Yeah, but what has ML ever done for neuroscience?

This question has been going round the neurotwitters over the past day or so.

Let’s limit ourselves to ideas that came from machine learning that have had an influence on neural implementation in the brain. Physics doesn’t count!

  • Reinforcement learning is always my go-to though we have to remember the initial connection from neuroscience! In Sutton and Barto 1990, they explicitly note that “The TD model was originally developed as a neuron like unit for use in adaptive networks”. There is also the obvious connection the the Rescorla-Wagner model of Pavlovian conditioning. But the work to show dopamine as prediction error is too strong to ignore.
  • ICA is another great example. Tony Bell was specifically thinking about how neurons represent the world when he developed the Infomax-based ICA algorithm (according to a story from Terry Sejnowski). This obviously is the canonical example of V1 receptive field construction
    • Conversely, I personally would not count sparse coding. Although developed as another way of thinking about V1 receptive fields, it was not – to my knowledge – an outgrowth of an idea from ML.
  • Something about Deep Learning for hierarchical sensory representations, though I am not yet clear on what the principal is that we have learned. Progressive decorrelation through hierarchical representations has long been the canonical view of sensory and systems neuroscience. Just see the preceding paragraph! But can we say something has flowed back from ML/DL? From Yemins and DiCarlo (and others), can we say that maximizing the output layer is sufficient to get similar decorrelation as the nervous system?

And yet… what else? Bayes goes back to Helmholtz, in a way, and at least precedes “machine learning” as a field. Are there examples of the brain implementing…. an HMM? t-SNE? SVMs? Discriminant analysis (okay, maybe this is another example)?

My money is on ideas from Deep Learning filtering back into neuroscience – dropout and LSTMs and so on – but I am not convinced they have made a major impact yet.

Neuroscience podcasts

I have a long drive to work each day so I listen to a lot of podcasts. I have been enjoying the new Unsupervised Thinking podcast from some computational neuroscience graduate students at Columbia. So far they have discussed: Blue Brain, Brain-Computer Interfaces, and Neuromorphic Computing. Where else would you find that?

I also found out that I got a shout-out on the Data Skeptic podcast (episode: Neuroscience from a Data Scientist’s perspective).

Update: I should also mention that I quite like the Neurotalk podcast. The grad students (?) interview neuroscientists who come to give talks at Stanford. Serious stuff. Raw Data was also recommended to me as up-my-alley but I have not yet had a chance to listen to it. YMMV.

Is the idea that neurons perform ‘computations’ in any way meaningful?

I wrote this up two months ago and then forgot to post it. Since then, two different arguments about ‘computation’ have flared up on Twitter. For instance:

I figured that meant I should finally post it to help clarify some things. I will have more comments on the general question tomorrow.

Note that I am pasting twitter conversations into wordpress and hoping that it will convert it appropriately. If you read this via an RSS reader, it might be better to see the original page.

The word ‘computation’, when used to refer to neurons, has started to bother me. It often seems to be thrown out as a meaningless buzzword, as if using the word computation makes scientific research seem more technical and more interesting. Computation is interesting and important but most of the time it is used to mean ‘neurons do stuff’.

In The Future of the Brain (review here), Gary Marcus gives a nice encapsulation of what I am talking about:

“In my view progress has been stymied in part by an oft-repeated canard — that the brain is not “a computer” — and in part by too slavish a devotion to the one really good idea that neuroscience has had thus far, which is the idea that cascades of simple low level “feature detectors” tuned to perceptual properties, like difference in luminance and orientation, percolate upward in a hierarchy, toward progressively more abstract elements such as lines, letters, and faces.”

Which pretty much sums up how I feel: either brains aren’t computers, or they are computing stuff but let’s not really think about what we mean by computation too deeply, shall we?

So I asked about all this on twitter then I went to my Thanksgiving meal, forgot about it, and ended up getting a flood of discussion that I haven’t been able to digest until now:

(I will apologize to the participants for butchering this and reordering some things slightly for clarity. I hope I did not misrepresent anyone’s opinion.)

The question

Let’s first remember that the very term ‘computation’ is almost uselessly broad.

