Hacker News Reader

Comment by NalNezumi | Hacker News Reader

There's three things I've always been baffled by the lack of interest in the current deep learning based AI field when it comes to parallels with biological brain:

1. Biological plausibility of back prop.

2. The lack of interest/consideration of time-continuous input on network. They are currently discrete and "learning" and inference is done separately. That's not how most organisms work.

3. The lack of consideration how brains (architecture, not weight) grows.

I might just be me missing something but I really have hard time seeing how things would scale in real world (ex: in Robotics applications of Neural nets) without those things addressed

View on HN

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.

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

https://arxiv.org/abs/2006.04182

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).

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.

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.

re 2: are you saying that the amount of signal isn't continuous, or that the time of the signal is not continuous? If the latter, I'm not sure why being made of particles would prevent the signal from being continuous in time. Err, not "a continuous function of time", just in the sense of "not a discrete-in-time thing", unlike things on a computer that are synced with a global clock cycle.

the amount.

also as a side-note, discrete-in-time does not imply a global clock cycle

Thank you for the correction regarding discrete-in-time. I appreciate it.

In the case of the amount of signal being discrete because it is made of particles, this is also true of the floats that computer neural nets are computing with anyway, isn't it?

I think the top level comment (by NalNezumi) was talking about continuous-ness in time anyway, (and, I think the strength of firing of biological neurons is thought to be at least approximately binary anyway? not sure about that.) so I'm not sure I see the bearing of them being unable to be truly continuous in strength on the question?

hmm. went back and reread the comment and... i think you’re right. Saw discrete and did not fully parse everything.

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.

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.

Reply on news.ycombinator.com