18 ms·
Transformers seem to mimic parts of the brain
- _s_a_m_ 4y agoNonsense, as always from AI research. Not even single neurons are properly understood and hardly possible to simulate. Don't even get me started about all this marketing bullshit.
- vanderZwan 4y agoSincere question inspired purely by the headline: how many important ML architectures aren't in some way based on some proposed model of how something works in the brain? (Not intended as a flippant remark, I know Quanta Magazine articles can generaly safely be assumed to be quality content, and that this is about how a language model unexpectedly seems to have relevance for understanding spatial awareness)
- edgyquant 4y agoIt is my opinion that pretty much all architectures already exist in the brain for some use or another. Otherwise we wouldn’t be able to reason about them
- dunefox 4y ago> Otherwise we wouldn’t be able to reason about them That's a bold claim.
- hackernewds 4y agoI agree but not for the "otherwise" reasoning I think it speaks to the complexity of the brain almost like every combination of numbers exists in pi
- coldtea 4y agoWe can reason about all kinds of things that don't exist in the brain...
- robotresearcher 4y agoWhen we're reasoning about them, where are they?
- coldtea 4y agoStill outside. It's our thoughts about them, models, and concepts that are in the brain. Plus, even if it was true, this wasn't the parent's point (that things exist in the brain while we're reasoning about them). His point (also wrong) was that different architectures must exist as structures in the brain (not as concepts we think and memories etc., but as parts of brains matter organization and wiring) for us to be able to reason about it.
- edgyquant 4y agoFor one You can’t actually prove this. We use brains to experience everything so we (humans as a species) can never know.
- coldtea 4y agoWhat does that even mean? We can reason about fusion reactors, yet there's no such mechanism in our brains... How's that for a proof?
- tomrod 4y agoI disagree. No one built cars in 5000 BC but the confluence of ideas then have led to cars now.
- edgyquant 4y agoThis is a non-sequitur
- tomrod 4y agoOn the contrary, I view it as a counterexample to "all architectures already exist in the brain for some use or another," which disproves your point. Let's not make the mistake of a fallacy fallacy here! Perhaps you would like to expound or clarify your point to rule out the edge case of cars not existing in 5000 BC, but the models to derive cars 5000 years later suddenly came into being?
- abeppu 4y agoI think the response to this has two prongs: - Some families of ML techniques (SVMs, random forests, gaussian processes) got their inspiration elsewhere and never claimed to be really related to how brains do stuff. - Among NNs, even if an idea takes loose inspiration from neuroscience (e.g. the visual system does have a bunch of layers, and the first ones really are pulling out 'simple' features like an edge near an area), I think it's relatively uncommon to go back and compare specifically what's happening in the brain with a given ML architecture. And a lot of the inspiration isn't about human-specific cognitive abilities (like language), but is really a generic description of neurons which is equally true of much less intelligent animals.
- soraki_soladead 4y ago> I think it's relatively uncommon to go back and compare specifically what's happening in the brain with a given ML architecture. Less common but not unheard of. Here's one example, primarily on focused on vision: http://www.brain-score.org/ http://www.brain-score.org/ DeepMind has also published works comparing RL architectures like IQN to dopaminergic neurons. The challenge is that its very cross-disciplinary and most DL labs don't have a reason to explore the neuroscience side while most neuro labs don't have the expertise in DL.
- westurner 4y agoIs it necessary to simulate the quantum chemistry of a biological neural network in order to functionally approximate a BNN with an ANN? A biological systems and fields model for cognition: Spreading activation in a dynamic graph with cycles and magnitudes ("activation potentials",) that change as neurally-regulated heart-generated electron potentials (and,) reverberate fluidically with intersecting paths. And a partially extra-cerebral induced field which nonlinearly affects the original signal source through local feedback; Representational shift. Representational shift: "Neurons Are Fickle. Electric Fields Are More Reliable for Information" (2022) https://neurosciencenews.com/electric-field-neuroscience-20178/amp/ https://neurosciencenews.com/electric-field-neuroscience-201... Spreading activation: https://en.wikipedia.org/wiki/Spreading_activation https://en.wikipedia.org/wiki/Spreading_activation Re: 11D (11-Dimensional) biological network hyperparameters, ripples in (hippocampal, prefrontal,) association networks: https://news.ycombinator.com/item?id=18218504 https://news.ycombinator.com/item?id=18218504 M-theory String theory is also 11D, but IIUC they're not the same dimensions Diffusion suggests fluids, which in physics and chaos theory suggests Bernoulli's fluid models (and other non-differentiable compact descriptions like Navier-Stokes), which are part of SQG Superfluid Quantum Gravity postulates. Can e.g. ONNX or RDF with or without bnodes represent a complete connectome image/map? Connectome: https://en.wikipedia.org/wiki/Connectome https://en.wikipedia.org/wiki/Connectome
- fzliu 4y agoSaying that neural networks are "similar to" or that they "mimic" the human brain can be misleading. Today's architectures are the byproduct of years of research and countless GPU-hours dedicated to training and testing architecture variants. Many neuroscience-based architectures that mimic the brain better than transformers end up performing much worse. The Quanta article is overall pretty reasonable, but I've unfortunately seen other news outlets regurgitate this kind of blanket statement for the better part of a decade. The very first models were perhaps 100% inspired by the brain, but today's ML research more or less follow a "whatever works best" principle.
