On Thu, Feb 21, 2019 at 3:10 AM Rob Freeman <[email protected]> wrote:
> > couldn't resist. I do know the frustration > I say things that hurt peoples feelings. Usually as a side-effect. > actually scored a couple of home runs with me. > Thank you! > He has identified linearities in vector models as a key weakness, resolved > the problem as one of reassembling parts (jigsaw pieces?), and even dropped > a mention of category theory, which speaks to formal incompleteness. > (Though I'm not sure he's thought that through, because it actually becomes > an argument for distributed representation and against symbolism, against > any fixed symbolism, anyway.) > How can I avoid writing a long email? "Formal incompleteness" and "category theory" are two different things. When you say "formal incompleteness", I think that what you mean is "the deep learning people made a breakthrough, but they do not understand why it actually works". And that "they are not even trying to understand" or maybe "they are looking in the wrong place for that understanding". And maybe that's true; I dunno. When something works, people rarely ask why. Category theory... has 120-year-old roots in "Universal algebra" and started life as a means of "formalization". But it isn't that, any more; its evolved into something completely different. Lots simpler in some ways, but clashes with what you learned in school. That clash leads to a mental block: like a fun-house mirror, its "obvious" once you get it, but it's perversely "hard to get". Here's what it is, today. It is a theory of dots an arrows, and a set of rules of how dots and arrows can be assembled. And a discussion of what you get when you assemble them. Some structures have a tiny handful of dots and arrows: zero, one, span, cospan, equalizer, pushout, pullback. Some have a countable infinity of dots and arrows: limit and colimit. Some have uncountable infinities of arrows going back and fourth: adjoint functors. There's more. As it happens, almost everything - literally - more or less everything in mathematics can be reduced to dots an arrows. Huh. Who would have guessed? Perhaps this smells like "formalization", depending on your background and experience. To me, (and I think to probably most working category theorists) its more like "gee look at the pretty pictures with dots and arrows" and "huh, what does that remind me of? Oh, right it reminds me of xyz". Back in the heyday, xyz was "the simplicial set" and stuff like that. It's less about "formalization", than it is "hey guys, look at the neat thing that happens when I do this!" Off-topic: 100 years ago (and even today) people thought that they could reduce all of mathematics to "just set theory", which sounds crazy cause sets are these absurdly trite, trivial things. Here, we just swap out sets and replace by trite, trivial dots and arrows. Whoopee! Despite the triteness of sets, the set theorists seem to have a lot on their hands. And very few of them actually work on "formalization". (whatever that might mean). Same deal for category theorists. its trite in the same way that 1+1=2 is trite, and then one day you wake up and go "hang on... prime numbers! WTF!?!?!" The phrase "distributional semantics" keeps popping up in this conversation. Lets pick it apart. A "distribution" is that thing from probability: a bunch of balls distributed into bins. Frequency of dice-rolls and gambling-card deals. "Semantics" is "meaning", the study of "meaning". Put together, what it means is that you can collect statistics on N-grams, for N=3 or N=5, and mash those statistics into K=100 or 200 or 300 bins (balls into bins) and then you notice that: "oh hey, check it! Each different distribution is kind of like the meaning of a word! Oh wow: they're like .. additive -- linear, like 'King - Man + Woman = Queen' holds as an approximate equation between the distributions for the words King, Man, Woman, Queen. What's the other famous example? 'Berlin - Germany + France = Paris'. So this is the idea of "distributional semantics". I think if you resurrected a dead linguist from the 1950's, they'd say something like "no shit, sherlock, what else did you figure out while I was dead?" The linguists always knew about distributional semantics, they just didn't have the compute power, the algorithms. The deep-learning guys have compute power and algorithms, but never really thought about linguistics before. Perhaps they think that linguistics is trite. You know, like 1+1=2 is trite. For me, its just more like a playground, or a candy store. So I crack open a book on linguistics, and its got dots and arrows in it, and like oh, cool! That really really reminds me of a simplicial set, except that its a simplicial set with extra prongs pointing in other directions, so its not that, but a fun-house mirror of one. So then I read about neural