone's attention. Sure. Yet how many resources are thrown at predicting the weather and it is usually still WRONG!! The utility of accurate prediction is so high even useless attempts have value due to spin-off technologies and incidentals and there is psychological value.. In all my efforts to get this issue onto people's mental agenda, my goal is to make them realize that they would NEVER say such a silly thing about the vast majority of complex systems (nobody has any idea how to build an analytical theory of the relationship between the patterns that emerge in Game Of Life, for example, and that is one of the most trivial examples of a complex system that I can think of!). But whereas most mathematicians would refuse to waste any time at all trying to make a global-to-local theory for complex systems in which there is really vicious self-organisation at work, AI researchers blithely walk in and say We reckon we can just use our smarts and figure out some heuristics to get around it. That's what makes engineers engineers. If it is not conquerable it is workaroundable. Still though I don't know how much proof that there is a CSP. The CB example you gave reminds me of a dynamical system. Proving the CSP exists may turn heads more. I'm just trying to get people to do a reality check. Oh, and meanwhile (when I am not firing off occasional broadsides on this list) I *am* working on a solution. Yes, and your solution attempt is :) Please feel free to present ideas to the list for constructive criticism :) John --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
Terren, Obviously, as I indicated, I'm not suggesting that we can easily construct a total model of human cognition. But it ain't that hard to reconstruct reasonable and highly informative, if imperfect, models of how humans consciously think about problems. As I said, artists have been doing a reasonable job for centuries. Shakespeare, who really started the inner monologue, was arguably the first scientist of consciousness. The kind of standard argument you give below - the eye can't look at itself - is actually nonsense. Your conscious, inner thoughts are not that different from your public, recordable dialogue. (Any decent transcript of thought, BTW, will give a v. good indication of the emotions involved). We're not v. far apart here - we agree about the many dimensions of cognition, most of which are probably NOT directly accessible to the conscious mind. I'm just insisting on the massive importance of studying conscious thought. It was, as Crick said, ridiculous for science not to study consciousness - (it had a lot of rubbish arguments for not doing that, then) - it is equally ridiculous and in fact scientifically obscene not to study conscious thought. The consequences both for humans generally and AGI are enormous. Terren: Mike, This is going too far. We can reconstruct to a considerable extent how humans think about problems - their conscious thoughts. Why is it going too far? I agree with you that we can reconstruct thinking, to a point. I notice you didn't say we can completely reconstruct how humans think about problems. Why not? We have two primary means for understanding thought, and both are deeply flawed: 1. Introspection. Introspection allows us to analyze our mental life in a reflective way. This is possible because we are able to construct mental models of our mental models. There are three flaws with introspection. The first, least serious flaw is that we only have access to that which is present in our conscious awareness. We cannot introspect about unconscious processes, by definition. This is a less serious objection because it's possible in practice to become conscious of phenomena there were previously unconscious, by developing our meta-mental-models. The question here becomes, is there any reason in principle that we cannot become conscious of *all* mental processes? The second flaw is that, because introspection relies on the meta-models we need to make sense of our internal, mental life, the possibility is always present that our meta-models themselves are flawed. Worse, we have no way of knowing if they are wrong, because we often unconsciously, unwittingly deny evidence contrary to our conception of our own cognition, particularly when it runs counter to a positive account of our self-image. Harvard's Project Implicit experiment (https://implicit.harvard.edu/implicit/) is a great way to demonstrate how we remain ignorant of deep, unconscious biases. Another example is how little we understand the contribution of emotion to our decision-making. Joseph Ledoux and others have shown fairly convincingly that emotion is a crucial part of human cognition, but most of us (particularly us men) deny the influence of emotion on our decision making. The final flaw is the most serious. It says there is a fundamental limit to what introspection has access to. This is the an eye cannot see itself objection. But I can see my eyes in the mirror, says the devil's advocate. Of course, a mirror lets us observe a reflected version of our eye, and this is what introspection is. But we cannot see inside our own eye, directly - it's a fundamental limitation of any observational apparatus. Likewise, we cannot see inside the very act of model-simulation that enables introspection. Introspection relies on meta-models, or models about models, which are activated/simulated *after the fact*. We might observe ourselves in the act of introspection, but that is nothing but a meta-meta-model. Each introspectional act by necessity is one step (at least) removed from the direct, in-the-present flow of cognition. This means that we can never observe the cognitive machinery that enables the act of introspection itself. And if you don't believe that introspection relies on cognitive machinery (maybe you're a dualist, but then why are you on an AI list? :-), ask yourself why we can't introspect about ourselves before a certain point in our young lives. It relies on a sufficiently sophisticated toolset that requires a certain amount of development before it is even possible. 2. Theory. Our theories of cognition are another path to understanding, and much of theory is directly or indirectly informed by introspection. When introspection fails (as in language acquisition), we rely completely on theory. The flaw with theory should be obvious. We have no direct way of testing theories of cognition, since we don't understand the
Re: [agi] the uncomputable
The standard model of quantum computation as defined by Feynman and Deutsch is Turing computable (based on the concept of qubits). As proven by Deutsch they compute the same set of functions than Turing machines but faster (if they are feasible). Non-standard models of quantum computation are not widely accepted, and even when they could hypercompute many doubt that we could take any from continuum entangling to perform computations. Non-standard quantum computers have not yet being well defined (and that is one of the many issues of hypercomputation: each time one comes up with a standard model of hypercomputation there is always another not equivalent model of hypercomputation that computes a different set of functions, i.e. there is no convergence in models unlike what happened when digital computation was characterized). Hypercomputational models basically pretend to take advantage from either infinite time or infinite space (including models such as infinite resources, Zeno machines or the Omega-rule, real computation, etc.), from the continuum. Depending of the density of that space/time continuum one can think of several models taking advantage at several levels of the arithmetical hierarchy. But even if there is infinite space or time another issue is how to verify a hypercomputation. One would need another hypercomputer to verify the first and then trust in one. Whether you think hypercomputation, the following paper is a most read for those interested on the topic. Martin Davis' articulates several criticisms: The myth of hypercomputation, in: C. Teuscher (Ed.), Alan Turing: Life and Legacy of a Great Thinker (2004) Serious work on analogous computation can be found in papers from Felix Costa et al.: http://fgc.math.ist.utl.pt/jfc.htm