The axon membrane could be analogous to this ribbon cable.
Chapter two
The Corduroy Neuron

Here is how I think it probably works. The neuron is a multichannel device. Its membrane is functionally analogous to the ribbon cable shown in this photo.

There is no encoding and no decoding. There is no code, no signal processing, and no need to notice and compare the distinct arrival times of a clocking pulse and a sensory pulse.

In a multichannel neuron each spike communicates a number, an integer, which is instantly meaningful upon arrival at the brain.

The model explains why the brain is so fast. It uses numbers.

The brain will read incoming channel numbers as levels — analog increments — not as crisply inscribed Arabic numerals. But it will know what to do with them.

The model increases the channel capacity of the human nervous system from 1011 up to perhaps 1013 or 1014.

A relatively modest change with striking theoretical consequences.

From the beginning
If the impulse moves slowly, and it certainly does, and we are quick and smart – and yes we are – then each impulse must be freighted with complete information.

To arrive at a fresh model of the neuron, it is necessary to stay within two rules. 1) because it is so slow, the impulse must carry a heavy load of information and 2) whatever the trick, the secret variable, it must elude detection by all the instruments commonly used to study nerves.

Let’s start with the notion that each individual nerve impulse communicates finely graded information that is instantly readable and meaningful to the brain. How to model this?

The corduroy membrane:
Suppose a model axon has 300 discrete longitudinal transmission channels. On the sensory end of this neuron model, a voltage source can be applied to stimulate the neuron in a range between 0 and 300 mV.

In this model, each longitudinal channel corresponds to an increment of stimulus voltage. Channel #1 means 1 mV. Channel #2 means 2 mV. Channel #27 “means” 27 mV, and so on up through Channel #300.

The neuron is now stimulated at some level of intensity, say 35 mV. An all-or-none nerve impulse of the familiar type is triggered and goes chugging down the axon. It looks, to a laboratory instrument, like every other nerve impulse. But it is traveling along a specific longitudinal channel, Channel 35, and in this way it preserves the original meaning of the graded voltage stimulus (i.e., 35 mV) all the way to the end of that channel.

Then what? Probably a separate and distinct synapse for each channel. Perhaps a chemically encoded channel identity, such as a unique peptide, packaged with the neurotransmitter.

There is a semantic difficulty with the model because it depends on the concept of channels — and ion channels are so central to our understanding of the nerve impulse that there is a potential for confusion between the postulated “longitudinal channel” and the ion channels.

In fact, they are the same channels.
Na channels linked by protein to create a longitudinal signal channel the length of the axon.
Linked receptors and other linked membrane structures are commonplace in biochemistry. The basic idea is to biochemically link adjacent sodium ion channels in a long line of succession the full length of the axon. This produces one longitudinal transmission channel as long as the axon. Repeat this structure in order to form about 300 longitudinal channels. A corduroy membrane.

In the model, the ion channels are connected with protein links, represented here by white spheres, since they are abstractions. Each single sodium channel (as conventionally understood) is represented in purple by its four homologous transmembrane protein domains.

The only novel element in the model is the conjectural protein link between channels — the white sphere. The link could be in the membrane, under the membrane in the cytosol; it could be cytoskeletal — or not. There is no specification necessary in this model, beyond the notion that a link between individual sodium channels exists. Thanks to this link the individual sodium channels have been, in a manner of speaking, polymerized or concatenated to form a long chain.

Na channels are in effect polymerized to form this longitudinal channel, one of ~300.
Note that one could wrap these longitudinal channels around the long axis of the neuron, forming a helix embedded in the axon’s cell membrane.

In this helical version of the model, conduction speed would be a function of the period of the helix.

Consequences — at first glance.
The model has the virtue that when a single nerve impulse arrives on channel 27, for example, it is instantly and completely meaningful. It means 27. It need not be counted, clocked, accumulated and averaged or otherwise processed to extract this meaning.

In detail it means, “27 mV were being applied to the sensory end of the nerve at the instant when the impulse was fired down this channel.” In a motor version of the nerve model, the channel number would correspond to a precise positional instruction. Channel 27 means, “bend your elbow 27 degrees.”

The sensory channel number and millivoltage have been made identical here for the purpose of this explanation. A real world Channel #1 would “mean” a voltage level corresponding to the threshold required for the nerve to fire. The neuron could be coarse or fine in resolution, depending on how finely the channels are incremented.