nets, and they talk about vectors, except that they aren't really vectors (there's no rotational symmetry; its not Euclidean space, why the heck do the neural net people say "vector" when *obviously* they are not vectors? Do they seriously think that 'Berlin - Germany + France = Paris' implies that distributions are vectors?) Anyway, distributions obey various axioms (and so do vectors), and those axioms are dot-and-arrow-like, and so there's yet more fun-n-games of "gee, lookit what happens when one does this with that". Mostly, I think one can go pretty darned far with just "plain old statistics", you just have to break out of the mind-set of N-grams and K=300-dimensional "vectors". I don't even think its that hard to do. You just play the usual game of turning it sideways. Email too long, again. Sheesh. --linas > > Formal incompleteness might actually be the biggest of those from a > theoretical point of view. It can be a kind of theoretical unifying factor, > reaching down even to deep learning, then extending it. I don't think Linas > would see it that way. But that makes it even more a situation which is > ripe to have more eyes on it, to examine exactly what it means. > > So I see a short distance, and concrete changes which could be made, to > make a big difference. > > I'm criticizing the status quo, definitely the deep learning work which > sparked this thread, but also even more what I now learn OpenCog has come > to, not because it is bad, but because I see the solution as very close, > and I want to span that last gap. > > -Rob > > P. S. Linas, I see you have posted more documentation. As a first > impression I think I'm going to suggest you simplify the way your "atoms" > connect, very, very simply, and ring it like a bell to find sheaves among > them simply too. It may just not work yet because serial hardware will be > too slow. But I'll look in more detail. > > -Rob > > On Thu, Feb 21, 2019 at 6:34 PM Nanograte Knowledge Technologies < > [email protected]> wrote: > >> Rob >> >> So much of what you're saying makes so much sense, that it's almost >> scary. The years of theoretical discussions I had with Prof Honeycutt bears >> out what you're saying. Why I constantly say it remains an issue of design, >> is exactly what I see in your communications. It is because if one designs >> and implements, one simply becomes stuck with that design and trying to >> make it work. Very few social groups are willing to throw away a design >> that is clearly inappropriate - even after a few years' of investment - to >> use that learning for a radical redesign. Karl Mannheim referred to such >> elitist practices as being rooted in Ideological and Utopian thinking. >> Clearly, a society of developers need new ways of thinking about AGI. >> >> My contention is; to develop AGI, one must first become AGI. To my mind, >> the artifact of such a transformation would be represented by an AGI >> blueprint. A strong truth is often most unpopular. I know my perspective is >> disturbing a number of staid veterans in the field, but it does not >> conclude that what I'm saying has no relevance. >> >> Rob, I'd like to venture to say that even what you're discussing, I have >> theoretically researched beyond with my mentor. There are a number of >> levels still above what you propose, which we have scant theory for. >> However, you are already thinking about what the integration and outcomes >> of such integration would mean. Such is systems thinking. >> >> I'm grateful for those who are diligently clawing away - with seemingly >> bare hands at times - at the AGI coalface, but I think it would've been >> much better had "we" first agreed on a the continuous development approach >> of an appropriate toolkit. There remains a chronic absence in consensus and >> I see no end to it. A failure of society to organize? >> >> One cannot use methodology, which has proven to be limited (eventually) >> to try to address futuristic solutions that we have to still develop "new" >> theory for. As such, I propose, at first, a radical redesign, supported by >> a pragmatic, next-step approach. We need case-based successes to learn >> from. >> >> In its absence, a number of us would simply follow our own idea of what >> an AGI system would become. >> >> Last, a critical thought. I have no idea why anyone would spend a useful >> life on building submarines, when those have already been perfected and are >> not what is needed in the world. To my mind, that is just playing it safe. >> What it is not, is assuming industry leadership and stepping out to define >> what AGI should become. To do so takes very-specific personality, and >> character. It requires a historical bigness, not petty nit picking and >> ridicule. This, I'm pointing to at our learned friend (and similar others >> who I have encountered here) who seem to