Spikes: Exploring the Neural Code (Computational Neuroscience) [Paperback]

Neural coding has traditionally been assumed to be one of rate coding, ie, the stronger the stimulus, then the more action potentials per second that a sensory neuron transmits, and so on throughout the nervous system. However, this book begins by pointing out that in various sensory systems there appears to be sparse temporal neural coding, ie, the timing of action potentials transmits information, and in fact does so quite efficiently. A mathematical basis is built up throughout the reference in order to support these claims. However, the general reader who has prior reading of other neurobiological references listed above and below, will nonetheless find the descriptive portions of this reference informative and reasonable to read. If a neuron can fire 100 spikes (ie, action potentials) per second, then it would appear that many biological phenomena are coded by no more than one or two spikes. For example, bat echolocation occurs on a time scale of 5-20 milliseconds (enough time for coding by a maximum of one or two spikes). For example, in the fly, movements across its visual field can cause it to generate a flight torque in less than 30 milliseconds (ie, enough time for only a few spikes). For example, in the rat hippocampus signaling about position is performed on the order of one or two spikes per neuron. The fact that single spikes are carrying information in these examples indicates that at least in some parts of the nervous system, a temporal neural coding exists. As well, the issue of neuron reliability is considered in detail. Traditionally, it has been considered that individual neurons are unreliable (for example, repeated presentations of the same sensory stimulus does not cause a sensory neuron to generate the same spike train each time), and that it is only in the context of the large network of neurons of the nervous system that perception is reliable (for example, an animal running through the woods at a high speed does not collide with trees). However, it is not so clear how the different spike trains generated each time by the sensory neuron in response to the same stimulus should really be quantified, and there is much evidence showing individual neurons to be quite reliable. For example, in human vision in very dim light individual photosensitive sensory neurons are detecting single photons. The fact that the many neural circuits after the photosensitive sensory neuron add little noise to the sensory neuron output, indicates that the neural computation involved must be very reliable. The fact that hyperacuity (ability to detect sensory stimuli beyond, albeit generally just somewhat beyond before it is truly impossible to do so, the threshold of physical reliability) exists also indicates the existence of a very reliable neural computation. For example, echolocating bats resolving jitter in the echoes on an order of 10 nanoseconds, or weakly electric fish resolving signal shifts on the order of 100s of nanoseconds, or human observers with a theoretical visual acuity threshold of 0.01 degree able to discriminate 0.002 degrees. Most of this reference analyzes single trains of spikes (ie, the action potentials being generated by a single neuron), and shows clearly that very few spikes can represent very precise computations. The last chapter of this book considers briefly more recent research on spike trains of multiple neurons.

A decade ago, computational neuroscientists and some neurophysiologists were twittering with excitement about information theory. Finally, a tool that could decode the "noise" observed when we record neuronal spike signals!

These days...information theory has become part of the standard toolkit in a few types of experiments. But we're not much closer to understanding the neural code(s) than when this book was written. Nevertheless, Bialek's group of mostly physicists turned neuroscientists continue to develop information theoretic tools. Perhaps they'll come up with one that's not just another hammer.

The authors of Spikes may still turn out to have been ahead of their time (just like Barlow, MacKay and McCulloch, who originally applied information theory to neurons). Or their research program may turn out to have been a detour, a misguided attempt to find a particular physical universal in evolutionarily contingent biological systems.

If you're interested in theoretical neuroscience, I would definitely recommend Dayan and Abbott's textbook. van Hemmen and Sejnowski's "23 Problems in Systems Neuroscience" also has good bits. If you really want to read about information theory, David MacKay's new book is available on the web.

This book asks: How does a nerve convey information about the world toward the brain? It is a crucially important question - one of the most important questions in human history, in fact -- because before one can make realistic theories about how a brain works, one must know what sorts of signals it receives and acts upon.

We were all told, in basic biology, that this question was answered decisively in the 1920s: The nerve encodes and transmits information about the world in the form of frequency modulated pulse trains. The more intense the stimulus, the higher the pulse frequency, and the closer together the pulses in the train. In this system, a single impulse, or "spike", is trivial, in the sense that it is blank. It cannot convey any information alone. It takes at least two pulses to encode sensory meaning. The information that is read by the brain (meaning, say, a level of light, or the intensity of a musical tone) is encoded as the interval between pulses. And so as students we ate this FM story. And answered the inevitable, standardized questions about it on exams.

Now we learn that this familiar, ingrained bedrock idea is not actually true. Somehow, a single spike is - after all -- capable of conveying information to the brain. This news was not revealed in some single egregious experiment but, rather, by a substantial body of experimental results that have filtered into the literature recently. This book gathers and pivots around this unexpected (and probably very unpopular) body of research work, and I suggest that you initially skip all the introductory material and go straight to pages 54-60, where the experimental literature is summarized.

A nice example comes from studying the decision making time of bats. The animal uses echolocation to navigate in flight. An experimental question is this: How many nerve impulses can the creature's brain have decoded before it suddenly decides to swerve? The answer is on the order of one spike. One. Uno.

At this point in the book, the answer is already transparent. The secret of the neural encoding is that there is no code. A single spike conveys information. The information is explicit. No computation is required to extract it.

Ah, but not so fast. On page 4, the authors reiterate the all-or-none law, declaring that: "... incoming stimuli either produce action potentials, which propagate long distances along the cell's axon, or they do not. There are no intermediate signaling mechanisms. This means that a single neuron can provide information to the brain only through the arrival times of the spikes."

Evidently they still want to keep this absolute intact, and so they go on to recreate, in lieu of the familiar FM neural code, another more sophisticated code. This book is their proposal for a new code.

But it seems to me that having driven such wonderfully high piton (their assertion that the FM code isn't one) the authors proceed to rappel down the mountain very fast. Retreating, perhaps, into their alternative code theory.

Instead of following them to lower, safer ground, you might pause to consider this: There might exist, after all, "intermediate signaling mechanisms." The pulse cannot be amplitude modulated (this really is an absolute). But it can surely do many other clever things that would elude detection by the instruments used to study nerve impulses. (Voltage clamps, patch clamps, probes). Like what? It could spin. It could and probably does travel up the axon membrane in one of many discrete longitudinal channels, formed by protein links between adjacent ion channels. In such a nerve the information, or sensory increment level, is inherent in the channel number.

Neurobiology, as an industry, is somewhat at risk to ideas of the type that are let loose in this remarkable book. If one were to follow up on them, one might arrive at a theory of the brain that actually made sense. Well understood structures like the synapse would have to be explained in new ways, etc. There might be uproar.

Also take a look at Findings and Current Opinion in Cognitive Neuroscience, by Squire and Kosslyn. Chapter 25 reviews some the ideas presented in Spikes, and competing explanations offered by other authors in an effort to elucidate the so called "sparse code." One spike. Very sparse indeed. By all means get a copy of Spikes. It would be a shame to miss out on the scientific revolution it so strongly augers.