Multichannel neuron axon membrane. A single concatenated sodium channel marked in orange is firing. Multichannel neuron membrane. A single channel marked in orange is firing.

What’s missing?
A channel selection mechanism. The multichannel nerve model differs from the familiar, conventional one-channel neuron in an important way. A one-channel neuron has only one firing threshold. A multichannel neuron requires a succession of incrementally higher and higher firing thresholds, each corresponding to a channel number. The thresholds could vary linearly but it seems more desireable that they would vary logarithmically.

The channel selector can be modeled as a passive device, simply assuming a staircase of thresholds, but the model becomes more fruitful and intuitive if channel selection is made active, as though by a moving commutator. A shifting commutator model readily reproduces waveforms commonly detected in the lab on real neurons.

Section of model neuron at the axon hillock. Section through the model neuron at the axon hillock, showing one model of a channel selector. This device works like a commutator within its normal range from channel 0 to channel 300. For a given stimulus, it shifts from channel to channel, firing each in rapid succession, until it arrives at a channel where the intensity of the stimulus is matched by the channel number. There it stops.

The channel selector can move in either direction in response to intensity changes. It can also halt, which puts an end to firing. But note that there is no mechanical stop at zero. Because there is no stop, the system is perfectly capable of “motoring” if it is overdriven. This motoring effect could generate and account for Adrian’s firing rate code.

This is a mechanical model of a simple and familiar electrical machine, a commutator. The model does not, however, operate by “making electrical contact” with each channel in turn. Instead it operates chemo-mechanically on each channel terminal, in some sense prising open the initial sodium channel and thus launching an action potential.

The commutator is a metaphor for some real, probably cyclical process that would operate at the molecular level. It is a biochemical machine, and so it is all about binding, conformational changes, and conformational responses to binding. These effects have ultimate consequences we can measure with electronic instruments, like an influx of ions. But the mechanism cannot be fully understood if we insist on regarding it as a purely electrical device. The underlying biochemical twists and shifts and hooks and grabs cannot be detected electrically.

Note that in a mechanical model, the commutator can be made to act as an attenuator by adding a spring that resists the pointer’s rotation. Out of range stimuli could be in effect reined in by a spring whose tension could be varied. This suggests a metaphor for adaptation. It could be that each signal transmitted by the multichannel neuron model has two components, like a logarithm. One number specifies the attenuation needed to achieve adaptation (X5, X10, X50). The second number specifies the position within a reporting range (e.g. a channel number from 0 to 300).

Biochemistry is fast and extremely mechanical, a molecular watchworks. The circulating crawler/selector above is reminiscent, in principle, of an enzyme finding and binding its active site, or a ribosome ratcheting along, or a polymerase at work on a loop. Another example at the molecular level is ATP Synthase, which actually uses free rotation, as shown in this animation depicting the work of Nobel Laureate Dr. John E. Walker, Medical Research Council, Dunn Human Nutrition Unit, Cambridge, UK.

In sum, ratcheting, site seeking and recognition, circularized strands, circular strand following and even free rotation are not novelties in nature. One can model a commutator without resorting to mechanisms that feel “unbiological.”

The commutator model can be used to interpret the known firing patterns of neurons, but I will re-emphasize that it is a conjecture, a model. No one has ever looked for such a mechanism.

If you like the corduroy neuron model, therefore, you must simply assume that some sort of threshold sensitive channel-selector and impulse launcher exists on the front end of the nerve, at or ahead of the multichannel axon hillock. This device would point to and trigger off the specific channel that is numerically appropriate to the intensity of an applied stimulus.

A test for the model: reproduce Adrian
Although we are now free to seek alternatives to Adrian’s firing rate code, which was his interpretation of his experimental results — any realistic model must be able to account for and faithfully replicate Adrian’s discovery: Spike frequency must vary as a function of stimulus intensity.

The model does not have to follow this pattern all the time, but it must have in its repertoire the behavior Adrian (and thousands since) observed.

For the multichannel model, if lots of meaningful impulses should happen to travel the axon in rapid succession, it means the stimulus is changing rapidly on the sensory end. In this version of events, the frequency does not indicate the intensity of the stimulus. Rather, it indicates the rate at which the stimulus is changing – rarely faster than in that moment when the stimulus is first applied; or when the stimulus is removed or turned off.

