TLD tracker and other recent object trackers do single-event learning actually
Modern deep learning systems are not just demos, they are often commercially deployed successful products... I don't think these deep learning systems are the key to AGI, but they're part of the picture and they are genuine progress... far more so than Eliza or other chatbots, for sure... FWIW my thoughts on how the recent deep learning work fits into the OpenCog universe can be found at the following links... Somewhat like pylearn2, I would define network structure in OpenCog, but do the heavy lifting in CUDA via python and Theano... Some general thoughts: http://blog.opencog.org/2015/11/01/putting-deep-perceptual-learning-in-opencog/ A fairly specific design suggestion: http://wiki.opencog.org/wikihome/index.php/Deep_Learning_Perception_in_OpenCog (edited) -- Ben On Thu, Nov 5, 2015 at 2:05 AM, Steve Richfield <[email protected]> wrote: > Ben, > > On Wed, Nov 4, 2015 at 12:50 AM, Ben Goertzel <[email protected]> wrote: > >> Hmmm... >> >> With all the recent progress in deep learning for image processing, >> > > ... which the last I heard remains unable to do single-event learning like > we do. And, even if it could do single-event learning, there is a VAST > chasm between learning and thinking in terms of what has been learned. If > anything, it seems to me that image processing is even further removed from > the process of thinking - whatever THAT is. > > How long do YOU think it will take for image processing to "evolve" > whatever is needed to "think"? > > Now, multiply that number by your historical overoptimistic ratio. 8-:D> > > >> and near-term similar progress in deep learning for video processing, how >> can you claim only text-based AGI is feasible in the near future?? >> > > The present situation with image and video processing is VERY similar to > the 1960s situation with text processing, Eliza, etc. Some exciting demos > have been done, but no one sees a clear path to the goal, though there are > some clouded paths. All that seemed to be needed were enough rules and a > fast enough computer to run them at practical speeds. What they missed was > that computers might NEVER become fast enough to do this by brute force - > because of its parallel to the chess playing problem that will clearly > never be solved by simple brute force. Just like "tree trimming" gets past > this in chess, techniques like LFU-driven evaluation of rules was needed in > text processing for pretty much the same reasons. > > We don't even know what barriers there are to be overcome are in image > processing, but we DO know that it takes a LOT more processing to compute > with pictures than to compute with words, and a way to compute fast enough > with words is VERY new. It took half a century to find a method that worked > for words. I wonder how long it will take to find a method that works for > images? > > Sure there might be a set of tricks to get past this, accompanied by > people smart enough to come up with those tricks in the "near" future, but > I see NO reason to believe that this is true, and the history of AI > strongly suggests that it is NOT true. > >> >> Current deep learning architectures are not AGI-capable, but they do >> demonstrate that awesome stuff can be done with videos and images on >> current computing infrastructure. >> > > Agreed. However, to my mind, this has NOTHING to do with AGI beyond > suggesting avenues for investigation. > >> >> Hardware has come a long way ;-) >> > > Back to the chess-playing parallel. Using brute force techniques, the > difference between clunky old vacuum tube computers (I have years of > experience on IBM-650 and IBM-709 computers in the 1960s) and modern-day > multi-core PCs is only between 1.5 and 2 moves of lookahead in chess. For > text the situation is even more extreme because there are many more words > than moves in chess, the ratio being between 1 and 2 word longer phrases > that can be understood at the same speed. Further, the 4GHz "barrier" of a > decade ago seems to be remaining pretty much in place. Sure, adding more > cores and refining the architecture helps a LOT - the asymptote being maybe > 2 orders of magnitude faster than a 2004 single-core PC, but not like the > ~4 orders of magnitude that LFU-driven evaluation (on a single processor) > brings. > > My point here is that you definitely will not live long enough; and the > universe might not exist long enough; for hardware improvements alone to > make brute force methods applicable to AGI, though like the original Eliza > program, I suspect you will eventually produce some intriguing demos that > suck others into expending their lives in this quest. > > To make AGI work, you MUST solve TWO problems: > 1. Figure out how to "think" on a computer. YOU are an existence proof > that this is possible. > 2. Figure out how to make #1 run fast enough for practical application. > There is NO proof that this is even possible, and some good arguments (like > my bidirectional processing discussion) that