jsenn

No, there really is a neural predisposition to language--not particular sounds, but language. I looked at the Wikipedia page on feral children. Many in fact do learn language when they're rescued. The ones that don't may bark like a dog, but that's not something that requires a language faculty!

That said, there is as you say an example of a boy raised in a room with birds and neglected (not exposed to human language) who did in fact chirp and flap his arms like wings! This isn't relevant to the point I was making above though. If the language faculty were not hard-wired, then normal children would be just as likely to display these behaviours as to speak their parents' native language. As far as I know that happens in not 1 percent or 1 tenth of a percent, but in 0 percent of cases. To me, this indicates a hard-wired tendency, rather than a perfectly general system with some optimizations or soft predispositions.

Actual paper is here: https://www.science.org/doi/10.1126/science.add8091. Very cool result. This is the gist:

> We introduce a theoretical approach that uses temporal coarse-graining akin to Einstein’s kinetic theory of Brownian motion (34). Our averaging scheme is valid in the case of nonjamming mixtures of hard particles, where the dynamics is dominated by pairwise interactions, which is a good approximation for typical pedestrian flows as well as dilute colloids. We recover and unify in a systematic manner the fundamental insights of Helbing and Vicsek (22) as well as Vissers et al. (11) and Klymko et al. (23) by showing that undulation-induced drift and diffusion can both contribute to lane nucleation. We also demonstrate that diffusive processes suppress the formation of very narrow lanes, thereby providing a dynamical selection mechanism that favors the nucleation of lanes of a particular width. We provide explicit formulas for the propensity of a given system to nucleate lanes, and we present a simple approximate rule that lanes emerge at a rate proportional to the product of agent speed, density, and an effective parameter related to the average magnitude of lateral displacement in agent-agent collisions.

So, there are 2 main mechanisms proposed in the literature on lane formation: "drift" and "diffusion". Drift arises from the tendency of people to have a preferred direction to turn to avoid a collision (i.e. right- or left-bias). When facing an opposing flow, repeated conflicts will cause you to slowly drift sideways in your preferred direction. Diffusion is the random Brownian Motion-like jostling that arises from conflicts with opposing pedestrians, which has no directional bias.

In both cases, you're getting jostled around more when there's an opposing flow in front of you than when you're in a lane of people going in the same direction, so there will be a tendency for lanes to form spontaneously ("nucleate"), even from a completely homogeneous initial configuration.

The point of the paper is that they derive, with minimal assumptions, a quantitative relationship that incorporates both mechanisms, and makes specific non-obvious predictions about the shapes and widths of lanes. It also predicts that the rate of lane formation has a simple form (density * speed * average jostle displacement from conflicts with other pedestrians). They do some real-world experiments to validate some of their predictions.

As someone who works on pedestrian simulation software, I don't see how to use this to write the simulation code, but it could be an interesting way to validate the software (i.e. verify that lane formation obeys the predicted relationship).

Something that bothers me about a lot of these papers though is they often perform simulations with periodic boundary conditions. I get why they do that, but it seems to me that that will cause spatial correlations to show up that wouldn't happen in real life. Would the rate of lane formation be different in a system with different boundary conditions?

Another limitation mentioned in the paper that makes this model difficult to apply to crowds of pedestrians is that they assume the interaction between 2 pedestrians depends only on their relative displacement. However, it's known that they key parameter in pedestrian interactions is actually the Time To Collision--i.e., pedestrian collision avoidance is fundamentally anticipatory [1]. Presumably this would complicate the model too much though (now you have to take into account displacement and relative velocity).

[1] http://motion.cs.umn.edu/PowerLaw/

It may be true that the brain can route around damage to be able to produce language using different brain structures than are typically used for that purpose, but what will never happen (as far as I know) is that the brain will spontaneously learn to produce something other than language instead. An example I gave in a sibling thread is a baby raised next to a forest full of birds: the baby will effortlessly learn its parents' native language, but no amount of listening to the birds will cause the baby to produce birdsong instead. GPT-3 would happily produce birdsong the same way it happily produces sequences of chess moves and other stuff in its training data, which means it's doing something different from the human brain.

> But yeah they won't produce birdsong. There's no utility to that. I guess it's a matter of environment.