Click to enlarge, Back to return.Spikes recorded from a frog retinal ganglion by Hartline. Tufts appear when stimulus is turned on and then off.Notice the two “tufts” of impulses at the onset and offset of stimulus in this oscillogram. In terms of the model, the first tuft indicates that the commutator is winding up rapidly in response to the light stimulus, firing at each increment. The second tuft appears when the commutator winds back down after the light is turned off, firing at each decrement. The illustration is an interpretation of an oscillogram recorded by Hartline in 1938 from a retinal ganglion cell of a frog. It demonstrates one of three patterns he observed and named: “ON,” “ON-OFF,” and “OFF.”

The tufts produced automatically by the multichannel neuron model begin to explain Adrian’s observation, but the hypothesis does not yet account for firing rate variations observed in long pulse streams. To produce this result, it turns out the multichannel model must be overdriven, as discussed below:

What if the medium is not the message?
In this multichannel model, what does the detection of a rapidfire stream of impulses really tell us? It lets us know that the stimulus is changing, and that the last recorded impulse conveyed the most recent value – though we cannot guess that value. A pulse stream could mean 1 2 3 4 5. It could mean 9 8 7 6. It could make an undetectable turn in mid passage, and mean 1 2 3 4 5 4 3 2 1, for a net change of zero.Speaking of zero, in a multichannel system that has assigned one, a stream of impulses could mean 0, 1, 2, 3, 2, 1, 0, -1, -2,-3.The system might even break into an oscillatory response that would be completely cryptic — trilling up and down the number line in sawtooth fashion while, to the observer, simply appearing to be firing one spike after another in rapid succession.

For example, a nerve that “fires like a machine gun” could be a nerve that is for the moment overdriven and overwhelmed. This means it is poorly scaled to the size of, and incremental changes in, the received stimulus.

This neuron could run up through its full range of channel numbers, plummet back to zero, and then repeat the cycle again and again.

Click to enlarge, Back to return.
Overdriven neuron. Undetected sawtooth waveform underlies the steady spike stream.
Say 300 is the topmost channel number, and 0 is the next number in a rotation. The commutator increments past 300, seeking a higher channel number, but the next number in turn is zero. The undetected pattern that could be plotted from channel numbers is a sawtooth, but the observed pattern on an oscilloscope screen is just a continuous high frequency pulse stream.

The higher the intensity of the stimulus, the steeper the ramp of each sawtooth and the higher the observed pulse frequency of the continuous pulse stream on the scope. Conversely, the lower the intensity of the stimulus, the gentler the ramp of each sawtooth, and the lower the observed spike frequency, as indicated here:
This neuron's spikes mimic Lord Edgar Adrian's rate code but the code is meaningless here.This hypothetical looping or “motoring” effect fits the model more completely to Adrian’s results. It also suggests that adaptation means scaling, attenuation, zero-positioning. Adaptation stops the motor.

Once adapted, neuron can then operate normally within its range of channel numbers, in this example 0-300. Within this normal operating range, a single spike is sufficient to communicate, as a channel number, the intensity of the stimulus.

Note that once a multichannel neuron is overdriven and breaks into oscillation, it essentially loses its power to quickly communicate meaningful information with a single impulse. The output becomes a blur of channel numbers. It could communicate by other means — Adrian’s code for example.

However, a better engineering solution in the event of oscillation might be to simply shift the system’s attention to a nearby neuron better scaled to the magnitude of the input. If the neuron cannot quickly adapt its range to the input, it could be inhibited, i.e., shut off.

It is possible, in sum, that what Adrian’s code signals, in a long stream of spikes, is a neuron that is temporarily overdriven and indisposed while it adapts.

The overdriven neuron is not just an artifact. The cycling effect could be quite useful in the normal operation of the nervous system, since it marks a re-scaling operation in progress. And it is experimentally useful in the same way a pinned meter is useful. It lets us know that the stimulus has exceeded the neuron’s accustomed operating range.

Finally, note that the model could be overdriven in either direction. A neuron scaled to a strong stimulus may break into oscillation if the stimulus is suddenly removed.
A neuron may be overdriven by the sudden removal of stimulus. An effect called motoring ensues.Notice that the sawteeth are flipped (i.e., with descending ramps) and that the transition or overrun occurs in the 0-to-300 direction, rather than the 300-to-0 direction. The commutator increments toward lower channel numbers, seeking a value that is smaller than 1 or is negative — but it instead finds channel 300. “Motoring” ensues.