- trhway 4y agoThe convolution kernels in the first layers of AlexNet and all its DL image processing descendants converge to the Gabor filters (or some variation of) which are the response functions of the neurons in the first layer of the visual cortex. About 15 years before AlexNet there were works showing that such type of filter is kind of mathematically optimally encoding for the feature based image processing. (So, theoretically one could have just pre-generated the first layers in the net and use them fixed thus cutting significant time/effort on the training - i myself wanted to do it 20 years ago, yet just didn't get to it :) I'm pretty sure that for middle layers in image DL as well as for transformers in lang we have a kind of similar optimality, ie. something like maximum entropy filter(separator/aggregator at higher levels?) at a given level of granularity/scale like Gabors at the first feature level.
- data_maan 4y agoCould you provide references for your statements > the first layers of AlexNet and all its DL image processing descendants converge to the Gabor filters (or some variation of) and > 15 years before AlexNet there were works showing that such type of filter is kind of mathematically optimally encoding for the feature based image processing ?
- trhway 4y agoFor the first - you can just look at the original AlexNet paper. The kernels are unmistakably strikingly Gabor-like. Some differences, like cross-color, are kind of giving rise to possibly interesting questions - is it improvement or deficiency(i.e. more training would correct) over biology? or may be it is just real-valued projection from the [plausible] fact that the optimal is complex-valued? >? I don't have that specific reference i had in mind that was published 15-20 years ago, yet you can trace that line of thought development through the works like these for example (there have been a bunch of them in the 199x and into 200x) : 1990 - https://opg.optica.org/josaa/abstract.cfm?uri=josaa-7-8-1362 https://opg.optica.org/josaa/abstract.cfm?uri=josaa-7-8-1362 1998 - https://pubmed.ncbi.nlm.nih.gov/12662821/ https://pubmed.ncbi.nlm.nih.gov/12662821/
- macrolocal 4y agoTransformer networks have deeper connections to dense associative memory. For example, the update rule to minimize the energy functional of these Hopfield networks converges in a single iteration and coincides with the attention mechanism [1]. [1] https://arxiv.org/abs/1702.01929 https://arxiv.org/abs/1702.01929
- macrolocal 4y agoMore accessible references: https://mcbal.github.io/post/an-energy-based-perspective-on-attention-mechanisms-in-transformers/ https://mcbal.github.io/post/an-energy-based-perspective-on-... (Modern continuous Hopfield networks section) https://arxiv.org/abs/2008.02217 https://arxiv.org/abs/2008.02217 Note that the connection to Hebbian learning hinges on the softmax function, in particular its exponential!
- rdedev 4y agoI think at some point similarities will naturally emerge. Smart moves in design space. That being said these similar designs will probably be minuscule compared to the overall architecture
- mtlmtlmtlmtl 4y agoAlphaGo is a pretty good example here. It uses a neural net for evalutation, and that's vaguely inspired by the brain, sure. But it employs a Monte Carlo based game tree search which is probably very different from how humans think. In addition it learns by iterated amplification and distillation: it plays games against itself, where one player gets more time, hence will be a stronger player(amplification). The weaker player then uses this stength differential as a fitness function to learn(distillation). Rinse and repeat. That's really nothing like how humans learn these games. While playing stronger players and evaluating is a huge part of becoming stronger, there's also a lot of targeted exercises, opening/endgame theory, etc. Humans can't really do that type of training at all.
- kadoban 4y ago> But it employs a Monte Carlo based game tree search which is probably very different from how humans think. It's not _that_ different from how humans play. We have pattern matching that points out likely places, we read out what happens and try to evaluate the result. Humans are just less methodical at it really.
- svnt 4y agoParent said “different from how humans think”, not play, which seems key. Your description is very broad. These machines don’t seem to carry narratives or plans yet (if they would benefit from them or be encumbered by them seems to be an open question). Watching the machines play they have zero inertia. If the next opportunity means a completely inverted play strategy has a marginally better chance of winning, they will switch their entire approach. Humans don’t typically do this, although having learned from machines that it can produce better outcomes perhaps we will start moving away from this local maximum.
- kadoban 4y ago> Watching the machines play they have zero inertia. If the next opportunity means a completely inverted play strategy has a marginally better chance of winning, they will switch their entire approach. In Go, especially at the high level, this isn't that far outside of the norm. In particular, you see players play in other areas (tenuki) at what to a weaker player would look like pretty random times, depending on what's most urgent or biggest. Computer go players aren't too chaotic. They're just _very_ good at some things that are already high-level-player traits. A computer will just give you what you want, but suddenly it's just not actually that good. It feels like Bruce Lee's flow/adaptation based fighting style applied to a go board.
