|A book by|
William H. Calvin
UNIVERSITY OF WASHINGTON
SEATTLE, WASHINGTON 98195-1800 USA
The Throwing Madonna|
Essays on the Brain
Copyright 1983, 1991 by William H. Calvin.
You may download this for personal reading but may not redistribute or archive without permission (exception: teachers should feel free to print out a chapter and photocopy it for students).
Scanned, OCR'ed, and webbed -- but NOT proofread (14 Jan 97)
Computing Without Nerve Impulses
Small Is Beautiful
Title of a book by E. F. SCHUMACHER,
subtitled Economics as if People Mattered.
(Neurophysiologists still think large is beautiful,
but they are finding that small is different.)
With the fantasy world populated by personable robots named "R2D2" and the like, it becomes
easier to think of our brain cells as fancy machines (there is even one neuron, in a seagoing slug,
that was named "R2" back in pre-Star Wars days). Building a robot that can perform
sophisticated computations about its environment via pattern recognition is a problem more
easily appreciated than its converse: figuring out how an evolution-designed computer works. At
the heart of this problem is understanding the computational abilities of the individual cells that
comprise the brain.
Our usual first line of defense is to make analogies with foods: the special molecules called neurotransmitters, so popular with neurochemists and nonscientists alike, are indeed "tasted" (if not consumed) by the neuron. Their variety seems endless (there are now many dozens of types; once there were only a few), and they are often close relatives of drugs and dietary substances. But to understand computation by a neuron, the somewhat more foreign subject of electricity is unavoidable.
However important chemicals may be for energy and for communicating between neurons, it may be said that computation within a neuron runs on electricity. To set off an impulse, the inputs impinging upon the neuron shift the membrane voltage about 10 millivolts until the threshold is reached. Like pulling the trigger on a gun, nothing happens until a threshold is crossed. Then something happens. In the case of the neuron, the something is the impulse (en famille, we call it the "spike"). Because various inputs must combine their actions to produce enough voltage (via the actions of chemical neurotransmitters) to cross the threshold, the take-home message perceived by most nonneurophysiologists (remember the "physicist's fallacy"?) is that the neuron functions in the manner of an AND gate or coincidence detector. This is unfortunate, as only a few specialized neuron types may actually operate in that digital manner.
Although some neurons are functioning entirely without producing impulses (remember those 99 percent of the neurons in the eyes reading this sentence), impulses are always used for long-distance communication, such as between eye and brain. (The distances within the hair-thickness of the retina do not often seem to require the amplification aspect of the impulse.) Voltage is still the common currency of the "nonspiking" neurons; it just regulates the release of neurotransmitter directly, rather than using the spike as a middleman.
When the strength of an input doubles, the voltage increases. This releases neurotransmitter at a faster rate. Inhibitory inputs decrease the voltage, thereby reducing the release rate. Thus the neuron can release neurotransmitter at a rate determined by the sum of many positive and negative inputs, just as a savings account pays interest depending upon the balance of deposits and withdrawals. This is analog computation, not digital. So, is the difference between spiking and nonspiking neurons the difference between digital and analog computers? No, because most spiking neurons are also analog in their computation, just using spikes for long-distance transmission of the result. The spectrum of neural computation and transmission processes can perhaps best be appreciated by first examining the so-called nonspiking neurons.
To function without spikes, a neuron must be small--and thus not one of the neurobiologist'sfavorite neurons whose large size makes them an easier target for inserting probes to measure internal voltages. Elongation over several millimeters usually means that the neuron uses impulses. But the brains of humans, as well as of our favored research animals, are all filled with cell types that fulfill the small-size criterion; most have yet to be studied. One attempts to extract principles of operation from the study of those neurons that can be reliably studied and then extrapolate them to the many situations, such as human sensory pathways through brain stem and thalamus, where they cannot be studied directly. The mud puppy, a primitive vertebrate, has a retina with exceptionally large cells; since John Dowling and Frank Werblin showed its extensive use of nonspiking neurons in 1968, many other vertebrate and invertebrate examples of nonspiking neurons have been reported.
Lessons from the Lobster's Stomach
But a better place to study nonspiking cells turns out to be located near the stomach of the lobster: the stomatogastric ganglion has just thirty neurons, enabling the researcher to get to know them as individuals and give them names. With a population of a classroom rather than anonymous millions as in the retina, one has the ability to study the interaction between a pair of neurons, just as a classroom teacher soon knows which children are exchanging secret messages under the desk tops.
At least one of the thirty stomatogastric neurons is nonspiking, refusing to fire a spike even under extreme voltage changes. However, by observing the-downstream neuron, one can easily see a response. Because the synapse (the site of the functional connection between two neurons) is inhibitory, a positive voltage change in the "presynaptic" neuron causes a negative voltage change in the "postsynaptic" neuron.
