Why does Gary Marcus hate computational neuroscience?

OK, this story on the BRAIN Initiative in the New Yorker is pretty weird:

To progress, we need to learn how to combine the insights of molecular biochemistry…with the study of computation and cognition… (Though some dream of eliminating psychology from the discussion altogether, no neuroscientist has ever shown that we can understand the mind without psychology and cognitive science.)

Who, exactly, has suggested eliminating psychology from the study of neuroscience? Anyone? And then there’s this misleading paragraph:

The most important goal, in my view, is buried in the middle of the list at No. 5, which seeks to link human behavior with the activity of neurons. This is more daunting than it seems: scientists have yet to even figure out how the relatively simple, three-hundred-and-two-neuron circuitry of the C. Elegans worm works, in part because there are so many possible interactions that can take place between sets of neurons. A human brain, by contrast, contains approximately eighty-six billion neurons.

As a C. elegans researcher, I have to say: it’s true there’s a lot we don’t know about worm behavior! There’s also not quite as many worm behavioralists as there are, say, human behavioralists. But there is a lot that we do know. We know full circuits for several behaviors, and with the tools that we have now that numbers going to explode over the next few years.

But then we learn that, whatever else, Gary Marcus really doesn’t like the work that computational neuroscientists have done to advance their tools and models:

Perhaps the least compelling aspect of the report is one of its justifications for why we should invest in neuroscience in the first place: “The BRAIN Initiative is likely to have practical economic benefits in the areas of artificial intelligence and ‘smart’ machines.” This seems unrealistic in the short- and perhaps even medium-term: we still know too little about the brain’s logical processes to mine them for intelligent machines. At least for now, advances in artificial intelligence tend to come from computer science (driven by its longstanding interest in practical tools for efficient information processing), and occasionally from psychology and linguistics (for their insights into the dynamics of thought and language).

Interestingly, he gives his own field, psychology and linguistics, a pass for how much more they’ve done.  So besides, obviously, the study of neural networks, let’s think about what other aspects of AI have been influenced by neuroscience. I’d count deep learning as a bit separate and clearly Google’s pretty excited about that. Algorithms for ICA, a dimensionality reduction method used in machine learning, were influenced by ideas about how the brain uses information (Tony Bell). The role of dopamine and serotonin have contributed to reinforcement learning. Those are just the first things that I can think of off the top of my head (interestingly, almost all of this sprouted out of the lab of Terry Sejnowski.) There have been strong efforts on dimensionality reduction – an important component of machine learning – from many, many labs in computational neuroscience. These all seem important to me; what, exactly, does Gary Marcus want? He doubles down on it in the last paragraph:

There are plenty of reasons to invest in basic neuroscience, even if it takes decades for the field to produce significant advances in artificial intelligence.

What’s up with that? There are even whole companies whose sole purpose is to design better algorithms based on principles from spiking networks. Based on his previous output, he seems dismissive of modern AI (such as deep learning). Artificial intelligence is no longer the symbolism we used to think it was: it’s powerful statistical techniques. We don’t live in the time of Chomskian AI anymore! It’s the era of Norvig. And the modern AI focuses on statistical principles which are highly influenced by ideas neuroscience.

Advertisements

6 thoughts on “Why does Gary Marcus hate computational neuroscience?

  1. Come now, I don’t really hate computational neuroscience, I just don’t think we are there yet. For sure, we need to develop theories of how the brain works, and computational neuroscience will be a key player in doing that. But we may need more complex theories than we currently have, to truly capture the complexity of the brain.

    There are lots of efforts to use the brain to inform AI, but I don’t think any of them can compete (yet?) with the mix of good-old-fashioned symbolic AI and Bayes that is Watson. Numenta, for example, hasn’t published many convincing results. To take one example, the Google cat detector (essentially deep learning and best in breed) has a long way to go when it comes to handling invariance, and nobody has successfully applied techniques like that to the kind of NLP that Watson can do.