This is the crux of the issue. GPT-3 would happily learn birdsong instead of human language, just like it has learned to produce snippets of code or a sequence of chess moves or various other things found in its training data. For that reason, it's not by itself useful as a model of human cognition. Which is not to say it isn't interesting, or that studying LLMs won't lead to interesting insights into the human mind--I suspect it will!

Sure, but the type of "language" that includes HTML and C++ is very different from the type of "language" that includes English and French. Chomsky's point is that there's something special about human brains that makes it very easy for them to learn English and French, even with very sparse and poorly-defined inputs, but doesn't necessarily help them learn to produce other types of structured sequences. For example, a baby raised next to a forest will effortlessly learn to speak their parents' native language (you couldn't stop them from doing that!) but won't learn to produce the birdsong they hear coming from the forest. This indicates that there's something special about our brains that leads us to produce English and not birdsong.

Similarly, it's true that some humans can, with lots of cognitive effort, produce HTML and C++, but some can't. Even the ones that can don't do it the same way that they produce English or French.

My understanding is that they work well as arbitrary sequence predictors. For example, they can write HTML markup or C++ code just as easily as they can write English sentences. If you trained them on character sequences other than text from the internet, they would likely perform just as well on that data.

Interesting paper, thanks for linking. I think this is slightly tangential to the Norvig-Chomsky controversy though. What Mumford is saying is that probability and statistics is a more useful basis for modelling natural phenomena than classical logic. I don't think Chomsky would disagree with this! What he disagrees with is the idea (also raised by Mumford) that merely reproducing surface-level aspects of a given natural phenomenon (like human language) with no insight or understanding is sufficient for a scientific theory. It isn't sufficient: one also has to show that the theory cannot produce phenomena that are not natural, and give some insight into what's going on. In fact, it's not even necessary: Galileo advanced physics by imagining a frictionless plane. This doesn't and can't exist in the real world, but it helps us understand what does happen in the real world.

As an example, Mumford talks about how particle filters work much better than any "classical AI" technique for certain tracking tasks. This is true, and particle filters are still important in engineering for this reason, but it's not a theory of how humans accomplish this for the simple reason that particle filters can just as happily do other things that humans don't/can't.

This is true, and interesting, but doesn’t address Chomsky’s concerns. While a LLM has structure, it’s still not as structured—or structured in the same way—as the human language faculty. This is easy to see by observing that LLMs can and do just as easily learn to produce things that are not human language as things that are. For something to count as a model of human language, it has to be able to produce language and not produce non-language.

This is interesting stuff--thanks for writing it out.

> ...super-Turing computational power.

Could you expand on this? The Wikipedia page claims that the deterministic version can be efficiently simulated by a Turing machine.

The Wikipedia page and page you linked claim that NP problems can be solved in linear time. It does point out the caveat that it would require exponential space, but I think even then something's being swept under the rug. Namely, it assumes that an exponential number of basic computational steps can be performed in a single "time step". In a physical cell (and presumably any other instantiation of this model in the real world), repeated application of a rule like a -> aa would require an exponentially-increasing expenditure of energy, which isn't being taken into account in the analysis.

Here's a paper that shows how to augment a recurrent NN with a stack: http://arxiv.org/abs/1503.01007.

I'm not an expert, so some or all of this may be wrong, but my understanding of how it works is this. You have a stack memory, which is just a list of numbers (or maybe a list of N-dimensional vectors). You add the contents of the stack at time t-1 to the input of the network at time t. The hidden layer contains a special number that controls whether to push or pop from the stack (or leave it alone). The values in the stack are determined from the values in the hidden layer at the time they were pushed and the special push/pop control parameter. The new value of the stack is passed into the network on the next time step, and so on.

The tricky thing is that the stack operations have to be differentiable, otherwise there's no way to optimize the push/pop control and the set of weights that determines how to turn the hidden layer into a stack value. So the special control variable in the hidden layer is a continuous value that controls "how much" push vs. pop to do. The value of the top of the stack is a weighted average of the values in the hidden layer and the current value on the top of the stack.