it is NOT possible on present > hardware architectures, because the difficulty grows faster than linearly > with complexity. Rather than work on more powerful architectures, which > seems to me to be the obvious direction needed for ALL applications, most > people here continue in a quasi-religious belief that bidirectional > processing is NOT needed, when neurons clearly do this sort of computation. > Of course, we can't really discuss the details of what is needed to make #1 > run fast enough, until someone makes #1 run at all, so we may have to set > this particular discussion aside for a few decades. > > Note in passing what I am **NOT** saying: > 1. That TAGI is even possible. I suspect it may NOT be possible. > 2. That even if TAGI is possible, that it will have great social value. I > suspect that, like human savants with photographic memories, it would have > limited value. I had a savant friend John who made a good living doing what > Google now does for a Beltway Bandit company - reading everything in a > field and then sitting in on technical meetings and reciting passages from > stacks of books he had read during preceding days. While John was able to > do this, he was NOT able to advance the technologies from which he was > reciting. Similarly, I wouldn't expect any Singularity miracles from TAGI, > at least NOT in the "near" future. > > I am only saying that right now, given present-day hardware and CS > knowledge, that my LFU-driven evaluation in TAGI is the ONLY game in town > that might have ANY (albeit thin) chance of near-term success. Other > methods are at least decades away from practical application, not because > problem #1 (figuring out how to "think") is necessarily so far off, but > because problem #2 (how to do it at useful speeds) looks pretty hopeless > for the foreseeable future. > > Steve > ========================== > >> >> On Wed, Nov 4, 2015 at 1:57 PM, Steve Richfield < >> [email protected]> wrote: >> >>> Ben, >>> >>> On Tue, Nov 3, 2015 at 1:26 PM, Ben Goertzel <[email protected]> wrote: >>> >>>> >>>> Hi Steve, >>>> >>>> As for taking >$10M in venture money, yes you have to do the legal >>>> paperwork right and the $$ can come only from accredited investors, but >>>> there is no rule about billionaires being involved; the criteria for being >>>> a accredited investor are much weaker... >>>> https://en.wikipedia.org/wiki/Accredited_investor#United_States >>>> >>> >>> Quoting myself: "To make this work, a company MUST have someone on board >>> whose friends collectively have >$1B. In my own case," You apparently >>> missed the work "collectively", If you are looking for $100M, if your >>> friends' assets totaled $1B, then each would have to invest 10%, which is >>> probably TOO MUCH for such a speculative investment. >>> >>>> >>>> Also -- I do know a couple billionaires personally, as it happens (and >>>> a number of others close to that level); they are interested in AGI but >>>> evidently I haven't convinced them to fund OpenCog amply at this point.... >>>> One of the secrets of being friends with the wealthy is not constantly >>>> asking them for money. By the same token, I also have some friends who are >>>> immensely less financially well-off than me and I am happy that they do not >>>> constantly ask me for funds, only would do so in an emergency situation ;/ >>>> .... Human life has its complexities... >>>> >>> >>> Most billionaires have someone else handling their investments. Your >>> representative should approach THAT person, adding that the billionaire >>> already knows you. >>> >>>> >>>> The key issue is that, even if someone has massive wealth and believes >>>> AGI is feasible and important, they still don't necessarily have a way to >>>> know WHICH of the many approaches to AGI is really likely to succeed. They >>>> are not experts. And if they choose to donate/invest in AGI they will tend >>>> to ask the profs at the nearest big-name university what to donate/invest >>>> in; and these profs tend to be fairly conservative minded.... >>>> >>> >>> Here is my own analysis: >>> >>> With current hardware, only textual AGI (new term: TAGI) seems >>> approachable, so if anyone is going to put out an AGI product in the next >>> couple/few years it will have to TAGI, until we have many additional orders >>> of magnitude in speed. >>> >>> The ONLY known approach to possibly doing TAGI faster than typing speed >>> is to trigger processing on Least Frequently Used (LFU) words, and my >>> patent covers the only known method of implementing such analysis. >>> >>> Apparently, the only difference between "weak" AI as I am shooting for >>> and TAGI is that a TAGI implementation would have self-modifying tables, >>> modified according to an algorithm that has yet to be discovered. The >>> remainder of a TAGI absolutely **MUST** either: >>> >>> 1. be built on the same engine as the "weak" AI system, or >>> 2. operate according to unknown principles which have remained >>> undiscovered since the 1960s despite hundreds of bright PhDs working on >>> the >>> problem. >>> >>> So, not only do I apparently have the only game in town for the >>> near-term, but I have a pretty good argument that the situation is going to >>> stay that way. >>> >>> EVERYONE interested in TAGI should be working on the self-modifying >>> table problem. Even if other methods are found, they can NOT be executed at >>> useful speeds on current hardware. >>> >>> This is really simple - as not only is there is no other viable path, >>> there is a good argument that no other path will be found in the >>> foreseeable future. >>> >>>> >>>> The easiest way to break through this -- if one wants to secure >>>> substantial, stable funding for AGI rather than for development and >>>> marketing of associated commercial products -- will be to make a cool >>>> enough demonstrated achievement with one's in-progress AGI system, that >>>> folks with spare wealth are willing to bet on your system as "the one" (or >>>> at least a one) that will get there first. But of course this is difficult >>>> because AGI is a hard problem and specifically because of the reasons I >>>> outlined in >>>> http://multiverseaccordingtoben.blogspot.kr/2011/06/why-is-evaluating-partial-progress.html >>>> some time ago... >>>> >>> >>> Or my plan - build a commercial AI system to fund further efforts, and >>> then devote resources to piercing the self-modifying table problem that >>> will run on the SAME platform. >>> >>>> >>>> My current 3 year plan for OpenCog is aimed at overcoming these issues >>>> and finally making enough practical progress to make a working demo system >>>> that shows enough early-stage AGI to overcome the doubts of possible >>>> donors/investors who are psyched about AGI in principle... >>>> >>> >>> Dare I ask: What might this look like? >>> >>> >>>> This should be easier now than 2-3 years ago due to the increased >>>> popularity of AI overall, and the embrace of AGI as a concept by Google >>>> Deep Mind and others with big names... and also because we have already >>>> made loads of progress within OpenCog, even if not yet shinily >>>> demonstrable. I think we can cobble together the resources to get the 3 >>>> year plan done, but we'll see... ask me in 9 months if we're on the right >>>> track or not ... or it may be obvious ;) ... >>>> >>>> OK, I'll mark my calendar. >>> >>> Steve >>> ============================= >>> >>>> >>>> On Wed, Nov 4, 2015 at 1:13 AM, Steve Richfield < >>>> [email protected]> wrote: >>>> >>>>> Ben, >>>>> >>>>> On Mon, Nov 2, 2015 at 2:58 PM, Ben Goertzel <[email protected]> wrote: >>>>> >>>>>> >>>>>> >>>>>>> >>>>>>> >>>>>>>> >>>>>>>>> which is TOO BIG to fund without an established cash flow, unless >>>>>>>>> you can demonstrate magic like Facebook has done. Anything over $10M >>>>>>>>> MUST >>>>>>>>> be done as a public offering that requires SEC scrutiny that is there >>>>>>>>> to >>>>>>>>> BLOCK companies with big dreams but nothing to show. >>>>>>>>> >>>>>>>> >>>>>>>> That's untrue, actually... >>>>>>>> >>>>>>> >>>>>>> I'll watch. >>>>>>> >>>>>> >>>>>> >>>>>> Silicon Valley companies routinely raise more than $10M in private >>>>>> venture funding (e.g. Vicarious Systems' last round, to name an AI >>>>>> example), so I'm not sure what you're talking about... >>>>>> >>>>> >>>>> Yes - there is a narrow exception to this rule, derived from the 1st >>>>> Amendment (where money is a form of "speech") and enshrined in SEC >>>>> regulations covering "sophisticated investors" as defined as those who >>>>> have >>>>> "won and lost money" in the type of business raising the money. You are >>>>> only allowed to approach 20 such investors. To cover their butts, >>>>> Vicarious >>>>> Systems actually lists their investors in their Wikipedia article. I once >>>>> went through this same process with my own startup Remote Time-Sharing >>>>> Corp. To make this work, a company MUST have someone on board whose >>>>> friends >>>>> collectively have >$1B. In my own case, I had two key board members - one >>>>> of whom owned a local chain of banks, and both of whom were partners in a >>>>> local stock brokerage firm that had a seat on the NYSE. >>>>> >>>>> This further makes my point that you can NOT raise such money from >>>>> crowdfunding (to NON-sophisticated investors by SEC definition) or though >>>>> any sort of public publication (that might be seen by non-sophisticated >>>>> investors). Instead, if you want to raise >$10M without going through a >>>>> public offering (that the SEC would probably block where it involved AGI >>>>> R&D), you MUST do it by *approaching* <20 very rich people. To avoid >>>>> exhausting your 20 on non-investors, "initial" discussions are usually >>>>> "hypothetical". Note that for Vicarious Systems, several of those people >>>>> own/control major corporations, but that those corporations were NOT >>>>> investors - as this would have violated SEC rules because the investors in >>>>> THOSE corporations were NOT "sophisticated". >>>>> >>>>> My statement was slightly over-broad, but my point is unchanged. If >>>>> you are expecting