My master's thesis was on the subject so if you are interested in getting an electronic copy just let me know. It is in French though. On Wed, Jul 2, 2008 at 11:15 AM, Abram Demski [EMAIL PROTECTED] wrote: So yes, I think there are perfectly fine, rather simple definitions for computing machines that can (it seems like) perform calculations that turing machines cannot. It should really be noted that quantum computers fall into this class. This is very interesting. Previously, I had heard (but not from a definitive source) that quantum computers could compute in principle only what a Turing machine could compute, but could do it much more efficiently (something like the square root of the effort a Turing machine would need, at least for some tasks). Can you cite any source on this? But I should emphasize that what I am really interested in is computable approximation of uncomputable things. My stance is that an AGI should be able to reason about uncomputable concepts in a coherent manner (like we can), not that it needs to be able to actually compute them (which we can't). On Tue, Jul 1, 2008 at 2:35 AM, Linas Vepstas [EMAIL PROTECTED] wrote: 2008/6/16 Abram Demski [EMAIL PROTECTED]: I previously posted here claiming that the human mind (and therefore an ideal AGI) entertains uncomputable models, counter to the AIXI/Solomonoff model. There was little enthusiasm about this idea. :) I missed your earlier posts. However, I believe that there are models of computation can compute things that turing machines cannot, and that this is not arcane, just not widely known or studied. Here is a quick sketch: Topological finite automata, or geometric finite automata, (of which the quantum finite automata is a special case) generalize the notion of non-deterministic finite automata by replacing its powerset of states with a general topological or geometric space (complex projective space in the quantum case). It is important to note that these general spaces are in general uncountable (have the cardinality of the continuum). It is well known that the languages accepted by quantum finite automata are not regular languages, they are bigger and more complex in some ways. I am not sure what is known about the languages accepted by quantum push-down automata, but intuitively these are clearly different (and bigger) than the class of context-free languages. I believe the concepts of topological finite automata extend just fine to a generalization of turing machines, but I also believe this is a poorly-explored area of mathematics. I beleive such machines can compute things that turing machiens can't .. this should not be a surprise, since, after all, these systems have, in general, an uncountably infinite number of internal states (cardinality of the continuum!), and (as a side effect of the definition), perform infinite-precision addition and multiplication in finite time. So yes, I think there are perfectly fine, rather simple definitions for computing machines that can (it seems like) perform calculations that turing machines cannot. It should really be noted that quantum computers fall into this class.
Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
2008/7/2 Terren Suydam [EMAIL PROTECTED]: Mike, This is going too far. We can reconstruct to a considerable extent how humans think about problems - their conscious thoughts. Why is it going too far? I agree with you that we can reconstruct thinking, to a point. I notice you didn't say we can completely reconstruct how humans think about problems. Why not? We have two primary means for understanding thought, and both are deeply flawed: 1. Introspection. Introspection allows us to analyze our mental life in a reflective way. This is possible because we are able to construct mental models of our mental models. There are three flaws with introspection. The first, least serious flaw is that we only have access to that which is present in our conscious awareness. We cannot introspect about unconscious processes, by definition. This is a less serious objection because it's possible in practice to become conscious of phenomena there were previously unconscious, by developing our meta-mental-models. The question here becomes, is there any reason in principle that we cannot become conscious of *all* mental processes? The second flaw is that, because introspection relies on the meta-models we need to make sense of our internal, mental life, the possibility is always present that our meta-models themselves are flawed. Worse, we have no way of knowing if they are wrong, because we often unconsciously, unwittingly deny evidence contrary to our conception of our own cognition, particularly when it runs counter to a positive account of our self-image. Harvard's Project Implicit experiment (https://implicit.harvard.edu/implicit/) is a great way to demonstrate how we remain ignorant of deep, unconscious biases. Another example is how little we understand the contribution of emotion to our decision-making. Joseph Ledoux and others have shown fairly convincingly that emotion is a crucial part of human cognition, but most of us (particularly us men) deny the influence of emotion on our decision making. The final flaw is the most serious. It says there is a fundamental limit to what introspection has access to. This is the an eye cannot see itself objection. But I can see my eyes in the mirror, says the devil's advocate. Of course, a mirror lets us observe a reflected version of our eye, and this is what introspection is. But we cannot see inside our own eye, directly - it's a fundamental limitation of any observational apparatus. Likewise, we cannot see inside the very act of model-simulation that enables introspection. Introspection relies on meta-models, or models about models, which are activated/simulated *after the fact*. We might observe ourselves in the act of introspection, but that is nothing but a meta-meta-model. Each introspectional act by necessity is one step (at least) removed from the direct, in-the-present flow of cognition. This means that we can never observe the cognitive machinery that enables the act of introspection itself. And if you don't believe that introspection relies on cognitive machinery (maybe you're a dualist, but then why are you on an AI list? :-), ask yourself why we can't introspect about ourselves before a certain point in our young lives. It relies on a sufficiently sophisticated toolset that requires a certain amount of development before it is even possible. 2. Theory. Our theories of cognition are another path to understanding, and much of theory is directly or indirectly informed by introspection. When introspection fails (as in language acquisition), we rely completely on theory. The flaw with theory should be obvious. We have no direct way of testing theories of cognition, since we don't understand the connection between the mental and the physical. At best, we can use clever indirect means for generating evidence, and we usually have to accept the limits of reliability of subjective reports. My plan is go for 3) Usefulness. Cognition is useful from an evolutionary point of view, if we try to create systems that are useful in the same situations (social, building world models), then we might one day stumble upon cognition. To expand on usefulness in social contexts, you have to ask yourself what the point of language is, why is it useful in an evolutionary setting. One thing the point of language is not, is fooling humans that you are human, which makes me annoyed at all the chatbots that get coverage as AI. I'll write more on this later. This by the way is why I don't self-organise purpose. I am pretty sure a specified purpose (not the same thing as a goal, at all) is needed for an intelligence. Will --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription:
Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