I have included the overdriven neuron hypothesis here in order to emphasize how much busy activity could be happening behind the scenes, and how utterly oblivious an oscilloscope might be to any and all of this activity.

However the system behaves and responds, in this model a nerve impulse is a medium. It carries a message — but it is not the message. The exact message, which is a channel number, cannot be readily deduced from the behavior of the medium.

Experiments (and this implicates all of our experiments) that detect and report the behavior of the nerve impulse as a medium can be perfectly repeatable yet perfectly inscrutable.

If the multichannel model were to prevail, this problem of confounding the medium with the message would raise many questions about the fundamental experiments in neurophysiology. It would also change somewhat the meaning and functions attributed to synapses and neurotransmitters.

 
The scale of the system
Take a look at Ion Channels of Excitable Membranes, which is the classic book on this subject by Bertil Hille.

One of the surprises in this splendid, fascinating book arises from Hille’s thumbnail history of the very idea of individual ion channels.

It is a much more recent idea than one might suppose. Not until the mid-1960s (well after the Hodgkin Huxley Katz voltage clamp work) did neurophysiologists finally arrive at the now commonplace image of an ion channel as an individual structure – an ion-specific porthole or passageway through the cell membrane.

Hille emphatically characterizes the individual channel as “a discrete entity,” and as “a distinct molecule.” By 1965 this concept had been in the air for a while, but it did not prevail or become the dominant picture until binding studies were conducted with tetrodotoxin and saxitoxin. Largely thanks to this work, by the late 1960s, the author recounts, the names “Na Channel” and “K Channel” began to be used consistently.

The familiar, orderly picture of individual channels embedded in the cell membrane was brought to us by the magic of long division. For example:

“Dividing specific binding by membrane area yields an average saxitoxin receptor density of 110 sites per square micrometer on the axon membranes of the vagus. We now know that the tetrodotoxin-saxitoxin receptor is a single site on the Na channel, so this experiment tells us how many Na channels there are in the membrane. Surface densities of 100 to 400 channels per square micron are typical …”

The picture you get is one of barrel like protein ports floating like buoys in the membrane, nicely regimented into rows and columns, anchored at the intersection points of an imaginary grid. It is, of course, an image made ideal by the arithmetic which originally produced it. Here is a diagram of the components of a single mammalian Na channel.

alpha & beta NaJust beneath the membrane the channels are in fact tethered to the cytoskeleton via Ankyrin, circled in red here. Our understanding of how the Na channels are positioned by these tethers is essentially two-dimensional. We know that in myelinated neurons, very dense local concentrations of the channels are maintained at the nodes of Ranvier and at the beginning of the axon. But the positioning of sodium channels in 3-space is not clear.

Hille concludes: “Now that we can record from single channels – and even purify them chemically, and sequence and modify their genes – there remains no question of their molecular individuality.”

Reading this, you get the idea there may have been a rather hot argument, years ago, about the “molecular individuality” of the Na channels. Maybe a question was raised about whether tetrodotoxin-saxitoxin receptor was an accurate index to the true number of Na channels. As in fact it turned out to be.

Whatever happened, I think there does remain after all a staggeringly important question about the molecular “individuality” of these ion channels. This is because the channels, though they are indeed distinct molecular entities — can be structurally and functionally linked. Clustered, paired, lined up in rows, bridged.

Linked cell-surface receptors are a commonplace of biochemistry and ion channel linkage is of course the anatomical basis of the multichannel neuron model we are exploring here.

To get a sense of the scale of the model, using Hille’s numbers, visualize a one micron square area of the neuron’s cell membrane, and draw 10 to 20 straight lines across it. The lines represent the passage of 10 to 20 longitudinal transmission channels — each longitudinal channel assembled as a row of 10 or 20 linked Na Channels.

This image shows us the working scale of the hypothetical corduroy membrane surface of a multichannel neuron. I doubt the dimensions are true to life – too neat.

The channels could be packed thick or thin, and they could run helically around the long axis of the axon, or longitudinally along it. They could even be closed rings around the neuron, linked with protein bridges – and the model neuron would still function as a multichannel device. But anyway, this is the place to start, at the square-micron level.

For certain configurations, e.g., the helically wound channels, one might try to detect a tiny phase shift in response to a substantial change in the magnitude of the stimulus, but it would probably wash out. The existence of multiple channels in whichever configuration will be difficult to detect at this scale, although it should help that they are probably periodic structures.

