I wrote a lay summary of this paper that compares and contrasts with the whole brain imaging and stimulation literature for those interested in a lighter read!
Great summary! At the end you mention the difficulty extrapolating beyond an HH neuron model. I think curious readers will find the work of Jim Smith (https://fcrc.acm.org/plenary-speakers/james-e-smith-plenary) interesting in this regard. His work starts with the possible information representation scheme (temporal coding <=> binary coding) and a compute unit (SRM0 neuron <=> transistor) and builds up the equivalent of Boolean logic/algebra from there.
As opposed to a neuroscientist understanding a processor, Jim is a computer architect using his techniques to understand the brain.
In contrast, I was underwhelmed. Neuroscientists typically try to develop tasks that create specific contrasts that highlight aspects of specific sorts of computations. They didn't do anything like that.
For example, in the field I work in, we very very reliably get signal that track reward prediction errors in the striatum (e.g. BOLD response in fmri) during reward learning tasks.
>For example, in the field I work in, we very very reliably get signal that track reward prediction errors in the striatum (e.g. BOLD response in fmri) during reward learning tasks.
Yeah, but that's a property of how the BOLD response correlates with the task structure, not a map of what computations the brain actually does.
Yeah, ok, I'm shortening things by a lot, but neuroimaging really is quite fraught in terms of what sorts of computational-level inferences we can draw from task-based experiments.
In these experiments, the BOLD signal reliably scales with reward prediction error, which is a computation, the difference between the reward expected and received. In short, we fit reinforcement learning models to in-task behavior, then correlate parameters of that model to neuroimaging acquired during that behavior.
fMRI does have quite a few limitations; BOLD signal is believed to reflect the amount of work being done, loosely, speaking, but it is not a computational signal in of itself; for reinforcement learning models we'd like to measure dopamine itself. Two decades ago, the pioneering work of Peter Dayan and Read Montague established that dopamine neurons report a prediction error signal (see: http://static.vtc.vt.edu/media/documents/SchultzDayanMontagu...), but recording of neurotransmitter release in humans is hard, and frankly, long thought impossible (given ethical constraints). However, recent work by Read Montague has done just that. See: https://www.pnas.org/content/113/1/200 Keep tuned.
Do you have any insight into why it took so long to do this in hmans? Fast-scan cyclic voltammeter was done in animals ~15 years ago by Wightman's group.
ethics, mainly. unlike macaques (etc.) we don't just stick probes in people. however, Parkinson's and epilepsy patients are getting implants for deep brain stimulation; Montague's group is working with hospitals performing these procedures.
I wasn't sure if there was some advance in voltammeter techniques, or if it just took forever to find a willing patient and convince ethics to green-light it.
DBS has been FDA-approved since the late 1990s (and even longer under IRB/IND stuff), so there have been lots of opportunities to stick probes in humans' heads.
There actually is a technical advance here; I just don't know enough about it to discuss intelligently (sorry). I would reach out to the researchers involved if interested.
"We present a computational model that predicts the functional magnetic resonance imaging (fMRI) neural activation associated with words for which fMRI data are not yet available. This model is trained with a combination of data from a trillion-word text corpus and observed fMRI data associated with viewing several dozen concrete nouns. Once trained, the model predicts fMRI activation for thousands of other concrete nouns in the text corpus,with highly significant accuracies over the 60 nouns for which we currently have fMRI data."
I think their lesion approach is a little awkward and probably doesn't map that well between the fields. Removing a transistor, IMHO, is more akin to removing an eye or some other body part. Yeah, it allows one to determine if that transitor was important, but its impact on behavior is due to damaged hardware. Plus, the software probably can't be written to compensate for missing transistors.
Instead of removing transistors from the CPU to determine if it was important to the "behavior" of the game, I think it would be more analogous to delete portions of memory and see if that alters game behavior. Even then, I think it's much more likely to cause a crash if done without some a priori information.
This article keeps coming up over and over; I suspect that it appeals to the HN crowd's background, but really, the article, and others like it, underestimate the sophisticated behavioral methods we cognitive neuroscientists employ, as well as the value of convergent lines of evidence taken from molecular to systems neuroscience.
A neural network doesn't understand anything but sure it could predict the program output sequence based on the input sequence.
Would there be any useful use case to such a technology?
There's no technological use for neuroscience analysis tools to model a microchip, nor for NLP modeling source code, but that isn't the point. It's a potentially useful target to hone the methodology itself on. I actually don't think something like BERT would do so well at learning to essentially execute code.
The counterexample is for a transformer with context width of N tokens, to try to model "def f(x) { ... }; <at least N more tokens here>; def g(x) { y = f(x) ...};". In this case, in trying to evaluate "g(4)", f appears as the literal token "f" and nothing more. To be able to learn to evaluate g, a model would need to be able to do (at least) distant coreference resolution.
Maybe for simple programs. Of course, an arbitrarily deep neural net can emulate any compiler (a function from plain text to machine instruction sequences). But predicting output is limited by the halting problem and complexity.
Whether a neural net could be trained to emulate a compiler seems like the more pertinent question.
I've had the displeasure of trying to explain classical learning hardness results to someone who kept repeating back to me "but RNNs are Turing complete!" If anything, the expressive power of the model you're trying to train pushes back against your efforts to train it to do something useful.
GPT-2 doesn't do a _great_ job at this, but it does to some extent seem to have learned some relationships in structures of text in different programming languages.
For example while toying around with it, I gave it a javascript function that referenced an undefined callback, and GPT-2 gave me back a function, with almost correct syntax and mostly random logic, but with the name and call signature of the callback.
These sorts of models do well at generating text, whether it's code or language. That's what you're talking about. They can perform the "execution" to some degree too for natural language (e.g. the squad tasks). But Q&A and text generation are very different tasks. I haven't seen anyone apply a transformer to source code based Q&A tasks.
You mean create a brain by convetion i.E. having intercouse of course.
Given raw cellular material, gtca biological code samples, design an organic chemical engine "brain" capable of self replication, repair and inference?
Its sorting through the 4.5 billion years of leftover legacy dna code that makes it tedious to understand.
In case anyone thinks the above joke is actually what the paper is doing: it is not. The work is trying to use the same techniques that neuroscientists use on neural systems on a processor. The reason is simple: we already understand processors backwards-and-forwards. The authors' hypothesis is that if the techniques we use on neural systems do not yield useful insights that match the already known-correct understanding of a processor, then we may have a problem with our current neural techniques.
It seems the paper is about using neuroscience techniques (designed for the brain which is hard to measure nondestructively) to examine microprocessors (which we already know all the physical measurements of) in order to evaluate how effective these techniques would be if we weren't so limited in our measurement of the brain. It's not as vapid as the title suggests.
http://www.neuwritewest.org/blog/can-we-reverse-engineer-the...