By doubling the voltage change in the nonspiking presynaptic neuron, one can almost double the voltage response in the postsynaptic neuron. But if one halves the original voltage in the nonspiking neuron, nothing may be seen postsynaptically. There is a threshold for neurotransmitter release, a voltage below which the release rate is undetectable. But once above the threshold, more voltage causes more neurotransmitter release, which causes a more negative postsynaptic voltage response. So this nonspiking neuron is analog, but with a threshold. From less direct evidence, it would seem that many retinal neurons have similar analog characteristics.
But Katherine Graubard, who extensively studied the lobster stomatogastric ganglion nonspiking neuron, demonstrated a far broader principle of nonspiking computation in collaboration with Daniel K. Hartline and Jonathan Raper. In experiments at the University of California San Diego and at the University of Washington in Seattle, it was found that even the spiking neurons of the ganglion were also using nonspiking neurotransmitter release. Because their threshold for releasing neurotransmitter was lower than their threshold for triggering spikes, fluctuations in net voltage could be communicated to other neurons without spikes--and when a spike was triggered, it added an additional squirt of neurotransmitter, just for emphasis. Altogether, quite a different picture of neuronal computation than the digital view of the "physicist's fallacy."
In addition to further depreciating an outworn early "principle" of neural functioning, the lobster experiments illustrate that the relevant dichotomy is not analog versus digital, nor spiking versus nonspiking neurons. It is spiking versus nonspiking computation. Some neurons, such as most cell types in the retina, use nonspiking computation exclusively. Others, such as the majority in the lobster stomatogastric ganglion, use a mixture of spiking and nonspiking methods. To broadly categorize cells as spiking" or "nonspiking" would be to miss an essential point (an otherwise excellent symposium volume was entitled Neurones Without Impulses, something of a takeoff on the famous Animals Without Backbones though hardly expressing as fundamental a dichotomy).
It would be tempting to say that spikes are merely a method for handling the long-distance problem: that whenever the input synapses are more than a few millimeters from the presynaptic neurotransmitter-releasing regions of the cell, spikes are used to initiate an all-or-nothing event which can be reproduced at a series of booster stations along the way. While this seems to be true, its converse is not. Some small cells use spikes, even though voltage attenuation over the cell's short length would not seem to pose a problem. As in the lobster cells and some vertebrate retinal neurons, spikes may also be used for emphasis, not merely for long-distance amplification.
Is the Neuron Ambidextrous?
The neuron's fundamental role is to control its rate of neurotransmitter release, regulating it in accordance with its microenvironment. For pacemaker neurons, the regulation is largely hormonal: special messenger molecules delivered by the bloodstream to the vicinity of the neuron and then diffusing the rest of the way. For most neurons, the special molecule is delivered up close: it is released from another neuron just a membrane's thickness away from the cell membrane in a complex called the synapse. At such a range, they can hardly miss their target. Binding to special receptor molecules on the cell surface, they either open up ionic channels through the membrane, or secondarily regulate internal chemical reactions. Neither, however, allows input strength to be compared between excitatory and inhibitory inputs, which use different neurotransmitters and receptor molecules. To avoid the adding-apples-and-oranges problem, neurons use a common currency--measured in volts rather than dollars.
If a neurotransmitter-releasing region is nearby, its release rate of neurotransmitter molecules will be changed as a result. But if it is farther away, the voltage change will be smaller, the attenuation depending upon the membrane's leakiness and the cell's geometry (branching patterns can be quite important). Several millimeters is merely a useful outside limit, beyond which some other process (such as the spike) is needed to keep the electrical message from dying out.
Some neurons have transmitter-releasing regions only at a long distance from the input synapses; the motor neurons of the spinal cord, which activate the muscles, operate in this manner and thus require spikes (the cat's spinal motor neurons are the "classic" example of neuron functioning, being the first to be extensively studied). Others have such distant release sites but also nearby ones, scattered among the same cell processes which receive the inputs from upstream neurons. The nonspiking neurons tend to have this nonsegregated arrangement which, if their transmitter release has a low voltage threshold, allows a single neuron in theory to perform many different computations, one at each of its release sites--there is nothing to make such nonspiking release uniform at each different release site.