Molecular models, proteins as playthings
Is there any evidence for linked or complex ion receptors? At the end of Chapter 5, in a literature summary, Hille remarks on the then newly discovered double barreled anionic channels, and notes some Cl channel electrophysiological data that seems to make it look “as though the channel were a cluster of pores – like a sieve or an aggregate of straws. An alternative would be that the pore fluctuates through frequent rearrangements of many constituent parts.”
This “pore that fluctuates through frequent rearrangments” is inspired, an admirable idea and a realistic approach to try to follow. Molecular modeling is something like toymaking.

If you put two pulls on an ordinary zipper, you can create a pore that travels. It is easier, not harder, to come up with this kind of mechanism by assembling protein subunits. You can also make starbursts, “cootie catchers”, “Jacob’s ladders” sliding anagram toys, and many other plaything analogs using protein repetitive units, links, foldings and conformational changes.

The molecular plaything metaphor for the ion channels of the nerve cell axon membrane is exactly what’s called for. It seems clear that for the multichannel model of the neuron, the protein links would probably serve to associate the Na channels. The image of a zipper with two pulls, one closely chasing the other, applies nicely. One also could link in, with a lag effect (owing to a conformational change), the potassium channels, though this is not absolutely required by the machine in order to work.

Note that the linked ion channels are now opening and closing under positive mechanical influence – in this model they are no longer free floating, each anchored in isolation. Nor is their opening simply triggered off by a passing wave of transmembrane potential changes. Our understanding of the action potential is rooted in the concept of voltage-gated ion channels, but this is not necessarily the whole story or the only story. In a model that invokes protein linkage between sodium channels, there is a definite order of succession in their operation, a mechanical or unlatching progression. Quite like a zipper.

At the synapse
Suppose the model is correct. Then for impulses measured as voltage changes across the axon membrane, we have been confusing the medium and the message. The nerve impulse is the medium. The message is the channel number.

What does this mean at the synapses? What about the synaptic potentials? Inhibition, excitation? The model doesn’t require that we change our present interpretation of what goes on at the synapse, but it suggests there is another level of meaning, of refinement. If the neuron is a multichannel device, then the synaptic potentials are gross effects. The fine effects – whatever they are — are not being detected or measured.

Effects we have called out as summation, or inhibition, or excitation – are descriptive terms we have imputed to our own voltage measurements. They imply a certain function and significance.

These interesting names our predecessors gave to the effects they were able to measure — may be obscuring from us some more finely resolved information (biochemical or electrical) that could be, with better or different instruments, dissected away from or teased out of the gross measurements.

In other words the familiar potentials and effects and phenomena mean something to us because we can detect and measure and interpret them. But they do not necessarily mean the same thing, or anything, to the nervous system.

The invention of Zero
It would help the multichannel model to have a zero setpoint, and a full number line range from positive 300 through zero and down to negative 300. If you split the clock like this, you can communicate finely resolved degrees of stimulation (or motor instructions) that have a direction. This makes it possible to traffic in signed concepts like forward and backward, up and down, more than zero and less than zero.

It is also possible within the model to bias the zero position for a neuron that needs, for its typical stimulus, a larger positive than negative range, or visa versa.

At the synapse, an excitatory potential is measurable electrically, but the actual messages – the finely resolved increments of excitation – are not detected.

On the axon the message is the channel number. Electrically the message is not detectable, at least not yet, but detection seems possible, promising. But at the synapse, the message is likely to be embodied as a biochemical messenger. It is difficult to detect a biochemical message with an electrically sensitive instrument. And we are looking for a subtler message – probably many of them – than our standard menu of neurotransmitters can provide.

In other words, the model calls for the successful transmission, across the synapse, of a channel number. This could be done electrically by re-creating or representing the potential originally perceived at the sensory end of the nerve. It could also be accomplished with a chemical messenger packaged into the synaptic vesicle along with the neurotransmitter. Maybe a peptide unique to each channel. The channels are physically distinct from one another, so one might look for specific cell surface receptors for these peptides (or whatevers) on the receiving nerve membrane. Finally, there could be some anatomical structure, or useful proximity, that helps conserve the channels’ identity as the message traverses the synapse. The simplest way to model it would be to allocate one dedicated synapse to each of 300 channels.

The problem of inhibition
Inhibition is a thoroughly studied effect, but in the multichannel model it may have an undetected aspect. In the multichannel model, inhibition essentially turns off the nerve — stops communication altogether. As in the traditional analysis of inhibition, hyperpolarization at the input of the neuron requires a stronger stimulus to overcome. Thus, an inhibitory potential raises the theshold level of the stimulus that must be applied for the neuron to resume firing.

