[HN Gopher] Artificial brains help understand real brains
___________________________________________________________________
Artificial brains help understand real brains
Author : Jeff_Brown
Score : 44 points
Date : 2023-06-09 19:43 UTC (3 hours ago)
(HTM) web link (www.economist.com)
(TXT) w3m dump (www.economist.com)
| phkahler wrote:
| For me the most interesting parallel is from (I think) GANs, and
| other generative AIs. This is similar to the idea in psychology
| that we are really doing a lot of projection with some correction
| based on sensory input - as opposed to actually perceiving
| everything around us.
|
| Also, real synapses are one of the most abundant features of real
| brains and are the direct inspiration for NN weights. I'm not
| sure the artificial brains help understand real ones, but they do
| seem to validate some ideas we have about real ones.
| kuprel wrote:
| The left hand in that photo should be a humanoid robot hand.
| Didn't actually read the article
| cosmojg wrote:
| As a computational neuroscientist, I find myself both terribly
| disappointed and unfortunately reminded of Gell-Mann amnesia.
| eep_social wrote:
| https://archive.is/ryTWs
| viableSprocket1 wrote:
| Oversimplifying for brevity (and there is definitely more nuance
| to this). This is basically the modeling approach:
|
| 1. Have a biological brain do a task, record neuronal data + task
| performance
|
| 2. Copy some of those biological features and implement in an ANN
|
| 3. Tune the many free parameters in the ANN on task performance
|
| 4. Show that the bio-inspired ANN performs better than SOTA
| and/or shows "signatures" that are more brain-like.
|
| The major criticisms of Yamins' (and similar) groups are either
| that correlation != causation, or correlation != understanding,
| or that it is tautological (bio-inspired ANNs will be more
| biological). I'm not sure how seriously this work is taken vs.
| true first principles theory.
| adamnemecek wrote:
| [flagged]
| civilized wrote:
| Isn't renormalization a technique by which divergent integrals
| can yield finite results? How would it be a mechanism by which
| anything operates?
| adamnemecek wrote:
| Read up on renormalization groups.
|
| There's this very fundamental problem in a lot of sciences,
| given some phenomena, find the patterns to compress the
| phenomena without knowing all the patterns a priori.
|
| It's like wavelet decomposition where your wavelets are
| updates as new data is coming in.
|
| That's renormalization.
| bigdict wrote:
| The fundamental problem is that you keep coming up with
| random bullshit to post on HN and you keep inviting people
| to your discord basement to discuss it.
| adamnemecek wrote:
| > The fundamental problem
|
| Problem for whom?
|
| > keep coming up with new random bullshit
|
| Elaborate, what's bullshit?
|
| > you keep inviting people to your discord basement to
| discuss it.
|
| Want to come? I can explain it to you.
| TaupeRanger wrote:
| Not really. The only real example given in the article is when
| you hook someone up to an fMRI machine, collect data about how
| the brain looks when it sees a certain image, and then have a
| computational statistics program (NOT an artificial brain, in any
| sense) do some number crunching and output the most likely thing
| it's looking at based on things you specifically trained it to
| recognize beforehand. We learn precisely nothing from this, no
| medical or computer science advances are made from it, and it
| doesn't remotely support the title of the article.
| moffkalast wrote:
| > We learn precisely nothing from this
|
| That's literally never true about anything.
| cosmojg wrote:
| >> We learn precisely nothing from this
|
| > That's literally never true about anything.
|
| But if this were true, data compression would be impossible.
| moffkalast wrote:
| OK you got me there I guess, but even learning that we've
| learned nothing is learning in itself. Otherwise the act of
| compressing wouldn't give any more information.
| foobarqux wrote:
| What have we learned specifically in this case?
| HarHarVeryFunny wrote:
| I think the Economist article is exactly right - that despite
| the massive differences between ANNs and the brain, ANNs are
| indeed highly suggestive of how some aspects of the brain
| appear to work.
|
| People can criticize the shortcomings of GPT-4, but it's hard
| to argue that it's at least capable of some level of reasoning
| (or functionally equivalent if you object to that word!). It's
| not yet clear exactly how a Transformer works other than at
| mechanical level of the model architecture (vs the LLM "running
| on" the architecture), but we are at least starting to glean
| some knowledge of how the trained model is operating...