Mike, That's a rather weak reply. I'm open to the possibility that my ideas are incorrect or need improvement, but calling what I said nonsense without further justification is just hand waving. Unless you mean this as your justification: Your conscious, inner thoughts are not that different from your public, recordable dialogue. How this amounts to an objection to my points about introspection is beyond me... care to elaborate? Terren --- On Wed, 7/2/08, Mike Tintner [EMAIL PROTECTED] wrote: Terren, Obviously, as I indicated, I'm not suggesting that we can easily construct a total model of human cognition. But it ain't that hard to reconstruct reasonable and highly informative, if imperfect, models of how humans consciously think about problems. As I said, artists have been doing a reasonable job for centuries. Shakespeare, who really started the inner monologue, was arguably the first scientist of consciousness. The kind of standard argument you give below - the eye can't look at itself - is actually nonsense. Your conscious, inner thoughts are not that different from your public, recordable dialogue. (Any decent transcript of thought, BTW, will give a v. good indication of the emotions involved). We're not v. far apart here - we agree about the many dimensions of cognition, most of which are probably NOT directly accessible to the conscious mind. I'm just insisting on the massive importance of studying conscious thought. It was, as Crick said, ridiculous for science not to study consciousness - (it had a lot of rubbish arguments for not doing that, then) - it is equally ridiculous and in fact scientifically obscene not to study conscious thought. The consequences both for humans generally and AGI are enormous. Terren: Mike, This is going too far. We can reconstruct to a considerable extent how humans think about problems - their conscious thoughts. Why is it going too far? I agree with you that we can reconstruct thinking, to a point. I notice you didn't say we can completely reconstruct how humans think about problems. Why not? We have two primary means for understanding thought, and both are deeply flawed: 1. Introspection. Introspection allows us to analyze our mental life in a reflective way. This is possible because we are able to construct mental models of our mental models. There are three flaws with introspection. The first, least serious flaw is that we only have access to that which is present in our conscious awareness. We cannot introspect about unconscious processes, by definition. This is a less serious objection because it's possible in practice to become conscious of phenomena there were previously unconscious, by developing our meta-mental-models. The question here becomes, is there any reason in principle that we cannot become conscious of *all* mental processes? The second flaw is that, because introspection relies on the meta-models we need to make sense of our internal, mental life, the possibility is always present that our meta-models themselves are flawed. Worse, we have no way of knowing if they are wrong, because we often unconsciously, unwittingly deny evidence contrary to our conception of our own cognition, particularly when it runs counter to a positive account of our self-image. Harvard's Project Implicit experiment (https://implicit.harvard.edu/implicit/) is a great way to demonstrate how we remain ignorant of deep, unconscious biases. Another example is how little we understand the contribution of emotion to our decision-making. Joseph Ledoux and others have shown fairly convincingly that emotion is a crucial part of human cognition, but most of us (particularly us men) deny the influence of emotion on our decision making. The final flaw is the most serious. It says there is a fundamental limit to what introspection has access to. This is the an eye cannot see itself objection. But I can see my eyes in the mirror, says the devil's advocate. Of course, a mirror lets us observe a reflected version of our eye, and this is what introspection is. But we cannot see inside our own eye, directly - it's a fundamental limitation of any observational apparatus. Likewise, we cannot see inside the very act of model-simulation that enables introspection. Introspection relies on meta-models, or models about models, which are activated/simulated *after the fact*. We might observe ourselves in the act of introspection, but that is nothing but a meta-meta-model. Each introspectional act by necessity is one step (at least) removed from the direct, in-the-present flow of cognition. This means that we can never observe the cognitive machinery that enables the act of introspection itself. And if you don't believe that introspection
Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
Will, My plan is go for 3) Usefulness. Cognition is useful from an evolutionary point of view, if we try to create systems that are useful in the same situations (social, building world models), then we might one day stumble upon cognition. Sure, that's a valid approach for creating something we might call intelligent. My diatribe there was about human thought (the only kind we know of), not cognition in general. This by the way is why I don't self-organise purpose. I am pretty sure a specified purpose (not the same thing as a goal, at all) is needed for an intelligence. Will OK, then who or what specified the purpose of the first life forms? It's that intuition of yours that leads directly to Intelligent Design. As an aside, I love the irony that AI researchers who try to design intelligence are unwittingly giving ammunition to Intelligent Design arguments. Terren --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
2008/7/2 Vladimir Nesov [EMAIL PROTECTED]: On Wed, Jul 2, 2008 at 2:48 PM, William Pearson [EMAIL PROTECTED] wrote: Okay let us clear things up. There are two things that need to be designed, a computer architecture or virtual machine and programs that form the initial set of programs within the system. Let us call the internal programs vmprograms to avoid confusion.The vmprograms should do all the heavy lifting (reasoning, creating new programs), this is where the lawlful and consistent pressure would come from. It is at source code of vmprograms that all needs to be changeable. However the pressure will have to be somewhat experimental to be powerful, you don't know what bugs a new program will have (if you are doing a non-tight proof search through the space of programs). So the point of the VM is to provide a safety net. If an experiment goes awry, then the VM should allow each program to limit the bugged vmprograms ability to affect it and eventually have it removed and the resources applied to it. Here is a toy scenario where the system needs this ability. *Note it is not anything that is like a full AI but illustrates a facet of something a full AI needs IMO*. Consider a system trying to solve a task, e.g. navigate a maze, that also has a number of different people out there giving helpful hints on how to solve the maze. These hints are in the form of patches to the vmprograms, e.g. changing the representation to 6-dimensional, giving another patch language that has better patches. So the system would make copies of the part of it to be patched and then patch it. Now you could give a patch evaluation module to see which patch works best, but what would happen if the module that implemented that vmprogram wanted to be patched? My solution to the problem is to allow the patch and non-patched version compete in the adhoc economic arena, and see which one wins. What are the criteria that VM applies to vmprograms? If VM just shortcircuits the economic pressure of agents to one another, it in itself doesn't specify the direction of the search. The human economy works to efficiently satisfy the goals of human beings who already have their moral complexity. It propagates the decisions that customers make, and fuels the allocation of resources based on these decisions. Efficiency of economy is in efficiency of responding to information about human goals. If your VM just feeds the decisions on themselves, what stops the economy from focusing on efficiently doing nothing? They would get less credit from the human supervisor. Let me expand on what I meant about the economic competition. Let us say vmprogram A makes a copy of itself, called A', with some purposeful tweaks, trying to make itself more efficient. A' has some bugs such that the human notices something wrong with the system, she gives less credit on average each time A' is helping out rather than A. Now A and A' both have to bid for the chance to help program B which is closer to the outputting (due to the programming of B), B pays a proportion of the credit it gets back. Now the credit B gets will be lower when A' is helping, than when A is helping. So A' will get less in general than A. There are a few scenarios, ordered from quickest acting to slowest. 1 ) B keeps records of who helps him and sees that A' is not helping him as well as the average, so no longer lets A' bid. A' resources get used when it can't keep up bidding for them. 2) A' continues bidding a lot, to outbid A. However the average amount A' gets is less than it gets back from B. A' bankrupts itself and other programs use its resources. 3) A' doesn't manage to outbid A' after a fair few trials, so gets the same fate as it does in scenario 1) If you start with a bunch of stupid vmprograms, you won't get anywhere. It can just go to nothingness, you do have to design them fairly well, just in such a way that that design can change later. Will --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] the uncomputable