More complicated things can happen with the model, however, because the multichannel neuron can be inhibited in a very direct and immediate way by freezing the commutator in one place — in effect, locking the brake.

What purpose is served by this different mode or understanding of inhibition? Why turn off a nerve, or lots of nerves? What could be its operational purpose in the brain, for example?

In an individual neuron that has been overdriven, and for some reason cannot successfully adapt (rescale), it might be useful to shut down this neuron to end the oscillation. In this model, continuous firing is a pure waste of energy. A momentary stop or hiatus might also be inherent in the normal process of adaptation.

Stopping certain neurons or systems of neurons from communicating can be a control system that helps focus attention, or improve a signal to noise ratio in certain areas. If inhibition is highly coordinated, it could provide the basis of a scanning (spotlight of attention) system or a multiplexer.

The failure or overactivity of a novel, unsuspected biochemical mechanism of nerve inhibition could be significant in understanding processes such as epilepsy, cortical depression or narcolepsy. I suppose one could even project the problem into large mysteries like intelligence or sleep.

To keep this on the ground, let’s note again that what we have here is a model neuron. This model can be turned off by freezing the motion of the channel selector. This Off-switch may operate when the nerve is hyperpolarized, as in the traditional, textbook sense of inhibition. But it might also switch off the neuron, or populations of neurons, for important reasons at other times as well. The problem leads to questions of cause and effect, since the cessation of firing can produce hyperpolarization. It appears from the model that this novel form of inhibition, by halting the commutator, could be made to act instantly.

Synaptic plasticity, LTP and AMPA receptors.
What about synaptic changes? These probably reflect processes of calibration and scaling, possibly the establishment of calculational pathways.A good engineering objective for a system like this would be a neuron that fires as rarely as possible. This argues for large increments, or steps, between channels, so that trivial changes in the stimulus do not trigger frequent firing. On the other hand, tiny changes in stimulus may matter a great deal if you are attempting very fine and delicate work, such as watchmaking or eye surgery.So this is an optimization problem. Sensitivity must be optimized. Optimization does not necessarily require an increase in sensitivity — it could just as readily require a decrease in sensitivity.

Optimization would require re-scaling and calibrating the neuron so that the response is appropriate to that of the most typical input stimulus. A too sensitive or too narrowly ranged neuron would fire too often or perhaps break into the overdriven looping or “motoring” mode discussed above. A too depressed neuron could miss something important. So the range of stimulus depicted by the range of channel numbers should probably be adjustable. One could do this — adjust the nerve’s resolution — by adding or subtracting channels, or by activating or deactivating every nth channel. The number of channels accessible and active would be reflected in the field potential.

Is this learning, this stretching and compressing and zero-positioning of a neuron’s operating scale? It is.

Is it capital “L” Learning, as in Learning and Memory?

No, I don’t think so. But the problem of what happens to these neurons at the synapse is clearly interwoven with the phenomenon we routinely read out as LTP.

In fact the LTP story might make clearer sense inside this slightly different logical framework of multichannel transmission.

For example, in the multichannel neuron there must exist a staircase of firing thresholds. Dialing the sensitivity of the neuron up or down, by raising or lowering the “steps” or thresholds, is a rather similar process, in effect, to LDP and LTP, respectively.

Some very sophisticated and beautiful biochemistry now depends from a curious assumption: the quaint concept of a synapse as a bad solder joint. Or perhaps as a variable resistor. The notion is based on telltale measurements of synaptic potentials. But metaphorically, the idea is that the synapse can be “strengthened” by improving the “contact” between neurons. It can also be weakened by reducing contact. Synaptic plasticity, that is, a change in synaptic strength, is supposed to reflect learning. Indeed it has been shown that synaptic changes do accompany behavioural changes. (The synaptic changes are necessary but not sufficient).

The synapses are thus nominated, in a considerable leap of faith actually, as the sites of memory storage.

This interpretation is based on the nearly universal concept of a neuron as a one-channel transmission system. In the hippocampus, where glutamate is the dominant neurotransmitter, “strengthening and weakening” of a given synaptic connection is accomplished by trafficking within the neuron in AMPA receptors (AMPARs). These receptors pool in the dendrites and are ultimately mounted in the dendritic spines. More AMPARs in action at the synapse means “more strength”. Fewer receptors in place means “less strength.”

