Artificial Neural Nets Finally Yield Clues to How Brains Learn

You could favorite the post and add a calendar reminder but I agree it would be a useful HN feature.

Reddit has the remindMe bot for that, HN should give us an exobrain too

You could make an account on ResearchGate and follow the authors of the paper if they're on there, see what they come up with next!

Check back the predictions of 2 years ago and compare to the reality of today.

As to 1, it has already been established that there's no biological plausibility of backprob whatsoever. You can only call the current models "neural" networks in the vaguest sense of analogy. There is significant academic interest in this intersection between AI and neuroscience, to design biologically plausible neural networks (see e.g. spiking networks). I guess the reasons there not very well known in the larger ML community is simply that these approaches don't work that well (as of yet).

Personally I don't believe chasing perfect biological plausibility will be very fruitful (in short term). An algorithm that runs efficiently on wetware will probably not be very efficient on current hardware like gpu's. The reason deep learning is so successful is for a large part that they are very good at exploiting the efficient linear algebra devices we have at our disposal (transformers are only the latest evidence of this).

You're describing the deep dissonance I was feeling a decade ago when I first stepped into AI research. I just kind of always assumed that studying AI would necessarily have a strong focus on how biological intelligence works. And boy was I wrong.

Knowing a bit more now, this gap makes some sense:

1. Neuroscience is really, really hard. Even with the unbelievable recent advances, we're still years away from having a clear understanding of the mechanics of learning and memory.

2. The drift between AI and the broader cognitive sciences started in the 70s, seemingly borne out of pragmatism and the difference in goals between engineer types and scientist types.

not an expert (more like a noob) by any means but:

1) from neuroscience point of view you have cortical columns with layers that are wired to send the input forward but to also propagate feedback. the layers constantly predict what is going to happen (by having neurons fire) and usually it’s the delta between what is predicted and what is coming from the sensory system that drives the reinforcement or the weakening of the connections. this sort of sounds like backpropagation to me (but again i may be super ignorant and would appreciate if you can educate me on this if you know more)

2) technically the “input” in the brain is not continuous. I don’t want to go into semantics but at the end of the day you have molecules, ions etc. so the input/transmission is not continuous. the size of the neurotransmitters is so small that it looks like it’s continuous. my point is that, if you take the current model and you have more computing power you could find out that some things translate between the 2 models (we definitely need a way better model of the neuron, but that’s another story)

3) this is a fair point.

Backprop isn't biologically plausible but predictive coding is and it approximates backprop.

https://arxiv.org/abs/2006.04182

1 backprop is not necessarily the only thing able to perform optim.. it could be something more parallel that try many path at once. a bit like quantum computing.. but we have not just found the algo yet

2 is basically sleeping.

The optimist in me likes to liken it to the difference between birds and planes. Same result but different principle.

For 2, it's worth looking at streaming audio input; we often use rnn's, which update state continually over time, and can make multi scale representations of the audio to help guide decisions. (eg, speaker embeddings which apply to whole sentences, and guide sample-level speech separation.). Audio's a great place to experiment with time; it's much cheaper than dealing with video, and there are lots of great data sets.

Transformers and attention are also tools for responding to the current context with a pretrained network.

A core problem is that if you update the weights during inference, there's seemingly very little garauntee of keeping prior quality high, and the models are already occult magic as it is. Federated learning might be another interesting area to look into for ways to address that issue.

Biological models are going to have specific in built heuristics to learn specific things like language more easily.

Seems to me we should be training DL networks to adjust models to resemble models trained via backprop, but without access to backprop, and see what kinds of heuristics it provides.

You're downvoted but this is correct. It is much like when the analogy was "springs and cogs", and an academic department created in that era "cog-nitive" science, would be the attempt to rotate enough gears in the right way.

Many presumptions are being made here in "computational cognitive science" which preclude including many relevant features of animal learning and animal biology.

Their whole world view is that "patterns of electrical signals in neurons" is where learning takes place. This is very likely to be false: it fails, for example, to note that the brain grows.

Organic growth isn't even scoped here. A brain is a time-evolving dynamic system, whose architecture is at every level dynamic. (& Not least, embedded in a motor system which has a profound effect on its structure ).

I would suggest you are missing something: the article shared a round up of advances in the area of effective biologically plausible learning algorithms. That is an area often missed by the field with its excitement about the advances associated with back propagation.

The title seemed a bit click-bait-y to be too though.