Hector Zenil said: and that is one of the many issues of hypercomputation: each time one comes up with a standard model of hypercomputation there is always another not equivalent model of hypercomputation that computes a different set of functions, i.e. there is no convergence in models unlike what happened when digital computation was characterized This is not entirely true. Turing's oracle machines turn out to correspond to infinite-time machines, and both correspond to the arithmetical hierarchy. On Wed, Jul 2, 2008 at 12:38 PM, Hector Zenil [EMAIL PROTECTED] wrote: The standard model of quantum computation as defined by Feynman and Deutsch is Turing computable (based on the concept of qubits). As proven by Deutsch they compute the same set of functions than Turing machines but faster (if they are feasible). Non-standard models of quantum computation are not widely accepted, and even when they could hypercompute many doubt that we could take any from continuum entangling to perform computations. Non-standard quantum computers have not yet being well defined (and that is one of the many issues of hypercomputation: each time one comes up with a standard model of hypercomputation there is always another not equivalent model of hypercomputation that computes a different set of functions, i.e. there is no convergence in models unlike what happened when digital computation was characterized). Hypercomputational models basically pretend to take advantage from either infinite time or infinite space (including models such as infinite resources, Zeno machines or the Omega-rule, real computation, etc.), from the continuum. Depending of the density of that space/time continuum one can think of several models taking advantage at several levels of the arithmetical hierarchy. But even if there is infinite space or time another issue is how to verify a hypercomputation. One would need another hypercomputer to verify the first and then trust in one. Whether you think hypercomputation, the following paper is a most read for those interested on the topic. Martin Davis' articulates several criticisms: The myth of hypercomputation, in: C. Teuscher (Ed.), Alan Turing: Life and Legacy of a Great Thinker (2004) Serious work on analogous computation can be found in papers from Felix Costa et al.: http://fgc.math.ist.utl.pt/jfc.htm My master's thesis was on the subject so if you are interested in getting an electronic copy just let me know. It is in French though. On Wed, Jul 2, 2008 at 11:15 AM, Abram Demski [EMAIL PROTECTED] wrote: So yes, I think there are perfectly fine, rather simple definitions for computing machines that can (it seems like) perform calculations that turing machines cannot. It should really be noted that quantum computers fall into this class. This is very interesting. Previously, I had heard (but not from a definitive source) that quantum computers could compute in principle only what a Turing machine could compute, but could do it much more efficiently (something like the square root of the effort a Turing machine would need, at least for some tasks). Can you cite any source on this? But I should emphasize that what I am really interested in is computable approximation of uncomputable things. My stance is that an AGI should be able to reason about uncomputable concepts in a coherent manner (like we can), not that it needs to be able to actually compute them (which we can't). On Tue, Jul 1, 2008 at 2:35 AM, Linas Vepstas [EMAIL PROTECTED] wrote: 2008/6/16 Abram Demski [EMAIL PROTECTED]: I previously posted here claiming that the human mind (and therefore an ideal AGI) entertains uncomputable models, counter to the AIXI/Solomonoff model. There was little enthusiasm about this idea. :) I missed your earlier posts. However, I believe that there are models of computation can compute things that turing machines cannot, and that this is not arcane, just not widely known or studied. Here is a quick sketch: Topological finite automata, or geometric finite automata, (of which the quantum finite automata is a special case) generalize the notion of non-deterministic finite automata by replacing its powerset of states with a general topological or geometric space (complex projective space in the quantum case). It is important to note that these general spaces are in general uncountable (have the cardinality of the continuum). It is well known that the languages accepted by quantum finite automata are not regular languages, they are bigger and more complex in some ways. I am not sure what is known about the languages accepted by quantum push-down automata, but intuitively these are clearly different (and bigger) than the class of context-free languages. I believe the concepts of topological finite automata extend just fine to a generalization of turing machines, but I also believe this is a poorly-explored area of
Re: [agi] the uncomputable
On Wed, Jul 2, 2008 at 1:30 PM, Abram Demski [EMAIL PROTECTED] wrote: Hector Zenil said: and that is one of the many issues of hypercomputation: each time one comes up with a standard model of hypercomputation there is always another not equivalent model of hypercomputation that computes a different set of functions, i.e. there is no convergence in models unlike what happened when digital computation was characterized This is not entirely true. Turing's oracle machines turn out to correspond to infinite-time machines, and both correspond to the arithmetical hierarchy. At each level of the arithmetical hierarchy there is a universal oracle machine (a hypercomputer), so there is no standard model of hypercomputation unless you make strong assumptions, unlike digital computation. There are even hiperarithmetical machines and as stated by Post's problem, intermediate non-comparable degrees at each level of the arithmetical and hiperarithmetical (that's why the Turing universe does not build a total order). On Wed, Jul 2, 2008 at 12:38 PM, Hector Zenil [EMAIL PROTECTED] wrote: The standard model of quantum computation as defined by Feynman and Deutsch is Turing computable (based on the concept of qubits). As proven by Deutsch they compute the same set of functions than Turing machines but faster (if they are feasible). Non-standard models of quantum computation are not widely accepted, and even when they could hypercompute many doubt that we could take any from continuum entangling to perform computations. Non-standard quantum computers have not yet being well defined (and that is one of the many issues of hypercomputation: each time one comes up with a standard model of hypercomputation there is always another not equivalent model of hypercomputation that computes a different set of functions, i.e. there is no convergence in models unlike what happened when digital computation was characterized). Hypercomputational models basically pretend to take advantage from either infinite time or infinite space (including models such as infinite resources, Zeno machines or the Omega-rule, real computation, etc.), from the continuum. Depending of the density of that space/time continuum one can think of several models taking advantage at several levels