|
| It seems that pairs of attention heads in consecutive layers
| are acting in coordination as "induction heads" that in one
| case are performing a kind of analogical(?) A'B' => AB match-
| and-copy type of operation. The induction head causes a context
| token A to be matched (via "attention" key query) with an
| earlier token A' whose following token B' then causes related
| token B to be copied to the residual stream at position
| following A.
|
| This seems a very basic type of operation, and no doubt there's
| a lot more interpretability research to be done, but given the
| resulting reasoning/cognitive power (even in absense of any
| working memory or looping!), it seems we don't need to go
| looking for overly complex exotic mechanisms to begin to
| understand how the cortex may be operating. It's easy to
| imagine how this same type of embedded key matching might work
| in the cortex, perhaps with cortical columns acting as complex
| pattern matchers. Perhaps the brain's well known ~7 item
| working memory corresponds to a "context" of sorts that is
| updated in same way as induction heads update the residual
| stream.
|
| Anything I've written here about correspondence between
| transformer and cortex is of course massive speculation, but
| the point is that the ANN's operation does indeed start to
| suggest how the brain, operating on similar sparse/embedded
| representations, may be working.
| amelius wrote:
| But (how) does the human brain perform backpropagation?
| HarHarVeryFunny wrote:
| It probably doesn't use backpropagation of gradients.
| Instead, the cortex appears to be a prediction engine that
| uses error feedback (perceptual reality vs prediction) to
| minimize prediction errors in a conceptually similar type
| of way. If every "layer" (cortical patch) is doing it's own
| prediction and receiving feedback, then you don't need any
| error propagation from one layer to the next.
| marcosdumay wrote:
| You just can't go changing the layers of a neural network
| independently of each other if you are doing guided
| optimization.
|
| For a start, it's not even a given if increasing some
| weight will increase or decrease the result. The neurons
| are all tightly coupled.
|
| You can do it on unguided optimization, what is one of
| the reasons I strongly suspect our brains use something
| similar to simulated annealing.
| HarHarVeryFunny wrote:
| Sure, we don't know the exact details (Geoff Hinton spent
| much of his career trying to answer this question), but
| at the big picture level it does seem clear that the
| cortex is a prediction engine that minimizes prediction
| errors by feedback, and most likely does so in a
| localized way. Exactly how these prediction updates work
| is unknown.
|
| Could you expand a bit on how you think simulated
| annealing could work?
| marcosdumay wrote:
| It's more that most other things couldn't.
|
| Simulated annealing works on a localized way (and over
| backward links), and is easy to appear in organic
| structures. It also can use a "neighborhood error signal"
| that is what our dispersion-based one looks like to me,
| and is completely coherent with the error distribution
| people have when learning a new physical movement
| (although this is compatible with a lot of other
| mechanisms).
| canjobear wrote:
| Approximately, via predictive coding
| https://arxiv.org/abs/2006.04182
| foobarqux wrote:
| This is just false. Outside of the visual cortex there isn't
| any evidence that brains work anything like GPT-4 or neural
| nets. Producing the same outputs isn't evidence of anything.
|
| At best you have just stated a hypothesis of how the brain
| might work, not any actual evidence supporting it.
| HarHarVeryFunny wrote:
| Sure - you can't just look at how an ANN works and assume
| that's how the brain does that too, but the ANN's operation
| can act as inspiration to suggest or confirm the way the
| brain might be doing something.
|
| It seems neuroscientists are good at discovering low level
| detail and perhaps not so good in general (visual cortex
| being somewhat of an exception) at putting the pieces
| together to suggest high level operations. ANNs seem
| complementary in that while their low level details are
| little like the brain, the connectionist architectures can
| be comparable, and we do know the top down operation (even
| though interpretation is an issue, more for some ANNs than
| others). If we assume that the cortex is doing some type of
| prediction error minimization then it's likely to have
| found similar solutions to an ANN in cases where problem
| and connectivity are similar.
___________________________________________________________________
(page generated 2023-06-09 23:00 UTC)