So, if the push vs. pop knob is turned all the way toward pop, then the value of the stack will be whatever the second-from-the-top value is. If it's turned all the way toward push, then the value at the top will be a weighted average of the values in the hidden layer. If it's 50% push and 50% pop, then the value at the top of the stack will be 0.5 * <hidden layer value> + 0.5 * <second-from-top value>, and so on. It's hard to think about what a half push, half pop would actually mean (at least for me), but it apparently works.

As a sibling commenter wrote, there are fancier architectures that provide a full differentiable RAM or other types of gadgets for the network to control.

* The Art of Probability by Hamming. An opinionated, slightly quirky text on probability. Unlike the text used in my university course its explanations were clear and rigourous without being pedantic. The exercises were both interesting and enlightening. The only book in this list that taught skills I've actually used in the real world.

* Calculus by Spivak. This was used in my intro calculus course in university. It's very much a bottom-up, first-principles construction of calculus. Very proof-based, so you have to be into that. Tons of exercises, including some that sneakily introduce pretty advanced concepts not explicitly covered in the main text. This book, along with the course, rearranged by brain. Not sure how useful it would be for self-study though.

* Measurement by Lockhart. I haven't read the whole thing, but have enjoyed working through some of the exercises. A good book for really grokking geometric proofs and understanding "mathematical beauty", rather than just cranking through algebraic proofs step by step.

* Naive Set Theory by Halmos. Somewhat spare, but a nice, concise introduction to axiomatic set theory. Brings you from nothing up to the Continuum Hypothesis. I read this somewhere around my first year in university and it was another brain-rearranger.

I know what you mean about the bad writing, but I don’t think that makes it a terrible book. Unfortunately most sci-fi has flat writing for the same reason most great literature recycles the same ideas and concepts: it’s hard to do everything at once.

Sci-fi generally sacrifices prose quality and character depth to pack in more interesting ideas; great literature generally makes the opposite tradeoff. If you read a sci-fi novel expecting great writing you’ll be frustrated.

Three Body Problem had enough interesting ideas to keep me turning the page, where most sci-fi novels don’t—I prefer short stories for sci-fi exactly because of the generally low quality writing.

If you haven't read Ted Chiang's short stories yet, do so. His first collection (Stories of Your Life and Others) was my favourite, but his more recent one (Exhalation) was also very good (the title story, Exhalation, is gorgeous). Stories of your Life is what the movie Arrival was based on and (as ever) the story was better than the movie.

As far as sci-fi novels, I have pretty generic tastes for this crowd: Dune, The Dispossessed, The Martian. The Crying of Lot 49 is one of my favourite books. Not exactly sci-fi but close.

* Three Body Problem. I've heard of it many times, but never read it. Picked it up a month or two ago, read the whole thing in a few sessions, one of my favourite sci-fi books ever. Will read the next book in the series.

* Immune: A Journey into the Mysterious System that Keeps You Alive. Best pop-sci book I've read in a while. Strikes the right balance between including pertinent information without being overwhelming, explaining things in a digestible way without dumbing down, etc. Inspired me to read a proper immunology primer (How the Immune System Works), which was also very good.

* How Rights Went Wrong. An excellent, level-headed take on the U.S. conception of rights, and how it leads to zero-sum thinking in supreme court cases. The author is so relentlessly reasonable that it's hard not to buy into his argument. Even though this book is about the U.S., it has lots of case studies where it contrasts with various other countries, which helped me understand my country's (Canada's) court system and system of rights better.

Let-downs:

* Seeing Like a State. The first few chapters are interesting, and do a good job of explaining the world-view of the author. Worth buying just for this. However, the last half or two-thirds of the book is a tedious re-hashing of the same ideas through various examples.

* The Precipice: Existential Risk and the Future of Humanity. Many interesting factoids in here (killer asteroids: solved problem). However, the book's central argument failed to convince me. Many of the analyses and probability estimates were disappointingly shallow and hand-wavy, especially for the #1 risk cited in the book--unaligned AI--which he thinks has a 10% chance of ending civilization this century.

You might also like Entangled Life by Merlin Sheldrake. It's focused on fungi rather than trees, but has a similar outlook. Overall it's a bit more down to earth and factual (or at least it seems that way to me, having no special knowledge of the subject matter), but no less eye-opening or fun to read.