The theoretical value of dialing up and down synaptic strength, or conductance, is clear and important in the textbook model of the synapse. Variable “strength” imparts a much needed analog quality to the stubbornly digital, all-or-none nervous system. Variable synaptic strength has been incorporated into various familiar models of memory. Historically, it has been used by theorists to create nervous systems within the nervous system — grooved-in pathways or, roughly from the 1930s into the 1970s, “reverberating circuits.” More recently it has been used to postulate neural networks based on distributed synaptic weightings.

In a multichannel neuron, which is the hypothesis we are discussing here, the nervous system appears to be digital to a probe on the axon, but it is actually analog throughout. Analog information can be readily detected and transferred intact along the axon with a single impulse. The content of memory is a channel number, and it exists and persists from the instant of sensory detection. The model is already wonderfully analog. It need not be un-digitized, or analogized, at the synapse.

The problem at the synapse is to transmit an incoming analog value, a channel number — across the cleft. Put another way, a synapse does not simply connect neurons. It marks the output of an individually numbered transmission channel.

One could conserve a channel number conventionally, by regenerating and re-presenting an analog voltage to the next neuron.

Or one could work more directly with the array of numbered output channels at the output end of the axon, and try to construct an hypothesis in which the channel number is communicated to the next neuron mechanically and biochemically.

Either way, one could guess that there exists a correspondence, at a given dendritic spine, between the population of AMPA receptors and the specific channel number of the active channel on the transmitting axon.

In this multichannel neuron model, spines are created and additional AMPARs are recruited to extend the numerical range or refine the resolution of the communicating neurons. This adapts the pathway to the amplitude or grain of the typical stimulus that is to be tracked and described.

Bear in mind that the stimulus that induces LTP is monotonous and, very probably, grossly exceeds the normal operating range of the stimulated neuron. The increase in AMPA receptor populations to accommodate the exaggerated new stimulus is unsurprising.

In normal (physiological) circumstances, more synaptic channels could accomplish one of two things: greater range, or greater resolution within a given range. One might also use AMPARs to block out an operating range within the full operating range. Say, channels 25-75 within an available range of 1-300.

In this very different view, synaptic changes are not about the “strength” of the connections between neurons. Synaptic changes are about resolution and range. Once these are properly set, the neurons should be able to communicate the magnitude of a typical stimulus with a single spike every now and then.

Recall Edgar Adrian’s frustration with a neuron that fired 10 impulses in a tenth of a second, and then “adapted.” Also remark the “tufts” observed by Hartline at the onset and offset of his light stimulus. The multichannel neuron will fire each active channel on its way up to the channel number that reflects the magnitude of the applied stimulus. A neuron set at high resolution will fire more frequently than a neuron set at low resolution.

If one were unaware of the multiple channels, a refinement in resolution might might make it appear, owing to more easily triggered and more abundant and higher frequency firing, that a synapse had been “strengthened.” An extension of range would have the same effect. More channels equals more firing.

In a multichannel nervous system, synaptic changes in range and scale can persist, but it would not usually be useful to make permanent or absolute changes at the synapse, since the typical input (stimulus) does not remain fixed in intensity.

So yes, the synapse contains a mechanism for retaining a memory, but we are not talking about Marcel Proust’s memory. In this multichannel model, in most nerves, the synaptic change might be better characterized as a setting than as a memory.

One can imagine multichannel models in which cortical synapses figure in the storage of visual memory, but the idea of changes in synaptic strength, which is so important in a nervous system built up from single channel neurons, has no place in these hypotheses.

Does the model predict any features in the LTP system? Not really, but it makes you wonder if perhaps the momentary population of AMPA receptors in the head of a given dendritic spine might be used to identify the channel number associated with that particular synapse. The idea would be to compare receptor counts in many neighboring dendritic spines to see if an order or progression could be recognized.

 
Where can we go with this?
One might reasonably ask why, in a system with up to 100 billion neurons at its disposal — why add 300 channels of fresh capacity to each single neuron? If you need 300 channels for whatever task, why not just recruit 300 one-channel neurons?

The overarching argument is speed-of-thought. By increasing the channel capacity of the nervous system by just two or three orders of magnitude, from 1011 up to 1014 for example — we can eliminate encoding and decoding and create a dazzlingly fast brain.