Neurons do compute stuff, but we don’t actually think about them like we do computers

Just because it ‘computes’, does that tell us anything worthwhile?

The idea helps distinguish them from properties of other cells

Perhaps we just mean a way of thinking about the problem

There are, after all, good examples in the literature of computation

We need to remember that there are plenty of journals that cover this: Neural Computation, Biological Cybernetics and PLoS Computational Biology.

I have always had a soft spot for this paper (how do we explain what computations a neuron is performing in the standard framework used in neuroscience?).

What do we mean when we say it?

Let’s be rigorous here: what should we mean?

A billiard ball can compute. A series of billiard balls can compute even better. But does “intent” matter?

Computation=information transformation

Alright, let’s be pragmatic here.

BAM!

Michael Hendricks hands me my next clickbait post on a silver platter.

Coming to a twitter/RSS feed near you in January 2015…

 

The bigger problem with throwing the word ‘computation’ around like margaritas at happy hour is it adds weight to

The public sphere of neuroscience

I have complained in the past about the lack of a blogosphere in neuroscience. And it’s not just bad for the community – it’s bad for the scientists, too. Here is a short selection from a piece on how twitter and blogs are not just an add-on to academic research:

A lot of early career scholars, in particular, worry that exposing their research too early, in too public a manner, will either open them to ridicule, or allow someone else to ‘steal’ their ideas.  But in my experience, the most successful early career humanists have already started building a form of public dialogue in to their academic practise – building an audience for their work, in the process of doing the work itself…

Perhaps the best example of this is Ben Schmidt, and his hugely influential blog: Sapping Attention.  His blog posts contributed to his doctorate, and will form part of his first book.  In doing this, he has crafted one of the most successful academic careers of his generation – not to mention the television consultation business, and world-wide intellectual network. Or Helen Rogers, whose maintains two blogs: Conviction: Stories from a Nineteenth-Century Prison – on her own research; and also the collaborative blog, Writing Lives, created as an outlet for the work of her undergraduates…The Many Headed Monster, the collective blog authored by Brodie Waddell, Mark Hailwood,  Laura Sangha and Jonathan Willis, is rapidly emerging as one of the sites where 17th century British history is being re-written.   While Jennifer Evans is writing her next book via her blog, Early Modern Medicine.

The most impressive thing about these blogs (and the academic careers that generate them), is that there is no waste – what starts as a blog, ends as an academic output, and an output with a ready-made audience, eager to cite it…But as importantly, blogs are part of establishing a public position, and contributing to a debate. Twitter is in some ways the same – or at least, like blogging, Twitter is good for making communities, and finding collaborators; and letting other people know what you are doing.  But, it also has another purpose.

Really, go read it all, it’s great.

Social media isn’t just a place to joke around and have fun – it’s a place to get into discussions and get your ideas out there. It’s a place to have an outsized voice if you have an outsized opinion. Papers are one way to get your ideas out there – but social media is more powerful. And a Latourian reading of science is that if your ideas don’t get out there, they don’t exist.

Although not in the influential category of the examples above, let me offer myself as an example. I often write about things that are on my mind. I put my thoughts and ideas out there to try to get them into a coherent form. And people interact and discuss my ideas with me, and help me refine them (even if they don’t know it!). I even found out that someone gave a lab meeting on one of my blog posts! Even more, I’ve found that over the past year, people will come up to me at conferences and tell me that they read my blog…which is honestly really weird for me (but it’s fine!). The point is: just being willing to talk on the internet has real-world consequences for your scientific ideas.

Someone published a comment in GenomeBiology today proposing a Kardashian Index: how many social media followers you have above what you’d expect from the number of scientific citations you have. It’s true to a certain extent: you pop the world “professor” into your twitter profile and it seems like an automatic boost in followers. But they make having an outsized following out to be a bad thing! It seems to me that means that you’re doing it right.

