Tuesday, March 5, 2013

Partha Mitra Makes No Sense (to me)

The Brain Activity Mapping project (BAM) has a lot to be critical of.  There's the dubious/wise choice of selling it by promising Alzheimer's cures.  There's the funding angst - will the Administration choose to reduce NIH budget to accommodate this?  (Something, by the way, that no one, including BAM planners, wants to see).  There's the particular scientific techniques they want to explore, including the far-out "molecular ticker-tape" strategies.

But in terms of overall scientific goals, BAM should be unimpeachable.

The aspirations of BAM folks are to develop new technologies that will bring them closer to the goal of recording as many neurons simultaneously as possible.  This, as I've written before, is essentially what a large chunk of neuroscience is about.  We want to know how the brain works, so we record from brain cells!

Enter Partha Mitra, a Cold Spring Harbor PI who does, essentially, mouse connectomics (anatomical brain mapping) as well as computation/theory.  He's written a new article appearing in Scientific American that mind-bogglingly challenges the very notion that we should even TRY to record from as many neurons as possible:

The fundamental problem with the goal of measuring every spike of every neuron is one of conceptual incoherence: the proposal does not stand up to theoretical scrutiny.

So his argument is going to be that there's no point to recording every spike from every neuron - it's a fruitless task that can only end in "incoherence."

Let's go point by point through some of his contentions, and see how they hold up:

[...] the authors imply that this correlated or collective behavior cannot be deduced from other levels of observation (including the circuitry), hence the imperative need for the “measure every spike” project.

This is grossly misleading.  In fact, the authors "imply" exactly the opposite in the article he's referencing: 

Correlating this firing activity with both the connectivity of the circuit and its functional or behavioral output could enable the understanding of neuronal codes and their regulation of behavior and mental states. 

Note "connectivity of the circuit".  The whole point of BAM is to develop new tools and record from as many neurons as possible.  Then, of course, take all that information into the BAM lair and look at it in isolation from everything else!  Oh wait, that's stupid.  Of course the anatomy of the circuit is important.  That's why they want to start with organisms like C. elegans which has a defined connectome and progress to preparations like the mouse retina, which is already the subject of ongoing mapping efforts.


First, brains do not exist in isolation. Spikes are driven by two sources: the intrinsic dynamics of the neuronal network, and external stimuli. Even if one recorded all spikes from all neurons (and for the entire life span of the organism), to make any sense of the data one would have to simultaneously record all external stimuli, and all aspects of behavior. It gets worse: there will be individual variation among animals, and each animal will have a different environmental history. The “comprehensive” measurement exercise would extend ad infinitum. 

This is all true and all irrelevant.  He's arguing against the strawman, essentially saying "You can't measure everything all at once!  There's too much!"  But does that imply that there's no value in trying to measure as many cells as possible?  In developing the technologies that would make that possible?  Keep in mind, one can use these technologies for whatever behaviorally-limited purpose one wants, just as we already do.

Of course he does admit this:

One could moderate the number of neurons being recorded from, control the environmental variables, and so on — and then one has returned to the realm of what neuroscientists are doing in any case and we have a specialized technology development project, not a moon shot or a genome project.

And yes, this IS what we have.  It's not moon shot or human genome - it is what it is, and that's tech dev -> massive simultaneous single-unit recordings.  So perhaps a lot of what he's writing has to do with how grandiose BAM has been made out to be so far.  Yet he still has an objection to even the idea of doing massive recordings - why?

Still, to understand whether one should focus all energies on greatly increasing the number of neurons being recorded from, we need to answer the theoretical question of what we gain by recording every neuron. If we cannot successfully argue that comprehensive neuronal recordings solve all our problems, then partial observations certainly won’t.

True but false.  We cannot argue that massive recordings solve "all our problems".  But they can help solve some of them.  It's equivalent to the "Well the connectome can't do X, so there's no point in doing it at all" argument that is clearly illogical on its face (and he should be aware of that, given his research!).  Spikes may be all or nothing, but the value of an experimental approach isn't.