But there is an additional argument for a multichannel neuron. A specialized sensory cell — a rod or cone for example — is badly served by a one-channel output line. Very badly served.

What would it mean to our understanding of sensory cells like the retinal rods or cones, or the hair cells of the organ of Corti – this fresh assumption that there exist 100 or 200 or 500 distinct channels of upstream transmission capacity in a single axon?

We notice, for example, that there are hundreds of light sensitive discs stacked into each individual rod cell of the retina. Lots of sensitive hairs on a single hair cell in the ear. Lots of discs center-tuned to a specific color wavelength stacked into a cone cell.

Yet these multiple repetitive sensory structures, begging for calibration, begging to tell us something sophisticated and finely incremented and highly detailed about space or frequency or intensity, perhaps even phase — are supposedly served by a single preposterous drainpipe of an output channel.

As presently understood, a cone cell is just a one-channel light meter. It reports only on its perception of the intensity of impinging light — and nothing more. Even if the cell had something more to tell us (that is, to tell the brain) it could not get its messages through in realtime or near realtime.

Information from a photoreceptor must ultimately be funneled into a retinal ganglion, an output neuron with the standard, simplistic all-or-none axon. All information about the retinal image that can be made available to the brain must be transmitted along the single channel, all-or-none axons of the retinal ganglion cells. There are in the retina of one eye about 125 million photoreceptors. From all these inputs, the output cable of axons from the retinal ganglion cells is just 1.2 million lines, essentially a 100-fold reduction. In this sense the optic nerve is a horrific theoretical bottleneck. For multichannel neurons, the bottleneck does not exist. The optic nerve could have a half-billion channels.

We know a cone cell of the retina produces an output potential which varies as a function of light intensity. Maybe it does much more. It might even produce and transmit, given a multichannel neuron, information about light intensity sensed at each disk within the cone cell. Such a cone could return to the brain a precis on intensity patterns (for example, standing wave interference patterns) at various planes along the cone cells’ z-axes in the fovea.

Other ways to apply this ability to detect intensity peaks and patterns along the z-axis of each photoreceptor would be in detecting chromatic aberration, or perhaps in remarking spatial phase.

You can limn suppositions like these about pressure sensors, acoustic sensors, olfactory sensors. In brief, there could be much more information emerging from sensory organs than we ever imagined. We should be more ambitions for these sensors, given a multichannel neuron.

The Good Address
Computer circuits depend absolutely on memory storage addresses. Are there addresses in the nervous system for the sources and sinks of information? It is problematical because nerves branch.

Visualize a nerve trunk served by 7 or 8 dendrites fanned across some tissue, the skin of the fingers, let’s say — each dendrite ending at a distinct point. An impulse glides up the central trunk. The brain must ask — Where did this impulse come from? Which finger, which branch, which dendrite was the original source?

Once the impulse reaches the central trunk, it would seem the source’s address is lost.In a corduroy neuron, however, the multiple channels can be used to conserve the identity of the source. In the model, a channel’s meaning is arbitrary. It can mean 25 mV, -25 mV, or +15 degrees or –15 degrees, or it can mean, “left little finger” – an address.

The brain needs a dual signal: of an address for the originating tissue, associated with a number representing an analog intensity level. It would be very computeresque. Probably evolution would have combined the address+information channels in a single neuron, by segmenting the available channels (say 300 of them) into five sets of 60 analog channels – each set of analog-increment channels corresponding to an originating dendritic branch and stimulus sourcepoint.

The address channels probably came first, and the (additional) information channels were probably added later, by duplication. Duplication is a favored process for effecting evolutionary change.

Where the idea came from.
This multichannel neuron model is one answer to the question, how could a single impulse convey information to the brain – inform a racing bat it’s time to swerve, just in time. Other models may suggest themselves, but I rather like this one.

It popped up when I was wondering how you might design an instrument to determine whether or not a nerve impulse was moving straight up the pole of the axon, or if it were, simultaneously, spinning around it, or wobbling.

The idea was to discover what degrees of freedom were still available to the “all or none” nerve impulse. We know it translates. This leaves vibration and rotation.

I have not yet solved the instrument, but when I started wondering why the impulse might want to oscillate or spin, and how it might be compelled to do so by some underlying membrane structure, it seemed a window suddenly cracked open on the problem created for us by Hermann von Helmholtz 167 years ago. We’ll see.

<<PREVIOUS CHAPTER              NEXT CHAPTER>>