CSHL cognition meeting (updated with a comparison to a similar meeting in 1990)

cshl cognition

For those that are unaware, CSHL is organizing a ‘supermeeting’ with an obnoxiously great list of invited speakers. Seriously, go check it out. They’ll be discussing cognition apparently, whatever that really means. Anyway, since this is such a list of luminaries, I was curious where they were being invited from. Ladies and gentlemen, the most prosperous universities for neuroscience apparently are:

1. MIT (6)

1. NYU (6)

3. UCSF (5)

4. Harvard (4)

4. CSHL (4)

6. UCSD (3)

6. Columbia University (3)

6. Princeton University (3)

9. Stanford University (2)

9. The Salk Institute (2)

9. University of Geneva (2)

But given that UCSD/Salk are essentially the same institute in terms of neuroscience – every faculty at Salk is associated with UCSD – I’d bump them up to number 3 along with UCSF 😉

Anyway, I’ll leave it to the audience to determine the amount of home cooking, though obviously there is a lesson there either way.

Update: It was suggested that I look at an old list; the last time CSHL had a symposium focused on the brain it was called The Brain (1990). I suppose they’ve gotten more specific with time, but the list itself is pretty interesting. Besides being much larger, it’s WAY more international (there are maybe 2-3 international speakers invited in the 2014 version). There were also more people invited from private industry. Not only were there tons of pharmaceutical companies and Bell Labs, but someone from General Motors came. GM! Anyway, here is the list which I’ve tried to be fair about (ie, Beth Israel+Mass Gen count as Harvard, etc). Once again, Salk + UCSD are affiliated so I’d put them up just behind UCSF at #5.

Update #the second: David Schoppik pointed me to the write-up of the 1990 meeting which is ‘fairly remarkable’.

Old list (1990):

1. Rockefeller University (22)

2. MIT (15)

2. Harvard University (15)

4. UCSF (14)

5. CalTech (11)

5. Columbia University (11)

5. NIH (11)

8. Salk Institute (9)

8. John Hopkins University (9)

10. Washington University (St. Louis) (8)

11. Stanford University (7)

12. Cornell University (6)

12. Max-Planck-Institut (6)

14. Berkeley (4)

14. Yale (4)

14. UCSD (4)

17. University of Oxford (3)

17. New York University (3)

Clay Reid & The Brain Institute

Sounds like a band name, huh? As I jet off to Cosyne, this article seemed appropriate:

As an undergraduate at Yale, he majored in physics and philosophy and in mathematics, but in the end decided he didn’t want to be a physicist. Biology was attractive, but he was worried enough about his mathematical bent to talk to one of his philosophy professors about concerns that biology would too fuzzy for him.

The professor had some advice. “You really should read Hubel and Wiesel,” he said, referring to David Hubel and Torsten Wiesel, who had just won the Nobel Prize in 1981 for their work showing how binocular vision develops in the brain…

“Torsten once said to me, ‘You know, Clay, science is not an intelligence test.’ ”Though he didn’t recall that specific comment, Dr. Wiesel said recently that it sounded like something he would have said. “I think there are a lot of smart people who never make it in science. Why is it? What is it that is required in addition?”…

He is studying only one part of one animal’s brain, but, he said, the cortex — the part of the mammalian brain where all this calculation goes on — is something of a general purpose computer. So the rules for one process could explain other processes, like hearing. And the rules for decision-making could apply to many more complicated situations in more complicated brains. Perhaps the mouse visual cortex can be a kind of Rosetta stone for the brain’s code.

It’s a fun read about the goals of Clay Reid and of the Brain Institute as a whole. I’m always dubious about using the visual system as a model for anything subcortical, and for implicitly assuming that non-cortex is less important than cortex. And what about long-time scale modulation? But for all that, they’re doing pretty cool stuff up there.

 

What is popular in top neuroscience?

As a reminder, here is the list of the top 10 most cited neuroscience papers from the last ten years:

1. EEGLAB: an open source toolbox for analysis of single-trial EEG dynamics including independent component analysis.

2. Epigenetic programming by maternal behavior.

3. Expanded GGGGCC Hexanucleotide Repeat in Noncoding Region of C9ORF72 Causes Chromosome 9p-Linked FTD and ALS.

4. Epigenetic regulation of the glucocorticoid receptor in human brain associates with childhood abuse.

5. Suppression of basal autophagy in neural cells causes neurodegenerative disease in mice.

6. Mutations in FUS, an RNA processing protein, cause familial amyotrophic lateral sclerosis type 6.

7. Microglia: active sensor and versatile effector cells in the normal and pathologic brain.

8. Separate neural systems value immediate and delayed monetary rewards.

9. The neural basis of economic decision-making in the ultimatum game.

10. TDP-43 mutations in familial and sporadic amyotrophic lateral sclerosis.

You can also go back to my previous post and see the longer (and more detailed) list of most-cited papers from the last ten years.