of the arithmetical hierarchy. But even if there is infinite space or time another issue is how to verify a hypercomputation. One would need another hypercomputer to verify the first and then trust in one. Whether you think hypercomputation, the following paper is a most read for those interested on the topic. Martin Davis' articulates several criticisms: The myth of hypercomputation, in: C. Teuscher (Ed.), Alan Turing: Life and Legacy of a Great Thinker (2004) Serious work on analogous computation can be found in papers from Felix Costa et al.: http://fgc.math.ist.utl.pt/jfc.htm My master's thesis was on the subject so if you are interested in getting an electronic copy just let me know. It is in French though. On Wed, Jul 2, 2008 at 11:15 AM, Abram Demski [EMAIL PROTECTED] wrote: So yes, I think there are perfectly fine, rather simple definitions for computing machines that can (it seems like) perform calculations that turing machines cannot. It should really be noted that quantum computers fall into this class. This is very interesting. Previously, I had heard (but not from a definitive source) that quantum computers could compute in principle only what a Turing machine could compute, but could do it much more efficiently (something like the square root of the effort a Turing machine would need, at least for some tasks). Can you cite any source on this? But I should emphasize that what I am really interested in is computable approximation of uncomputable things. My stance is that an AGI should be able to reason about uncomputable concepts in a coherent manner (like we can), not that it needs to be able to actually compute them (which we can't). On Tue, Jul 1, 2008 at 2:35 AM, Linas Vepstas [EMAIL PROTECTED] wrote: 2008/6/16 Abram Demski [EMAIL PROTECTED]: I previously posted here claiming that the human mind (and therefore an ideal AGI) entertains uncomputable models, counter to the AIXI/Solomonoff model. There was little enthusiasm about this idea. :) I missed your earlier posts. However, I believe that there are models of computation can compute things that turing machines cannot, and that this is not arcane, just not widely known or studied. Here is a quick sketch: Topological finite automata, or geometric finite automata, (of which the quantum finite automata is a special case) generalize the notion of non-deterministic finite automata by replacing its powerset of states with a general topological or geometric space (complex projective space in the quantum case). It is important to note that these general spaces are in general uncountable (have the cardinality of the
Re: [agi] the uncomputable
Yes, I was not claiming that there was just one type of hypercomputer, merely that some initially very different-looking types do turn out to be equivalent. You seem quite knowledgeable about the subject. Can you recommend any books or papers? On Wed, Jul 2, 2008 at 1:42 PM, Hector Zenil [EMAIL PROTECTED] wrote: On Wed, Jul 2, 2008 at 1:30 PM, Abram Demski [EMAIL PROTECTED] wrote: Hector Zenil said: and that is one of the many issues of hypercomputation: each time one comes up with a standard model of hypercomputation there is always another not equivalent model of hypercomputation that computes a different set of functions, i.e. there is no convergence in models unlike what happened when digital computation was characterized This is not entirely true. Turing's oracle machines turn out to correspond to infinite-time machines, and both correspond to the arithmetical hierarchy. At each level of the arithmetical hierarchy there is a universal oracle machine (a hypercomputer), so there is no standard model of hypercomputation unless you make strong assumptions, unlike digital computation. There are even hiperarithmetical machines and as stated by Post's problem, intermediate non-comparable degrees at each level of the arithmetical and hiperarithmetical (that's why the Turing universe does not build a total order). On Wed, Jul 2, 2008 at 12:38 PM, Hector Zenil [EMAIL PROTECTED] wrote: The standard model of quantum computation as defined by Feynman and Deutsch is Turing computable (based on the concept of qubits). As proven by Deutsch they compute the same set of functions than Turing machines but faster (if they are feasible). Non-standard models of quantum computation are not widely accepted, and even when they could hypercompute many doubt that we could take any from continuum entangling to perform computations. Non-standard quantum computers have not yet being well defined (and that is one of the many issues of hypercomputation: each time one comes up with a standard model of hypercomputation there is always another not equivalent model of hypercomputation that computes a different set of functions, i.e. there is no convergence in models unlike what happened when digital computation was characterized). Hypercomputational models basically pretend to take advantage from either infinite time or infinite space (including models such as infinite resources, Zeno machines or the Omega-rule, real computation, etc.), from the continuum. Depending of the density of that space/time continuum one can think of several models taking advantage at several levels of the arithmetical hierarchy. But even if there is infinite space or time another issue is how to verify a hypercomputation. One would need another hypercomputer to verify the first and then trust in one. Whether you think hypercomputation, the following paper is a most read for those interested on the topic. Martin Davis' articulates several criticisms: The myth of hypercomputation, in: C. Teuscher (Ed.), Alan Turing: Life and Legacy of a Great Thinker (2004) Serious work on analogous computation can be found in papers from Felix Costa et al.: http://fgc.math.ist.utl.pt/jfc.htm My master's thesis was on the subject so if you are interested in getting an electronic copy just let me know. It is in French though. On Wed, Jul 2, 2008 at 11:15 AM, Abram Demski [EMAIL PROTECTED] wrote: So yes, I think there are perfectly fine, rather simple definitions for computing machines that can (it seems like) perform calculations that turing machines cannot. It should really be noted that quantum computers fall into this class. This is very interesting. Previously, I had heard (but not from a definitive source) that quantum computers could compute in principle only what a Turing machine could compute, but could do it much more efficiently (something like the square root of the effort a Turing machine would need, at least for some tasks). Can you cite any source on this? But I should emphasize that what I am really interested in is computable approximation of uncomputable things. My stance is that an AGI should be able to reason about uncomputable concepts in a coherent manner (like we can), not that it needs to be able to actually compute them (which we can't). On Tue, Jul 1, 2008 at 2:35 AM, Linas Vepstas [EMAIL PROTECTED] wrote: 2008/6/16 Abram Demski [EMAIL PROTECTED]: I previously posted here claiming that the human mind (and therefore an ideal AGI) entertains uncomputable models, counter to the AIXI/Solomonoff model. There was little enthusiasm about this idea. :) I missed your earlier posts. However, I believe that there are models of computation can compute things that turing machines cannot, and that this is not arcane, just not widely known or studied. Here is a quick sketch: Topological finite automata, or geometric finite automata, (of which the quantum finite automata
Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