think no one else in this whole, >> wide world has much use to contribute to their version(s) of AGI. I think >> they're sorely mistaken, but only time would tell. >> >> Rgds >> >> Robert Benjamin >> >> >> >> ------------------------------ >> *From:* Rob Freeman <[email protected]> >> *Sent:* Thursday, 21 February 2019 12:38 AM >> *To:* AGI >> *Subject:* Re: [agi] openAI's AI advances and PR stunt... >> >> OK, that makes sense Ben. So long as you have a clear picture of how to >> progress the theory beyond temporary expediency, temporarily using the >> state-of-the-art may be strategic. >> >> So long as you are moving forward with some strong theoretical candidates >> too. If we get trapped without theory, we're blind. There are too few >> people with any broad theoretical vision for how to move forward. Too many >> script kiddies just tweaking blindly, viz, the "important step" this thread >> began with. >> >> I'm encouraged that it now appears you are deconstructing grammar and >> resolving it to a raw network level. That Linas is seeing the relevance of >> maths like category theory, which is motivated by formal incompleteness, >> speaks to this realization. (Though he may not be aware of the full import.) >> >> Deep learning does not realize this. It does not realize that formal >> description above the network level will be incomplete. I'm sure that is >> the key theoretical failure holding it back. I wish there were more people >> talking about it. If deep learning realized this they wouldn't still be >> trying to "learn" representations, whether in intermediate layers or other. >> (What was that article recently about the representation "bottle neck" idea >> in deep learning needing to be revised?) >> >> It's actually ironic that deep learning does not realize this idea that >> formal description (above the network) must always be incomplete, because >> it is also the key to the success of deep learning! The whole success of >> distributed representation is due to this. The field moved to distributed >> representation blindly, without theory, just because things started working >> better that way! But you still see articles where people say no-one knows >> why distributed representation works better! The failure of theoretical >> vision is extraordinary. >> >> But if you've deconstructed your dictionaries (throwing out your hand >> coded dictionaries?) and arrived back at the level of observation in a >> sequence network. And done it because of the theoretical realization that >> complete representation above the network level is impossible (or was it >> just an accident, trying to deconstruct symbolism to connectionism, and >> then accidentally noticing the relevance to variational theories of maths?) >> Then your group would be the only ones I've come across who have done (I >> think the Oxford thread of variational formalization, around Coecke et al. >> Grefenstette, were also seduced away by the short term effectiveness of >> deep learning on GPUs.) >> >> We need to keep (or get!) the theoretical vision. >> >> Even given a vision of formal incompleteness, you (and Pissanetzky?) may >> still be lacking a totally clear conception that the key problem is >> assembling elements in new ways all the time. >> >> Still, some focus on assembling elements in different ways (from a >> sequence network) is encouraging. There is scope to move forward. >> >> As a concrete, immediate, idea to explore moving forward, I hope you'll >> look at the idea of using oscillations to structure your sequence network >> representations. For it to be meaningful your networks will need to be >> connected in ways which directly reflect the ideas behind embedding vectors >> (without their linearities.) I don't know if that is true for your >> networks. But given that, implementation should be simple, if practically >> slow without parallel hardware. >> >> -Rob >> >> On Thu, Feb 21, 2019 at 12:03 AM Ben Goertzel <[email protected]> wrote: >> >> It's not that it's hard to feed data into OpenCog, whose >> representation capability is very flexible >> >> It's simply that deep NNs running on multi-GPU clusters can process >> massive amounts of text very very fast, and OpenCog's processing is >> much slower than that currently... >> >> *Artificial General Intelligence List <https://agi.topicbox.com/latest>* > / AGI / see discussions <https://agi.topicbox.com/groups/agi> + > participants <https://agi.topicbox.com/groups/agi/members> + delivery > options <https://agi.topicbox.com/groups/agi/subscription> Permalink > <https://agi.topicbox.com/groups/agi/T581199cf280badd7-M2dceba07a96e44cdbca6ecd0> > -- cassette tapes - analog TV - film cameras - you ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T581199cf280badd7-Meb085d2313898846eabef456 Delivery options: https://agi.topicbox.com/groups/agi/subscription
