With articles like this, I want a "check back in 2 years" reminder, to see how the science shakes out. I'm not smart or informed enough to judge these current events style updates for myself.
Surely this is more a deficit in your calendar software than it is anything else? Think of all of the work you could waste making new tooling, trying to replace something that already exists and serves the sole purpose of reminding you about things
I occasionally did check up on stories, but people rarely do follow-up reporting (especially for things that don't pan out), and Google searches usually just turn up 50 variations on the original story written from the original press release. It's a very unfortunate dynamic.
> Nonetheless, Hinton and a few others immediately took up the challenge of working on biologically plausible variations of backpropagation.
Trying to prove the plausibility of a theory is one approach to science I guess... The researchers have already concluded that brains are simply information processing machines and that AI techniques are a sufficiently representative model to use to learn what brains are like.
I don't see how this research could give us clues to anything other than what is already presumed to be true by the researchers.
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 ).
>Many presumptions are being made here in "computational cognitive science" which preclude including many relevant features of animal learning and animal biology.
This post doesn't actually seem to be citing computational cog-sci, which is usually a bit better about these things. Instead it's addressing the field of biologically plausible (ie: with Hebbian learning rules) deep learning.
> (& Not least, embedded in a motor system which has a profound effect on its structure ).
Sure, but that would expose how weak so much of the present AI work actually is when it comes to studying the motor system.
Other things current AI's are lacking besides growth: embodiment + the social and physical environment, ability to make interventions in the environment, self reproduction, learning from reward signals, autonomy, adaptation, radical open-endedness.
"Patterns of electrical signals in neurons" are just part of the picture. Yes, learning happens there, but learning is fed by signals from the body and environment. It would be silly to focus on the neurons while ignoring the actual content, then start wondering where meaning comes from, and if syntax is enough. Meaning doesn't come from mere neurons, it comes from being an embodied agent.
> Their whole world view is that "patterns of electrical signals in neurons" is where learning takes place
Actually, the mechanisms are chemical processes involving trophic factors (i.e. inputs to those processes) and alteration of the physical structures the signals are transmitted with. You say "the brain grows" but the alteration of its structure to strengthen our weaken transmission and connections in response to signals is how it grows usefully. Which was present in the work described by the article.
An infant brain and an adult brain do not differ by their "strength of connection and trasmission".
It is a hypothesis, and I think a false one, that this is identical to learning.
The dynamical number and arrangement of neuronal tissues is not incidental.
The micro (sub-neurone), medial (neurone) and macro structure (morphology) of the brain is time-varying, not merely its connection patterns (conditioned on fixed micro/medial/macro).
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.
This linear model doesn't seem to reference those memories when considering new memories. You'd need a secondary processing unit for addressing the memories based on the current situation or argument. This is a decent model for how cells develop and how memory cells are maintained. However, it's creation still seems to be very binary, relying on IO rather than variance.
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
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)
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.
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?
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
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.
Really nice to read a round up of advances in biologically plausible algorithms. The field, responding to incentives has, in my subjective opinion, undervalued this class of advancement. I expect once we've wrung the value of of the current techniques that this is the direction advancements will be made in.
"In 2007, some of the leading thinkers behind deep neural networks organized an unofficial “satellite” meeting at the margins of a prestigious annual conference on artificial intelligence. The conference had rejected their request for an official workshop; deep neural nets were still a few years away from taking over AI."
The author almost makes this sound nefarious or short sighted. Workshops and symposia get rejected all the time for a mundane reason: Too many submissions for the available schedule resources at the conference. Important research gets "rejected" all the time, and the selection committees are not saying your topic/research are silly, illegitimate, or fantasy.
I’m rather disappointed with the write-up. The way in which the author outlines these advances don’t really tell me what’s going on. I have some years of experience working with neural networks and I’m reasonably comfortable with the concepts.
Perhaps more surprisingly the mentioned ‘advances’ are not cited!
39 comments
[ 5.4 ms ] story [ 106 ms ] threadTrying to prove the plausibility of a theory is one approach to science I guess... The researchers have already concluded that brains are simply information processing machines and that AI techniques are a sufficiently representative model to use to learn what brains are like.
I don't see how this research could give us clues to anything other than what is already presumed to be true by the researchers.
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 ).
This post doesn't actually seem to be citing computational cog-sci, which is usually a bit better about these things. Instead it's addressing the field of biologically plausible (ie: with Hebbian learning rules) deep learning.
> (& Not least, embedded in a motor system which has a profound effect on its structure ).
Sure, but that would expose how weak so much of the present AI work actually is when it comes to studying the motor system.
Other things current AI's are lacking besides growth: embodiment + the social and physical environment, ability to make interventions in the environment, self reproduction, learning from reward signals, autonomy, adaptation, radical open-endedness.
"Patterns of electrical signals in neurons" are just part of the picture. Yes, learning happens there, but learning is fed by signals from the body and environment. It would be silly to focus on the neurons while ignoring the actual content, then start wondering where meaning comes from, and if syntax is enough. Meaning doesn't come from mere neurons, it comes from being an embodied agent.
Actually, the mechanisms are chemical processes involving trophic factors (i.e. inputs to those processes) and alteration of the physical structures the signals are transmitted with. You say "the brain grows" but the alteration of its structure to strengthen our weaken transmission and connections in response to signals is how it grows usefully. Which was present in the work described by the article.
It is a hypothesis, and I think a false one, that this is identical to learning.
The dynamical number and arrangement of neuronal tissues is not incidental.
The micro (sub-neurone), medial (neurone) and macro structure (morphology) of the brain is time-varying, not merely its connection patterns (conditioned on fixed micro/medial/macro).
The title seemed a bit click-bait-y to be too though.
Maybe this will help.
https://ieeexplore.ieee.org/document/9325353
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
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).
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.
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.
also as a side-note, discrete-in-time does not imply a global clock cycle
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?
https://arxiv.org/abs/2006.04182
2 is basically sleeping.
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.
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.
The author almost makes this sound nefarious or short sighted. Workshops and symposia get rejected all the time for a mundane reason: Too many submissions for the available schedule resources at the conference. Important research gets "rejected" all the time, and the selection committees are not saying your topic/research are silly, illegitimate, or fantasy.
Someone tried making a computer like this decades ago.
Ex-Machina had a plot device like this too, to make the robot’s transistor based brain.
Perhaps more surprisingly the mentioned ‘advances’ are not cited!