How do you assign credit to programs that are good at generating good children? Particularly, could a program specialize in this, so that it doesn't do anything useful directly but always through making highly useful children? On Wed, Jul 2, 2008 at 1:09 PM, William Pearson [EMAIL PROTECTED] wrote: 2008/7/2 Vladimir Nesov [EMAIL PROTECTED]: On Wed, Jul 2, 2008 at 2:48 PM, William Pearson [EMAIL PROTECTED] wrote: Okay let us clear things up. There are two things that need to be designed, a computer architecture or virtual machine and programs that form the initial set of programs within the system. Let us call the internal programs vmprograms to avoid confusion.The vmprograms should do all the heavy lifting (reasoning, creating new programs), this is where the lawlful and consistent pressure would come from. It is at source code of vmprograms that all needs to be changeable. However the pressure will have to be somewhat experimental to be powerful, you don't know what bugs a new program will have (if you are doing a non-tight proof search through the space of programs). So the point of the VM is to provide a safety net. If an experiment goes awry, then the VM should allow each program to limit the bugged vmprograms ability to affect it and eventually have it removed and the resources applied to it. Here is a toy scenario where the system needs this ability. *Note it is not anything that is like a full AI but illustrates a facet of something a full AI needs IMO*. Consider a system trying to solve a task, e.g. navigate a maze, that also has a number of different people out there giving helpful hints on how to solve the maze. These hints are in the form of patches to the vmprograms, e.g. changing the representation to 6-dimensional, giving another patch language that has better patches. So the system would make copies of the part of it to be patched and then patch it. Now you could give a patch evaluation module to see which patch works best, but what would happen if the module that implemented that vmprogram wanted to be patched? My solution to the problem is to allow the patch and non-patched version compete in the adhoc economic arena, and see which one wins. What are the criteria that VM applies to vmprograms? If VM just shortcircuits the economic pressure of agents to one another, it in itself doesn't specify the direction of the search. The human economy works to efficiently satisfy the goals of human beings who already have their moral complexity. It propagates the decisions that customers make, and fuels the allocation of resources based on these decisions. Efficiency of economy is in efficiency of responding to information about human goals. If your VM just feeds the decisions on themselves, what stops the economy from focusing on efficiently doing nothing? They would get less credit from the human supervisor. Let me expand on what I meant about the economic competition. Let us say vmprogram A makes a copy of itself, called A', with some purposeful tweaks, trying to make itself more efficient. A' has some bugs such that the human notices something wrong with the system, she gives less credit on average each time A' is helping out rather than A. Now A and A' both have to bid for the chance to help program B which is closer to the outputting (due to the programming of B), B pays a proportion of the credit it gets back. Now the credit B gets will be lower when A' is helping, than when A is helping. So A' will get less in general than A. There are a few scenarios, ordered from quickest acting to slowest. 1 ) B keeps records of who helps him and sees that A' is not helping him as well as the average, so no longer lets A' bid. A' resources get used when it can't keep up bidding for them. 2) A' continues bidding a lot, to outbid A. However the average amount A' gets is less than it gets back from B. A' bankrupts itself and other programs use its resources. 3) A' doesn't manage to outbid A' after a fair few trials, so gets the same fate as it does in scenario 1) If you start with a bunch of stupid vmprograms, you won't get anywhere. It can just go to nothingness, you do have to design them fairly well, just in such a way that that design can change later. Will --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
2008/7/2 Abram Demski [EMAIL PROTECTED]: How do you assign credit to programs that are good at generating good children? I never directly assign credit, apart from the first stage. The rest of the credit assignment is handled by the vmprograms, er, programming. Particularly, could a program specialize in this, so that it doesn't do anything useful directly but always through making highly useful children? As the parent controls the code of its offspring, it could embed code in its offspring to pass a small portion of the credit they get back to it. They would have to be careful how much to skim off so the offspring could still thrive. Will --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
2008/7/2 Vladimir Nesov [EMAIL PROTECTED]: On Wed, Jul 2, 2008 at 9:09 PM, William Pearson [EMAIL PROTECTED] wrote: They would get less credit from the human supervisor. Let me expand on what I meant about the economic competition. Let us say vmprogram A makes a copy of itself, called A', with some purposeful tweaks, trying to make itself more efficient. So, this process performs optimization, A has a goal that it tries to express in form of A'. What is the problem with the algorithm that A uses? If this algorithm is stupid (in a technical sense), A' is worse than A and we can detect that. But this means that in fact, A' doesn't do its job and all the search pressure comes from program B that ranks the performance of A or A'. This generate-blindly-or-even-stupidly-and-check is a very inefficient algorithm. If, on the other hand, A happens to be a good program, then A' has a good change of being better than A, and anyway A has some understanding of what 'better' means, then what is the role of B? B adds almost no additional pressure, almost everything is done by A. How do you distribute the optimization pressure between generating programs (A) and checking programs (B)? Why do you need to do that at all, what is the benefit of generating and checking separately, compared to reliably generating from the same point (A alone)? If generation is not reliable enough, it probably won't be useful as optimization pressure anyway. The point of A and A' is that A', if better, may one day completely replace A. What is very good? Is 1 in 100 chances of making a mistake when generating its successor very good? If you want A' to be able to replace A, that is only 100 generations before you have made a bad mistake, and then where do you go? You have a bugged program and nothing to act as a watchdog. Also if A' is better than time A at time t, there is no guarantee that it will stay that way. Changes in the environment might favour one optimisation over another. If they both do things well, but different things then both A and A' might survive in different niches. I would also be interested in why you think we have programmers and system testers in the real world. Also worth noting is most optimisation will be done inside the vmprograms, this process is only for very fundamental code changes, e.g. changing representations, biases, ways of creating offspring. Things that cannot be tested easily any other way. I'm quite happy for it to be slow, because this process is not where the majority of quickness of the system will rest. But this process is needed for intelligence else you will be stuck with certain ways of doing things when they are not useful. Will Pearson --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
Re: [agi] the uncomputable