I made some word clouds! I took my database and removed words like “neuroscience”, “neuron”, “cells”, etc. Here are the most common words from the top 100 most cited papers of the last ten years:

Here are the next 100:

and then 200-500:

therest_neuroscienceI don’t trust the data to go down any further. The lesson here is if you want an incredibly well-cited paper, work on human cognition! If you just want a really well-cited one, you can work on mechanisms, circuitry, or stem cells.  And unless you are working on the hippocampus, definitely don’t refer to what region of the brain you are studying unless you just want a pretty well-cited paper.

So much for making explicit that you work on rats and mice; people only want to know if you are working on humans or AIs!

Cosyne: Decision-making

I spent a week recently in Salt Lake City at the Cosyne (COmputational and SYstems NEuroscience); people had told me that it was their favorite conference, and now I understand why.  Other attendees have put up their reactions, so I figure it’s about time I got off the couch and did the same.

Probably the biggest effect this meeting had on me is that I started using twitter a bit more seriously – follow me at @neuroecology – and participated in my first “tweet-up” (is that really a thing?).  There are lots of great neuroscientists tweeting though there should be more!

For simplicity of organization, there will be three posts on Cosyne: one on a few decision-making talks, one on The Big Talks, and one on neural correlates of foraging.

Carlos Brody

On decision-making, the first (and longest) talk was by Carlos Brody.  His talk was focused on the question of how we make decisions in noisy environments.  In this case, rats had to sit around listening to two speakers emit clicks at random (poisson) intervals and decide which speaker, left or right, was clicking more.  We typically think of the way animals make these types of decisions is as a ‘noisy integrator‘: each point of information – each click – is added up with some noise thrown in there because we’re imperfect, forgetful, and the environment (and our minds!) are full of noise.  The decision is then made when enough information has been accumulated that the animal can be confident in going one direction or another.

One small problem with this is that there are a lot of models that are consistent with the behavioral data.  How noisy is the internal mind?  Is it noisy at all?  How forgetful are we?  That sort of thing.  The Brody lab fit the data to many models and found that the one that most accurately describes the observed behavior is a slow accumulator that is leaky (ie a bit forgetful) but where the only noise is from the sensory input!  Actually, I have in my notes that is ‘lossless’ but also that it is ‘leaky’ so I’m not sure which of the two is accurate, but the important detail is that once the information is in the brain it gets computed on perfectly and our integration is noiseless; all the noise in the system comes from the sensory world.

They also recorded from two areas in the rat brain, the posterior parietal cortex (PPC) and the frontal orienting fields (FOF).  The PPC is an area analogous to LIP where neural activity looks like it is integrating information; you can even look at the neural activity in response to every click from a speaker and see the information (activity) go up and down!  The rational expectation is that you’d need this information to make a decision, right?  Well, when he goes and inactivates the region there is no effect on the behavior.  The other region he records from is the FOF which is responsible orienting the head (say, in the direction of the right decision).  The neural activity here looks like a binary signal of ‘turn left’ or ‘turn right’.  Inactivating this area just prior to the decision interferes with the ability to make a proper decision so the information is certainly being used here, though only as an output.  Where the information is being integrated and sent from, though, is not clear; it’s apparently not the PPC (and then maybe not LIP)!

Kenway Louie

The second good talk was from a member of Paul Glimcher’s lab, Kenway Louie.  He was interested in the question of why we make worse decisions when given more choices.  Although he wanted to talk about value, he used a visual task as a proxy and explainer.  Let’s say you have two noisy options where you weren’t certain which option was better; if the options are noisy but very distinct, it is easy to decide which one you want. However, if they are noisy and closer together in value it becomes harder and harder to distinguish them both behaviorally and as a matter of signal detection.