    But by all means, write a post where you detail the biggest contributions that computational neuroscience has made so far to powerful AI systems; that would be very useful.

    — Gary Marcus

    • Yes, I apologize for the incendiary title, I woke up feeling a bit ornery when I wrote the post…

      I think that we are going to have a disagreement as to what counts as ‘AI’ (and also what is impressive). Obviously there will be some domains where what we know from neuroscience will have little to say to AI researchers – and language may very well be one of those. And it’s not to say that all of one system must use only neurally-inspired algorithms, nor that any particular algorithm must dominate a particular field; I think if the cat detector isn’t great at invariance, it’s totally fine if that’s added on top. But there’s a reason that people are excited by deep learning. We just don’t need algorithms to do everything all at once.

      I will have a more detailed post about where neuroscience has contributed to the area of ‘artificial intelligence and smart machines’, hopefully soon.

    • Best in breed? I thought the SUPERVISED video recognition work was the best, not that cat detector stuff in all the papers.

      I don’t myself happen to know what the best contributions of deep learning (or neural nets) are to NLP; but I do know a few recent applications:

      1. See this work on sentiment classification:

      http://nlp.stanford.edu/~socherr/EMNLP2013_RNTN.pdf

      (And, no, I’m not one of the authors of this paper.) It even is able to learn basic logic (e.g. negation and double-negation) without being explicitly taught (no hand-crafted feature detection).

      2. Here is a paper from back in 2011 which shows that neural nets can even learn to parse sentences:

      http://arxiv.org/pdf/1103.0398.pdf

      3. This is a neat demo of what one can do using vector space embeddings for words:

      https://github.com/dhammack/Word2VecExample

      The key ingredient is a package called word2vec written by some Google researchers. It’s maybe not quite correct to call it “deep learning”; a true deep learning method could probably give even better results. Curiously, word2vec can also be used to uncover analogies; e.g. vec(Paris) – vec(France) is very close to vec(Italy) – vec(Rome). It may not work so well with deep analogies; but it does appear to work with some of the simpler ones.

      4. These methods have recently been used to give improvements to machine translation:

      http://research.microsoft.com/pubs/200865/nn4smt.msrtechreport.pdf

      And there is this recent work by some researchers at Google:

      http://www.technologyreview.com/view/519581/how-google-converted-language-translation-into-a-problem-of-vector-space-mathematics/

      5. There is some work on using “neural tensor networks” (I wouldn’t exactly call these neural networks; but, ok):

      http://arxiv.org/abs/1301.3618

      It appears to be surprisingly accurate, given how simple it is.

      And those are just some of the applications of these methods to NLP. They’ve also been applied to things like robotic grasp detection (by some people at Cornell); pedestrian detection (by Yann Lecun et al); handwriting generation (by Hinton and his students, I think); super-human-level road sign recognition (by J. Schmidhuber); cancer detection (Schmidhuber again); Chinese character recognition; optical character recognition; and so on.

      Statistical methods, more broadly, have been applied to all kinds of NLP problems, to great effect.

  2. The role of dopamine and serotonin have contributed to reinforcement learning.

    Mmm…other way around, I’m afraid!

    …statistical principles which are highly influenced by ideas neuroscience

    I’d love to hear about statistical principles influenced by ideas in neuroscience. Note that statistical principles being used in analyses of neuroscientific data (which most of the examples you cite above fall under) is not the same thing.

    • What I was thinking about re: RL was really TD-learning. I went back to the Sutton&Barlow 1993 paper and he had this to say:

      “The TD model was originally developed as a neuron- like unit for use in adaptive networks (Sutton and Barto 1987; Sutton 1984; Barto, Sutton and Anderson 1983).”

      I had *thought* that thinking about serotonin and dopamine contributed to what went into TD-learning models, but you’re right, I’m no longer so sure this is true…

      In any case, I’m going to start working on a post that will hopefully have a lot of the history on this to clear things up.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s