- adamnemecek 4y agoThe underlying idea is the idea of fixed points (aka spectra, diagonalizations, embedding, invariants). By fixed point I mean something like the "Lawvere's fixed point theorem". https://ncatlab.org/nlab/show/Lawvere%27s+fixed+point+theorem https://ncatlab.org/nlab/show/Lawvere%27s+fixed+point+theore... Dennis Gabor's Holonomic brain theory also indicates something like that https://en.wikipedia.org/wiki/Holonomic_brain_theory https://en.wikipedia.org/wiki/Holonomic_brain_theory I have a linkdump on this https://github.com/adamnemecek/adjoint https://github.com/adamnemecek/adjoint I also have a discord https://discord.gg/mr9TAhpyBW https://discord.gg/mr9TAhpyBW
- data_maan 4y agoWhy do you invocate a categoric theorem when the theory that is discussed on Quanta has a manifestly non-category-theoretic flavor?...
- adamnemecek 4y agoBecause it ties it into a larger ecosystem.
- data_maan 4y agoWell... this is kind of a problem with category theory, isn't it? Almost every can be categorified (which, I know, is insanely popular these days). But outside of pure mathematics I'm having trouble seeing why such a categorification is really useful. Yes, it can be nice, but it often seems more like linguistic exercise than providing crucial insight.
- ericbarrett 4y agoThe first thing I thought of after reading this was the "mind palace" technique used since antiquity for memorizing things: https://en.m.wikipedia.org/wiki/Method_of_loci https://en.m.wikipedia.org/wiki/Method_of_loci
- hnplj 4y ago
- origin_path 4y agoI do wonder about the ethical issues that people are going to start facing in, I don't know, maybe 10-20 years? Maybe even sooner. These new DNNs create very human like outputs. And their structure is rather similar to the brain. Now we are learning that maybe they're even more brainlike than we thought. At what point do we cross a threshold here and encounter a non-trivial number of people who argue that a sufficiently large model actually is a brain, and therefore deserves rights? Blake Lemoine went there already but I wonder if that's going to be an isolated incident or a harbinger of things to come. It feels weird. On one hand, I know that these things are just programs. On the other hand, I also know that our brains are just molecules and cells. At the same time I feel that intelligent creatures deserve rights and the exact way the underlying brain works doesn't seem especially important. Especially if people start getting brain augmentations at scale a la Neuralink. At what point do you say the line is crossed?
- heyitsguay 4y agoThere are no actual ethical issues with contemporary ML architectures (including transformers) being too conscious or brainlike, it's just laypeople and demagogue chatter. It could be an issue in the future but only with computational systems that are utterly unlike the ones being created and used today. Actual practitioners are tired of the debate because getting to an informed answer requires grokking some undergraduate-level math and statistics, and nobody seems particularly willing to do that
- data_maan 4y ago> nobody seems particularly willing to do that This. The amount of mathematical illiteracy is staggering in ML.
- heyitsguay 4y agoThe professionals I've met actually working in ML R&D have basically all been very technically competent, including in mathematics. It's more the people who talk a lot about AI in grandiose and anthropomorphized terms that I was referring to.
- data_maan 4y agoThis whole area of research is more hype than substance. First note that what they tout as being "the brain" is actually just a very very simplified model of the brain. If you really want to model the brain there's a hierarchy of ever more complicated spiking neural nets. Even simulating a single synapse in full detail can be challenging. Having said that, the fact that some models used in practice have been found to be equal to a neuroscientific model is not really that impressive since it does neither explain the inner working of the brain, nor modern ML models. Unfortunately, Quanta magazine editors are riding too much on the hype wave to notice that. Note also that Whittingon's other work on predictive coding networks also is not really solid. It was a pretty irritaring experience to read some of his work. That makes me skeptical of how rigorous his claims are in this case.
- westoncb 4y ago> Even simulating a single synapse in full detail can be challenging I wonder how common it is for serious commentators on NN / brain relationship to hold this supposition that in order for the two to be called... let's say "functionally similar/equivalent", that there would have to be some kind of structural equivalence in their most basic parts. Neurons (and their synaptic connections etc.) developed in a biochemical substrate which is going to bring a certain amount of its own representational baggage with it, i.e. elements which are loosely incidental to "what really matters" in creating the magic of the brain—and we should not expect those features to reappear in artificial NNs (as they are by definition incidental): bringing them up could only establish a trivial non-equivalence imo. I'd like to see more discussions about NN / brain relationship mentioning which level/kind of equivalence they're refuting/confirming when refuting/confirming.
- lossolo 4y agoThat has never been thought by anyone but computer scientists who never looked at a biology textbook. To begin approximating what a lone spherical synapse would actually do you'd need to solve 2^n coupled second order differential equations where n is the number of ions used. That is before you throw in things like neuro transmitters and the physical volume of a cell. Simulating a single neuron accurately is beyond any super computer today. The question is how inaccurately can we simulate one and still get meaningful answers. Then how we do it 100e9 more times. source: https://news.ycombinator.com/item?id=32407028 https://news.ycombinator.com/item?id=32407028 There is interesting discussion there.
- intrasight 4y agoNext headline of course will be that our brains are now mimicking parts of Transformers. "This is your brain on Stable Diffusion"
- midoridensha 4y agoThe Transformers: more than meets the eye