On Wed, Jul 2, 2008 at 3:39 PM, Abram Demski [EMAIL PROTECTED] wrote: Yes, I was not claiming that there was just one type of hypercomputer, merely that some initially very different-looking types do turn out to be equivalent. You seem quite knowledgeable about the subject. Can you recommend any books or papers? Sure. Are you interested on hypercomputation or arguments against hypercomputation? For the latter, I already gave the reference to Martin Davis' paper on 'The Myth of Hypercomputation'. For serious work on hypercomputation I would recommend people doing real and analogue computation. The groups of Cris Moore and Felix Costa, et al. E.g. Cris Moore, Recursion Theory on the Reals and Continuous-time Computation. Theoretical Computer Science 162 (1996) 23-44. Bruno Loff, Jerzy Mycka and Felix Costa, The new promise of analog computation, invited paper, in S. Barry Cooper, Benedikt Löwe, and Andrea Sorbi (eds.), Third Conference on Computability in Europe, CiE2007, Siena, Italy, June 18--23, 2007, Computation and Logic in the Real World, Lecture Notes in Computer Science 4497: 189--195, Springer, 2007. Bruno Loff and Felix Costa, Five views of hypercomputation, International Journal of Unconventional Computing, Special Issue on Hypercomputation, to appear. Bruno Loff, Jerzy Mycka, and Felix Costa, Computability on reals, infinite limits and differential equations, Applied Mathematics and Computation, 191(2):353–371, Elsevier, 2007. On Wed, Jul 2, 2008 at 1:42 PM, Hector Zenil [EMAIL PROTECTED] wrote: On Wed, Jul 2, 2008 at 1:30 PM, Abram Demski [EMAIL PROTECTED] wrote: Hector Zenil said: and that is one of the many issues of hypercomputation: each time one comes up with a standard model of hypercomputation there is always another not equivalent model of hypercomputation that computes a different set of functions, i.e. there is no convergence in models unlike what happened when digital computation was characterized This is not entirely true. Turing's oracle machines turn out to correspond to infinite-time machines, and both correspond to the arithmetical hierarchy. At each level of the arithmetical hierarchy there is a universal oracle machine (a hypercomputer), so there is no standard model of hypercomputation unless you make strong assumptions, unlike digital computation. There are even hiperarithmetical machines and as stated by Post's problem, intermediate non-comparable degrees at each level of the arithmetical and hiperarithmetical (that's why the Turing universe does not build a total order). On Wed, Jul 2, 2008 at 12:38 PM, Hector Zenil [EMAIL PROTECTED] wrote: The standard model of quantum computation as defined by Feynman and Deutsch is Turing computable (based on the concept of qubits). As proven by Deutsch they compute the same set of functions than Turing machines but faster (if they are feasible). Non-standard models of quantum computation are not widely accepted, and even when they could hypercompute many doubt that we could take any from continuum entangling to perform computations. Non-standard quantum computers have not yet being well defined (and that is one of the many issues of hypercomputation: each time one comes up with a standard model of hypercomputation there is always another not equivalent model of hypercomputation that computes a different set of functions, i.e. there is no convergence in models unlike what happened when digital computation was characterized). Hypercomputational models basically pretend to take advantage from either infinite time or infinite space (including models such as infinite resources, Zeno machines or the Omega-rule, real computation, etc.), from the continuum. Depending of the density of that space/time continuum one can think of several models taking advantage at several levels of the arithmetical hierarchy. But even if there is infinite space or time another issue is how to verify a hypercomputation. One would need another hypercomputer to verify the first and then trust in one. Whether you think hypercomputation, the following paper is a most read for those interested on the topic. Martin Davis' articulates several criticisms: The myth of hypercomputation, in: C. Teuscher (Ed.), Alan Turing: Life and Legacy of a Great Thinker (2004) Serious work on analogous computation can be found in papers from Felix Costa et al.: http://fgc.math.ist.utl.pt/jfc.htm My master's thesis was on the subject so if you are interested in getting an electronic copy just let me know. It is in French though. On Wed, Jul 2, 2008 at 11:15 AM, Abram Demski [EMAIL PROTECTED] wrote: So yes, I think there are perfectly fine, rather simple definitions for computing machines that can (it seems like) perform calculations that turing machines cannot. It should really be noted that quantum computers fall into this class. This is very interesting. Previously, I had heard (but not from a
Re: [agi] WHAT SORT OF HARDWARE $33K AND $850K BUYS TODAY FOR USE IN AGI
On Thu, Jul 3, 2008 at 12:59 AM, William Pearson [EMAIL PROTECTED] wrote: 2008/7/2 Vladimir Nesov [EMAIL PROTECTED]: On Wed, Jul 2, 2008 at 9:09 PM, William Pearson [EMAIL PROTECTED] wrote: They would get less credit from the human supervisor. Let me expand on what I meant about the economic competition. Let us say vmprogram A makes a copy of itself, called A', with some purposeful tweaks, trying to make itself more efficient. So, this process performs optimization, A has a goal that it tries to express in form of A'. What is the problem with the algorithm that A uses? If this algorithm is stupid (in a technical sense), A' is worse than A and we can detect that. But this means that in fact, A' doesn't do its job and all the search pressure comes from program B that ranks the performance of A or A'. This generate-blindly-or-even-stupidly-and-check is a very inefficient algorithm. If, on the other hand, A happens to be a good program, then A' has a good change of being better than A, and anyway A has some understanding of what 'better' means, then what is the role of B? B adds almost no additional pressure, almost everything is done by A. How do you distribute the optimization pressure between generating programs (A) and checking programs (B)? Why do you need to do that at all, what is the benefit of generating and checking separately, compared to reliably generating from the same point (A alone)? If generation is not reliable enough, it probably won't be useful as optimization pressure anyway. The point of A and A' is that A', if better, may one day completely replace A. What is very good? Is 1 in 100 chances of making a mistake when generating its successor very good? If you want A' to be able to replace A, that is only 100 generations before you have made a bad mistake, and then where do you go? You have a bugged program and nothing to act as a watchdog. Also if A' is better than time A at time t, there is no guarantee that it will stay that way. Changes in the environment might favour one optimisation over another. If they both do things well, but different things then both A and A' might survive in different niches. I suggest you read ( http://sl4.org/wiki/KnowabilityOfFAI ) If your program is a faulty optimizer that can't pump the reliability out of its optimization, you are doomed. I assume you argue that you don't want to include B in A, because a descendant of A may start to fail unexpectedly. But if you reliably copy B inside each of A's descendants, this particular problem won't appear. The main question is: what is the difference between just trying to build a self-improving program A and doing so inside your testing environment. If there is no difference, you add nothing by your framework. If there is, it would be good to find out what it is. I would also be interested in why you think we have programmers and system testers in the real world. Testing that doesn't even depend on program's internal structure and only checks its output (as in your economy setup) isn't nearly good enough. Testing that you're referring to in this post (activity performed by humans, based on specific implementation and understanding of high-level specification that says what algorithm should do) has very little to do with testing that you propose in the framework (fixed program B). Anyway, you should answer on that question yourself: what is the essence of useful activity that is performed by software testing and that you capture in your framework. Arguing that there must be some such essence and that it must transfer to your setting isn't reliable. Also worth noting is most optimisation will be done inside the vmprograms, this process is only for very fundamental code changes, e.g. changing representations, biases, ways of creating offspring. Things that cannot be tested easily any other way. I'm quite happy for it to be slow, because this process is not where the majority of quickness of the system will rest. But this process is needed for intelligence else you will be stuck with certain ways of doing things when they are not useful. Being stuck in development is a problem of search process, it can as well be a problem of process A that should be resolved from within A. -- Vladimir Nesov [EMAIL PROTECTED] http://causalityrelay.wordpress.com/ --- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244id_secret=106510220-47b225 Powered by Listbox: http://www.listbox.com
[agi] WHAT PORTION OF CORTICAL PROCESSES ARE BOUND BY THE BINDING PROBLEM?
