Ah, I see. Yes, I saw that reference. But I interpreted it only to mean the general forms of a grammar. Do you think he means the mechanism must actually be a grammar?
In the earlier papers I interpret him to be saying, if language is a grammar, what kind of a grammar must it be? And, yes, it seemed he was toying with actual physical mechanisms relating to levels of brain structure. Thalamo-cortical loops? The problem with that is, language doesn't actually seem to be any kind of grammar at all. It's like saying if the brain had to be an internal combustion engine, it might be a Mazda rotary. BFD. It's not an engine at all. I don't know if the authors realized that. But surely that's the point of the HNet paper. That something can generate the general forms of a grammar, without actually being a grammar. I guess this goes back to your assertion in our prior thread that "learning" needs to be constrained by "physical priors" of some kind (was it?) Are there physical "objects" constraining the "learning", or does the "learning" vaguely resolve as physical objects, but not quite? I don't think vague resemblance to objects means the objects must exist, at all. Take Kepler and the planets. If the orbits of planets are epicycles, which epicycles would they be? The trouble is, it turns out they are not epicycles. And at least epicycles work! That's the thing for natural language. Formal grammar doesn't even work. None of them. Nested stacks, context free, Chomsky hierarchy up, down, and sideways. They don't work. So figuring out which formal grammar is best, is a pointless exercise. None of them work. Yes, broadly human language seems to resolve itself into forms which resemble formal grammar (it's probably designed to do that, so that it can usefully represent the world.) And it might be generally useful to decide which formal grammar it best (vaguely) resembles. But in detail it turns out human language does not obey the rules of any formal grammar at all. It seems to be a bit like the way the output of a TV screen looks like objects moving around in space. Yes, it looks like objects moving in space. You might even generate a physics based on the objects which appear to be there. It might work quite well until you came to Road Runner cartoons. That doesn't mean the output of a TV screen is actually objects moving around in space. If you insist on implementing a TV screen as objects moving around in space, well, it might be a puppet show similar enough to amuse the kids. But you won't make a TV screen. You will always fail. And fail in ways very reminiscent of the way formal grammars almost succeed... but fail, to represent human language. Same thing with a movie. Also looks a lot like objects moving around on a screen. But is it objects moving on a screen? Different again. Superficial forms do not always equate to mechanisms. That's what's good about the HNet paper for me. It discusses how those general forms might emerge from something else. The history of AI in general, and natural language processing in particular, has been a search for those elusive "grammars" we see chasing around on the TV screens of our minds. And they all failed. What has succeeded has been breaking the world into bits (pixels?) and allowing them to come together in different ways. Then the game became how to bring them together. Supervised "learning" spoon fed the "objects" and bound the pixels together explicitly. Unsupervised learning tried to resolve "objects" as some kind of similarity between pixels. AI got a bump when, by surprise, letting the "objects" go entirely turned out to generate text that was more natural than ever! Who'd a thunk it? Letting "objects" go entirely works best! If it hadn't been for the particular circumstances of language, pushing you to a "prediction" conception of the problem, how long would it have taken us to stumble on that? The downside to that was, letting "objects" go entirely also doesn't totally fit with what we experience. We do experience the world as "objects". And without those "objects" at all, LLMs are kind of unhinged babblers. So where's the right balance? Is the solution as LeCun, and perhaps you, suggest (or Ben, looking for "semantic primitives" two years ago...), to forget about the success LLMs had by letting go of objects entirely. To repeat our earlier failures and seek the "objects" elsewhere. Some other data. Physics? I see the objects, dammit! Look! There's a coyote, and there's a road runner, and... Oh, my physics didn't allow for that... Or could it be the right balance is, yes, to ignore the exact structure of the objects as LLMs have done, but no, not to do it as LLMs do by totally ignoring "objects", but to ignore only the internal structure of the "objects", by focusing on relations defining objects in ways which allow their internal "pattern" to vary. That's what I see being presented in the HNet paper. Maybe I'm getting ahead of its authors. Because that is the solution I'm presenting myself. But I interpret the HNet paper to present that option also. Cognitive objects, including "grammar", can emerge with a freedom which resembles the LLM freedom of totally ignoring "objects" (which seems to be necessary, both by the success of LLMs at generating text, and by the observed failure of formal grammars historically) if you specify them in terms of external relations. Maybe the paper authors don't see it. But the way they talk about generating grammars based on external relations, opens the door to it. On Fri, May 24, 2024 at 10:12 PM James Bowery <jabow...@gmail.com> wrote: > > > > On Thu, May 23, 2024 at 9:19 PM Rob Freeman <chaotic.langu...@gmail.com> > wrote: >> >> ...(Regarding the HNet paper) >> The ideas of relational category in that paper might really shift the >> needle for current language models. >> >> That as distinct from the older "grammar of mammalian brain capacity" >> paper, which I frankly think is likely a dead end. > > > Quoting the HNet paper: >> >> We conjecture that ongoing hierarchical construction of >> such entities can enable increasingly “symbol-like” repre- >> sentations, arising from lower-level “statistic-like” repre- >> sentations. Figure 9 illustrates construction of simple “face” >> configuration representations, from exemplars constructed >> within the CLEVR system consisting of very simple eyes, >> nose, mouth features. Categories (¢) and sequential rela- >> tions ($) exhibit full compositionality into sequential rela- >> tions of categories of sequential relations, etc.; these define >> formal grammars (Rodriguez & Granger 2016; Granger >> 2020). Exemplars (a,b) and near misses (c,d) are presented, >> initially yielding just instances, which are then greatly re- >> duced via abductive steps (see Supplemental Figure 13). > > Artificial General Intelligence List / AGI / see discussions + participants + > delivery options Permalink ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T682a307a763c1ced-M9f8daceca7b091a0b823481d Delivery options: https://agi.topicbox.com/groups/agi/subscription