One is not really interested in the particulars of a given animal’s history of all spikes in the brain: one is interested in characterizing the potential dynamics of neurons, under all possible circumstances .

First of all, "one" can be interested in all sorts of things.  But to his general point - again true but false.  We can be interested in characterizing the potential dynamics under some circumstances and still have a valuable contribution to make.

Here is the rub: what sets the laws of the neural network? Well, it is precisely the circuit connections and the physiology of single neurons that the authors have dismissed. The paper would focus all resources into multi-neuron recordings, without any plan to complete the outstanding task of mapping out the anatomical circuitry, itself a huge project, which we have only begun to seriously address and which provides a much closer analog to the Genome project. The physiological properties of neurons depend on carefully studying individual cells or pairs of cells, also not something that is on the agenda. Once the circuit and cellular physiology is known, we can in principle derive the pattern of every spike from every neuron, under every environmental stimulus. Network structure and cellular physiology determine the dynamical laws governing the neurons, and therefore drive the spiking activity .

Yes, there is the rub.  Focus on the bolded sentence:

(1A) You cannot know cellular physiology without recording from cells.
(1B) There are many cells in the circuit, each with a unique physiology, even if some belong to classes or types of similar cells.
(1C) Therefore recording from as many cells as possible gives you the most cellular physiology.
(1D) Therefore... BAM, which seeks to develop new tech to record from many neurons simultaneously, is crap?

(2A) The circuit is knowable at any one given moment, say from a connectome.
(2B) Circuits change over time.
(2C) Therefore it is impossible to know the pattern of all spikes "under every environmental stimulus" except for the circuit in the state it was measured to be in.
(2D) Therefore that sentence should be objectionable to Partha Mitra himself on the basis of the same criteria he laid out earlier: "The “comprehensive” measurement exercise would extend ad infinitum."

Additionally, when speaking of issues with Markram's brain-from-the-bottom-up project:
The way to resolve this is not to measure every spike from every neuron, but to map circuit connectivity and measure cellular physiology.


I think what's motivating Partha Mitra's response is NOT what BAM is, but rather how it's being presented (and maybe a little bit of saltiness that his mouse connectome efforts aren't included in it...).  The record-every-neuron thing IS unfeasible, at least in humans and probably in mice too.  But to argue that it's also wrong to want to record from every neuron?

I guess I just don't get it.  He says, for instance, that recording from a single neuron, or pairs of neurons is okay.  But "every" neuron is not theoretically sound.  So, presumably, somewhere between 2 and "every", which we'll call 90 billion, there's a number at which it's no longer theoretically justified to look at individual neuronal spiking?  What?

In fact, he even writes:
Over the last two decades many labs have gathered spiking data simultaneously from dozens to hundreds of neurons. This has not yet led to any tremendous new insight: in fact, much of the dynamics can be captured by the study of correlations between pairs of neurons.
So, yeah, all you guys and girls doing calcium imaging and multi-unit recordings and multi-electrode array recordings... guess what?  You haven't produced any "tremendous new insight".  Only non-tremendous ones, I guess, but everyone knows those don't count.  There's only the behavioral level, and the single-cell level.  What's the point trying to bridge the gap?

Finally, I've ignored most of what he wrote comparing neuroscience and physics on the grounds that all analogies suck and that I'm not really qualified to argue whether we can expect truly emergent properties from these kinds of recordings or not.  Maybe that's his objection - he doesn't think there are emergent properties, so BAM is bad.  But that's fallacious because then he's just arguing against the motivations for what's being proposed, not what's actually being proposed.


There are issues with BAM, related to cost, possible opportunity cost, and, especially, how its communication tends to rely on the grandiose.  All of that is up for debate.  But the merits of the scientific goals?  There's no theoretical point to recording from a massive number of neurons?  Better to just wait and build a simulation from the bottom up?


No comments:

Post a Comment