WHAT PORTION OF CORTICAL PROCESSES ARE BOUND BY THE BINDING PROBLEM? Here is an important practical, conceptual problem I am having trouble with. In an article entitled Are Cortical Models Really Bound by the 'Binding Problem'? Tomaso Poggio's group at MIT takes the position that there is no need for special mechanisms to deal with the famous binding problem --- at least in certain contexts, such as 150 msec feed forward visual object recognition. This article implies that a properly designed hierarchy of patterns that has both compositional and max-pooling layers (I call them gen/comp hierarchies) automatically handles the problem of what sub-elements are connected with which others, preventing the need for techniques like synchrony to handle this problem. Poggio's group has achieved impressive results without the need for special mechanisms to deal with binding in this type of visual recognition, as is indicated by the two papers below by Serre (the later of which summarizes much of what is in the first, which is an excellent, detailed PhD thesis.) The two works by Geoffrey Hinton cited below are descriptions of Hinton's hierarchical feed-forward neural net recognition system (which, when run backwards, generates patterns similar to those it has been trained on). These two works by Hinton show impressive results in handwritten digit recognition without any explicit mechanism for binding. In particular, watch the portion of the Hinton YouTube video starting at 21:35 - 26:39 where Hinton shows his system alternating between recognizing a pattern and then generating a similar pattern stochastically from the higher level activations that have resulted from the previous recognition. See how amazingly well his system seems to capture the many varied forms in which the various parts and sub-shapes of numerical handwritten digits are related. So my question is this: HOW BROADLY DOES THE IMPLICATION THAT THE BINDING PROBLEM CAN BE AUTOMATICALLY HANDLED BY A GEN/COMP HIERARCHY OR A HINTON-LIKE HIERARCHY APPLY TO THE MANY TYPES OF PROBLEMS A BRAIN LEVEL ARTIFICIAL GENERAL INTELLIGENCE WOULD BE EXPECTED TO HANDLE? In particular HOW APPLICABLE IS IT TO SEMANTIC PATTERN RECOGNITION AND GENERATION --- WITH ITS COMPLEX AND HIGHLY VARIED RELATIONS --- SUCH AS IS COMMONLY INVOLVED IN HUMAN LEVEL NATURAL LANGUAGE UNDERSTANDING AND GENERATION? The paper Are Cortical Models Really Bound by the 'Binding Problem'?, suggests in the first full paragraph on its second page that gen/comp hierarchies avoids the binding problem by coding an object through a set of intermediate features made up of local arrangements of simpler features [that] sufficiently constrain the representation to uniquely code complex objects without retaining global positional information. For example, in the context of speech recognition, ...rather than using individual letters to code words, letter pairs or higher-order combinations of letters can be used-i.e., although the word tomaso might be confused with the word somato if both were coded by the sets of letters they are made up of, this ambiguity is resolved if both are represented through letter pairs. The issue then becomes, WHAT SUB-SETS OF THE TYPES OF PROBLEMS THE HUMAN BRAIN HAS TO PERFORM CAN BE PERFORMED IN A MANNER THAT AVOIDS THE BINDING PROBLEM JUST BY USING A GEN/COMP HIERARCHY WITH SUCH A SET OF SIMPLER FEATURES [THAT] SUFFICIENTLY CONSTRAIN THE REPRESENTATION TO UNIQUELY CODE THE TYPE OF PATTERNS SUCH TASKS REQUIRE? There is substantial evidence that the brain does require synchrony for some of its tasks --- as has been indicated by the work of people like Wolf Singer --- suggesting that binding may well be a problem that cannot be handled alone by the specificity of the brain's gen/comp hierarchies for all mental tasks. The table at the top of page 75 of Serre's impressive PhD thesis suggests that his system --- which performs very quick feedforwad object recognition roughly as well as a human --- has an input of 160 x 160 pixels, and requires 23 million pattern models. Such a large number of patterns helps provide the simpler features [that] sufficiently constrains the representation to uniquely code complex objects without retaining global positional information. But, it should be noted --- as is recognized in Serre's paper --- that the very rapid 150 msec feed forward recognition described in that paper is far from all of human vision. Such rapid recognition --- although surprisingly accurate given how fast it is --- is normally supplemented by more top down vision processes to confirm its best guesses. For example, if a human is shown a photograph of a face, his eyes will normally saccade over it, with multiple fixation points, often on key features such as eyes, nose, corners of mouth, points on the outline of the face, all indicating the recognition of the face is normally much more than one rapid feed forward