to cross the $10M threshold, you had better have a >>>>> REALLY >>>>> GOOD plan. If you knew people who have billions of dollars, I presume this >>>>> field would now be MUCH better funded. I suggest expanding your circle of >>>>> friends. >>>>> >>>>> Steve >>>>> ================================ >>>>> >>>>> >>>>>> >>>>>> >>>>>> >>>>>>> But Linux has prevailed without selling out to big companies (and >>>>>>>> without Torvalds getting insanely rich, for that matter). And so may >>>>>>>> an >>>>>>>> OSS AGI initiative, if things go well... >>>>>>>> >>>>>>> >>>>>>> Linux is <<<1% of AGI and so makes a horrible model. There was >>>>>>> nothing to do in Linux but code, because there was UNIX to copy, and >>>>>>> UNIX >>>>>>> is SO much simpler than YOU are. >>>>>>> >>>>>> >>>>>> >>>>>> An OS is very different than an AGI, but still the Linux development >>>>>> model has things to teach us. >>>>>> >>>>>> The research aspect of AGI is different, but to help grapple with >>>>>> that we can reach out to academics who can help with an OSS project >>>>>> consistent with their university jobs, but couldn't join a commercial >>>>>> endeavor more than very-part-time without quitting... >>>>>> >>>>>> >>>>>> >>>>>>> >>>>>>> I suggest writing a business plan and putting it out there for >>>>>>> people to pick at, and then including an addendum addressing people >>>>>>> objections that you don't directly address in the plan. >>>>>>> >>>>>>> >>>>>> >>>>>> Well, we have put together a 3 year development plan for OpenCog (not >>>>>> as a business, as an OSS AGI project) and will be doing a corresponding >>>>>> fundraising push probably in Q2 next year, after we get our demos, >>>>>> documentation and tutorials cleaned up also..... But putting it out >>>>>> there >>>>>> for people to pick at doesn't seem productive at this stage.... It has >>>>>> already been critiqued by a number of experienced people, and tweaked a >>>>>> bit >>>>>> as a result... >>>>>> >>>>>> > Until then, it is all smoke and mirrors. >>>>>> >>>>>> If you wish to perceive it as smoke and mirrors, that's just fine >>>>>> with me, my friend ;) ... >>>>>> >>>>>> -- Ben >>>>>> >>>>>> >>>>>> *AGI* | Archives <https://www.listbox.com/member/archive/303/=now> >>>>>> <https://www.listbox.com/member/archive/rss/303/10443978-6f4c28ac> | >>>>>> Modify <https://www.listbox.com/member/?&;> Your Subscription >>>>>> <http://www.listbox.com> >>>>>> >>>>> >>>>> >>>>> >>>>> -- >>>>> Full employment can be had with the stoke of a pen. Simply institute a >>>>> six hour workday. That will easily create enough new jobs to bring back >>>>> full employment. >>>>> >>>>> *AGI* | Archives <https://www.listbox.com/member/archive/303/=now> >>>>> <https://www.listbox.com/member/archive/rss/303/212726-deec6279> | >>>>> Modify <https://www.listbox.com/member/?&;> Your Subscription >>>>> <http://www.listbox.com> >>>>> >>>> >>>> >>>> >>>> -- >>>> Ben Goertzel, PhD >>>> http://goertzel.org >>>> >>>> "The reasonable man adapts himself to the world: the unreasonable one >>>> persists in trying to adapt the world to himself. Therefore all progress >>>> depends on the unreasonable man." -- George Bernard Shaw >>>> *AGI* | Archives <https://www.listbox.com/member/archive/303/=now> >>>> <https://www.listbox.com/member/archive/rss/303/10443978-6f4c28ac> | >>>> Modify <https://www.listbox.com/member/?&;> Your Subscription >>>> <http://www.listbox.com> >>>> >>> >>> >>> >>> -- >>> Full employment can be had with the stoke of a pen. Simply institute a >>> six hour workday. That will easily create enough new jobs to bring back >>> full employment. >>> >>> *AGI* | Archives <https://www.listbox.com/member/archive/303/=now> >>> <https://www.listbox.com/member/archive/rss/303/212726-deec6279> | >>> Modify <https://www.listbox.com/member/?&;> Your Subscription >>> <http://www.listbox.com> >>> >> >> >> >> -- >> Ben Goertzel, PhD >> http://goertzel.org >> >> "The reasonable man adapts himself to the world: the unreasonable one >> persists in trying to adapt the world to himself. Therefore all progress >> depends on the unreasonable man." -- George Bernard Shaw >> *AGI* | Archives <https://www.listbox.com/member/archive/303/=now> >> <https://www.listbox.com/member/archive/rss/303/10443978-6f4c28ac> | >> Modify <https://www.listbox.com/member/?&;> Your Subscription >> <http://www.listbox.com> >> > > > > -- > Full employment can be had with the stoke of a pen. Simply institute a six > hour workday. That will easily create enough new jobs to bring back full > employment. > > *AGI* | Archives <https://www.listbox.com/member/archive/303/=now> > <https://www.listbox.com/member/archive/rss/303/212726-deec6279> | Modify > <https://www.listbox.com/member/?&;> > Your Subscription <http://www.listbox.com> > -- Ben Goertzel, PhD http://goertzel.org "The reasonable man adapts himself to the world: the unreasonable one persists in trying to adapt the world to himself. Therefore all progress depends on the unreasonable man." -- George Bernard Shaw ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