But now let’s add in a third object.  It also has some noisy value, but you only have to make a decision between the first two.  Should be easy right?  Let’s add in some neuroscience: in the brain, one common way to represent the world is ‘divisive normalization’.  Basically, the firing of a neuron is normalized by the activity of the other neurons in the region.  So now that we’ve added in the third option, the firing of the neurons representing the value of the other two objects goes down.  My notes were unfortunately…not great… so this is where I get a bit confused, because what I remember thinking doesn’t make total sense on reflection.  But anyway: this normalization interferes with the probability distributions of the two options making it more difficult to make the right choice, although it is nonlinear and the human behavior matches nicely (I think).  The paper is in press so hopefully I can report on it soon…

Paul Schrater

Paul Schrater gave a talk that was a mix of decision-making and foraging.  His first and main point was that many of the things that we refer to having value are in fact secondary sources for value; money only has value inasmuch as it can buy things that are first-order needs such as food or movies or such.  However, the same amount of money cannot always buy the same amount of goods so value is really an inference problem, and he claims it can of course be explained through Bayesian inference.

His next important point is that we really need to think about decision-making as a process.  We are in a location, in which we must perform actions which have values and must fulfill some need states which, of course, influence our choice of location.  Thinking about the decision-process as this loop makes us realize we need to have an answer to the stopping problem or, how long should I stay in a location before I leave to another location?  The answer in economics tends to come from the answer to the Secretary Problem (how many secretaries should I interview before I just hire one?) and the answer in ecology comes from Optimal Foraging; in fact, both of these rely on measuring the expected mean value of the next option and both of these are wrong.  We can instead think of the whole question as a question of learning and get an answer by reinforcement learning.  Then when we stop relies not just on the mean expected reward but also the variance and other higher-order statistics.  And how do humans do when tested in these situations?  They rely on not just mean but also variance!  And they fit quite closely to the reinforcement learning approach.

He also talked about the distractor problem that Kenway Louie discussed, but my notes here don’t make much sense and I’m afraid I don’t remember what his answer was…

Is neuroscience useful? (Updated)

I recently got a quadcopter and in pockets of my spare time I’ve been attempting to make it an autonomous drone. Yet reading this article on unmanned drones has me returning to some thoughts I’ve had while working on the project.  Basically: is neuroscience useful?  Much of the utility from drones comes from their autonomy and adaptability.  In my naive fantasies, I think that the work we do to understand the nervous system should inspire drone makers, hiring neuroscientists left and right to implant the lessons we’ve learned from the nervous system into these machines.

And yet – and yet I’m not aware of anyone doing this.  There are whispers and rumors emanating from the Brain Corporation that this is their mission but I have yet to see anything concrete come out of that (to be fair, they’re a relatively new company).  But even more we should be asking ourselves: are we going to be leap-frogged by those who are working in computer sciences – artificial intelligence, machine learning, vision processing?

That the drones are living in a newly created ecosystem, interacting and invading new niches, is undeniable.  Presumably an enterprising young scientist in ecology, neuroscience, (economic) decision-making should be perfectly suited to at least consulting on these projects.  I guess the question is: does that actually happen?  Outside of ‘explaining the brain’ for ‘medicine’, do we do anything that’s actually useful?  Or is that up to the engineers?

Update: Well here’s a good example of using animal behavior/reflexes to improve robotics.

RIP Huxley

I should have posted this earlier, but Nobel prize winner Andrew Huxley died on May 30, 2012.  Not only did Huxley (along with Alan Hodgkin) perform the original voltage clamp recordings on the squid giant axon neurons, but he also came up with a beautiful set of equations to describe the neural action potential.  These equations predicted the existence of sodium and potassium ion channels well before they were known to be the cause of neuronal spiking.  The Hodgkin-Huxley equations are the E=mc2 of neuroscience.

If you don’t know who Huxley is, read his obituary and maybe this article on the importance of the Hodgkin-Huxley work.  Then make sure to find out about the other two superheroes of neuroscience, Hubel and Wiesel.

And shame on the New York Times for not even mentioning Andrew Huxley’s death; it makes one a little sad to think that one of the greatest neuroscientists of all time, someone who practically created the modern field, would go unnoticed into death.