This breaks down the long-prized functional unity of the neuron. In spinal motor neurons, for example, the anatomic unit is also the functional unit because the release sites are all segregated at the far end of the cell in the muscle, all releasing together when a spike arrives. But nonspiking computation allows for multiple computations to be performed in a single cell, as each site sees a different weighting of inputs. Certainly a cell can combine spiking and nonspiking computational schemes, and the lobster neurons often provide good examples. They also show that nonspiking computation may occur in an elongated cell, just not in the elongated portion itself. This relaxation of the small-cell criterion means that many more types of neurons in both vertebrate and invertebrate brains may use nonspiking computation in the manner seen in the small nonspiking neurons, and that such a neuron may be communicating different messages to different cells downstream from it.
Spikes in a New Light
While nonspiking computation is usually analog (though with a threshold), this does not mean that spiking computation is digital, as in the "physicist's fallacy." Indeed, it too is usually analog with a threshold. Most cells are like pacemakers, in that they can produce a rhythmic discharge of spikes at a certain rate. Changes in the excitatory and inhibitory inputs, however, modify the rate. For the release sites far away from the inputs, the only voltage change they ever see is the spike. The rate of transmitter release is thus controlled by the spike rate. Which means that the net input voltage varies the release rate just as in the nonspiking neurons--but by using the spike rate as a middleman.
Thus analog computation in spiking neurons is controlled by the mechanisms that vary the spike production rate when the net input voltage goes up or down. As noted in Chapter 6, this works much like the control pedal of a sewing machine. For gentle presses on the pedal, nothing happens. When its threshold is reached, the machine starts stitching at a certain minimum rate; harder presses, and it speeds up proportionally to pedal pressure. So too the neuron regulates its spike firing rate with net voltage from the inputs, having a threshold below which it does nothing, above which it grades firing rate with input. Spinal motor neurons work this way, a fact known since the 1930s. More recent studies of this neural oscillator by Daniel Kernell in Amsterdam, myself, and many others have established how neurons encode information for long-distance transmission.
"Emphasis" is sometimes superimposed upon the usual voltage-to-rate code in the form of "double spikes," which release much more than twice the usual amount of neurotransmitter at the distant release sites. This patterning of the impulse train (common in many normal neurons but exaggerated in epilepsy and chronic pain disorders) may be more uniquely a property of spiking neurons. It is certainly common in the normal thalamus. But it took many decades of cellular neurophysiology before the patterning was appreciated--and the study of nonspiking computation is still only a decade old, though Ted Bullock foresaw it in 1947. [See 1997 update note at the end]
Studies of neuronal computation schemes are probably only beginning. Their value lies not in how to build a faster robot or computer--designing from scratch would surely be faster--but in understanding what goes on inside a brain and how it came to be. Evolution builds neural machines in stages, trying variations on what already works well enough to survive the vicissitudes of selection pressures (brains are, as someone once noted, juryrigged). Pittendrigh noted that adaptive organization is "a patchwork of makeshifts pieced together, as it were, from what was available when opportunity knocked, and accepted in the hindsight, not the foresight, of natural selection."
It looks as if both spikeless and spiking computations are basic stages which have survived well enough to form the foundations of the higher functions of the brain. They are both ancient solutions to the problem of adding apples and oranges.
1997 UPDATE: At least for the pyramidal neurons of the newer parts of cerebral cortex, there has been a substantial challenge to the voltage-to-rate view of spike coding that seems to work so well in spinal motorneurons: Softky and Koch (1993) have persuaded me that the cortical neurons just don't normally fire rhythmically for very long, as if all their spikes were generated by something like an AND gate. See theend notes for this chapter and my discussion in The Cerebral Code (1996).
The Throwing Madonna:
Essays on the Brain (McGraw-Hill 1983, Bantam 1991) is a group of 17 essays: The Throwing Madonna; The Lovable Cat: Mimicry Strikes Again; Woman the Toolmaker? Did Throwing Stones Lead to Bigger Brains? The Ratchets of Social Evolution; The Computer as Metaphor in Neurobiology; Last Year in Jerusalem; Computing Without Nerve Impulses; Aplysia, the Hare of the Ocean; Left Brain, Right Brain: Science or the New Phrenology? What to Do About Tic Douloureux; The Woodrow Wilson Story; Thinking Clearly About Schizophrenia; Of Cancer Pain, Magic Bullets, and Humor; Linguistics and the Brain's Buffer; Probing Language Cortex: The Second Wave; and The Creation Myth, Updated: A Scenario for Humankind. Note that my throwing theory for language origins (last 3 essays) has nothing to do with the title essay: THE THROWING MADONNA is a parody (involving maternal heartbeat sounds!) on the typically-male theories of handedness.
Many libraries have it (try the OCLC on-line listing, which cryptically shows the libraries that own a copy), and used bookstores may have either the 1983 or the 1991 edition.
- Powell's Books in Portland lists used copies in their web database.
Email || Home Page || The Calvin Bookshelf || Table of Contents || End notes for this chapter || Continue to read next chapter