Max’s recent post brings up the issue of Ray Kurzweil, a polarizing figure if there ever was one. First, I try to take his musings with a truckload of salt, but his proclamations about the progress of neuroscience seem to go so far beyond the pale that its shade has surpassed the visible light spectrum. He seems to completely trivialize the daunting task set before neuroscience: create a biologically and mathematically precise functional characterization of the human brain. In other words, how does it do what it does with what it has? It seems like Ray is fixated, naturally, on the exponential growth of information technology and its implications for the field. I’ll summarize his thesis, based on the 7 minute video linked by Max, thusly: computers will become insanely powerful, along with ways to measure the brain, and so we’ll have a brain simulation by 2020, or 2029, or sometime in the next fifty years.

Though it pains me to do so, I’ll grant the assumption that information technology, along some non-trivial dimension, will grow exponentially for the time being. Perhaps storage will be measured in googlebytes and some new exotic processor will compute in googleflops by the time I hit my golden years. We love to throw around anecdotes like, “A computer used to take up a whole room, now look at how far we’ve come!” I agree, we have come far in terms of computing firepower, and I have no doubt that we will go much, much further. Maybe I’ll connect to a nanocomputer in my brain and automatically integrate a set of differential equations I just dreamt up in less than a second, then see the results imprinted on my retina. Whatever happens, computational neuroscience will surely welcome the ability to simulate things much faster and at a grander scale. In short, the field may have enough processing power and storage to simulate a hundred billion neurons (a dangerously rough estimate of the number of cells in the human brain, see here) all connected to each other in the time it takes me to finish this sentence, though, I’ll grant that I can be a sluggish typist.

What do you do with this simulation? You may start by tweaking certain parameters in a complex multi-compartment model, like time constants, H currents, and the like. You could perhaps explore all possible parameters to see how the billions and billions of virtual cells dance about. You may find some data to fit and fit it. Then what? This simulation grants you some useful information, but if you think you’ve solved neuroscience then I have some land to sell you! The problem, completely sidestepped by Kurzweil, is that no experimental technique could hope to observe that many cells interacting at the proper resolution and at the proper time scale in vivo. Most neuroscientists know this, and I am sure Kurzweil knows this as well. His response to this line of attack seems to be: but experimental techniques are improving at the same rate as information technology. Nope, not even close. There are basically two ways of observing human brain activity in vivo: (1) fMRI (or PET), and (2) electrophysiology.  Neither of them, as far as I can tell, have improved exponentially, nor do they show signs of doing so in the near future.

You can’t simply state by fiat that a technique like fMRI will eventually resolve single cells at the millisecond timescale, since it will improve exponentially. As it stands, fMRI can indirectly resolve the hemodynamics of about a millimeter of cortical tissue averaged over a second, at best. Under a millimeter of cortex you find something around 100,000 neurons in humans, and spikes occur on a millisecond time scale. So, the technique is many orders of magnitude away from the desired resolution, and hasn’t shown signs of covering such a large gap any time soon, i.e. within the next fifty years. In fact, it’s hard to find any independent measure of the resolution or a precise point spread function of fMRI. If fMRI has improved exponentially, no one has the numbers to back this up.  Moreover, there is a necessary limit on how well hemodynamics reflect spiking activity.  In other words, the spatial resolution of fMRI has a theoretical ceiling, and it may be well under neuroscience’s desired height, i.e. the spiking activity of single cells.

Human electrophysiology in vivo is even dicier. Electroencephalography uses electrodes on the scalp to measure the activity of (supposedly) assemblies of neurons whose activity can leak through the skull and skin. Though the technique has millisecond precision, just what the signal is a signal of is anyone’s guess. It could be assemblies of cells that group to communicate important information through synchronous firing, or it could be a good deal of noise. Moreover, the spatial resolution of EEG (or MEG for that matter) hasn’t been precisely characterized, but it might resolve many centimeters of tissue, at best. You can improve the spatial resolution by placing the electrodes on the pia or cortex itself, in which case you need to work with special populations that require neurosurgery, like patients with medically intractable epilepsy. And maybe you can resolve the spiking activity of a large patch of cortex with electrocorticography (ECoG), but then you’re at the mercy of complex spike sorting techniques to figure out what the heck you’re looking at in the first place. The best outcome of this technique is to pick up the spiking activity of forty or fifty cells, maybe even a hundred if you’re lucky. This is a very powerful technique, perhaps the best we have for in vivo human studies, but we’re way off observing all 100 billion at the proper scale.  Further, like fMRI, we don’t have a precise measure of what its spatial resolution even is, so we have nothing to show that it has improved exponentially like the dear CPU.  In fact, since the technique is quite new, it’s hard to say if we have seen any improvement.   In any recording that happens extracellularly, you have no guarantee on how many cells you are recording from (admittedly, some scenarios make more sense than others), so we have no ground truth to compare it with.

Still, most physiology is done with non-human primates, and dual photon recording appears to be the most promising variant, as it can resolve the spiking activity of all cells in a 1 mm^3 patch of cortex. Unfortunately, the technique is extremely invasive (i.e. not tenable to do in humans), and is technically nigh-impossible, which explains why it’s only been done a handful of times and only in V1 (an area near the tip of the occipital lobe, the occipital pole). To put it in perspective, the estimated surface area of the human cerebral cortex is 2500 cm^2, with a depth of 1.5-4.5 mm.  Dual photon recording is certainly an improvement over traditional methods, but does it indicate that physiology, even primate physiology is improving exponentially, like information technology?  The answer clearly seems to be no.  It remains an exotic methodology, that, while powerful, has only been applied under a very narrow set of conditions.  Like most experimental methods in neuroscience, we’ve seen incremental growth in fits and starts.  Far from the smooth exponential curve envisioned by Kurzweil.

Additionally, areas like IT (the supposed locus of visual object recognition), my personal fave, have only been characterized at a coarse scale with single electrodes or fMRI.  That means we have data from single cells or small (i.e. 5-10) assemblies at the millisecond time scale recorded over a few hours, or from 100,000 cells averaged over a second. So, we have small time windows (i.e. a few hours each) to look at two vastly different scales.  This area provides a good case study on how difficult it is to gauge improvement in experimental neuroscience.  In certain respects, fMRI improves on the spatial scale of single cell recording. but it sacrifices temporal and spatial precision at the altar of the global, hemodynamic scale.  Should we consider this an improvement at all, let alone an exponential improvement over single cell recording?  If fMRI itself hasn’t improved exponentially, and it doesn’t seem to be an exponential improvement over classical physiology, then where’s the Kurzweilian growth?

So, experimental neuroscience is hard, really hard. The field has surely progressed over the past hundred or so years since the seminal work of Golgi and Ramon y Cajal, but to think that the field follows a nice mono-exponential function like the growth of processing power is misguided, to say the least. And to say we can simulate the whole thing in twenty years is just as misguided. Neuroscience isn’t hard because computers are too slow, it’s hard because doing the experiments and coming up with the right mathematical formalisms to describe the data is hard.  Even if information technology, in some important way, improves exponentially over the next fifty years, it cannot make up for the incremental, decidedly slower-than-exponential march of experimental neuroscience (has any non-trivial experimental science shown some sort of protracted exponential growth?).   In my experience, Kurzweil isn’t the only one to miss this point. Fast computers are wonderful, but to think that they’re going to make up for volumes of data that still need to be collected and the equations that need to be written is just plain wrong. This trivializes the hard work of experimental neuroscientists and computationalists who need more than a new CPU to crack the cosmic nut: the human brain.

Image from Flickr user Jonas B.