All scribes, however careful, are bound to make a few errors, and some are not above a little willful “improvement.” If they all copied from a single master original, meaning would not be greatly perverted. But let copies be made from other copies, wbicb in their turn were made from other copies, and errors will start to become cumulative and serious. We tend to regard
erratic copying as a bad thing, and in tbe case of human
documents it is bard to think of examples wbere errors can be described as improvements. I suppose tbe scbolars of tbe Septuagint could at least be said to bave started something big wben tbey mistranslated tbe Hebrew word for “young woman” into tbe Greek word for “virgin,” coming up witb tbe prophecy: “Behold a virgin sball conceive and bear a son. . . . ” Anyway, as we sball see, erratic copying in biological replicators can in a real sense give rise to improvement, and it was essential for tbe progressive evolution of life that some errors were made.
RICHARD DAWKINS, The Selfish Gene, 1976
THIS is THE HARDEST CHAPTER in the whole book. In it, I have to delve into the neuroanatomy and neurophysiology of cortical neurons, importing lessons from such seemingly unrelated subjects as synchronously flashing fireflies. By the end of this chapter, I will have shown how copying could arise in neocortex. By the end of the sixth chapter, the cortical equivalents of all the darwinian essentials and all the accelerating factors will
have been examined. But it gets easier, not harder, starting with the fourth chapter.
Neurophysiologists distinguish between cell properties and circuit properties, much as biologists distinguish between genotype and phenotype. Some phenomena are clearly due to the circuit rather than the cells involved, to the wiring rather than the components — a new property “emerges” from the particular combination. You won’t find it in any one neuron. The classical example of an emergent property involves lateral inhibition and it is the reason that Keffer Hartline got a Nobel Prize in 1967.
Thanks to local activity contributing to a ring of depression in surrounding neurons, lateral inhibition sharpens up fuzzy bound- aries. Compound eyes, the manynarrow-angle photoreceptors of which provide an extreme case of fuzzy optics, have a series of inhibitory interconnections that are capable of recreating a light- dark boundary in the environment, restoring much of what was lost.
But lateral inhibition also has a tendency to produce features where none exist, illusions such as the Mach bands that you see if looking through a narrow slit between your fingers. Georg von Bekesy, whose studies of such sideways interactions in the cochlea were the subject of his 1961 Nobel Prize, also produced illusions from skin surfaces, to illustrate the generality of the lateral inhibition principles. Antagonistic surrounds (“Mexican hats”)
are common in all the first half-dozenstages of the analysis of a visual image, though they become somewhat elongated and asymmetric (“Australian bush hats”) in primary visual cortex. Because of the many axon collaterals that branch laterally in neocortex,
lateral inhibition extends several millimeters.
Both the sharpening of fuzzy boundaries and the illusions are emergent properties of a laterally inhibiting neural network. What might be the emergent consequences of lateral excitation?
THERE IS GOOD REASON to worry about recurrent excitation. It is potentially regenerative, in the same sense as a string of fire
PyrwmUM neuron mama eond many
fPOGWVWtt GOmtCTtV VTWGnOB WtOIWtyto neighboring areas of cortex, the hash for both lateral Inhibition and
crackers. It is also the most prominent wiring principle of mam- malian neocortex.
A few words about cerebral cortex, the icing on the brain’s cake:. in this cake, it’s the frosting that has the appearance of layering! Six layers are usually identified on the basis of cell size
or axon packing density, though we sometimes subdivide it further (in primary visual cortex, one talks about layers 4a, 4b, 4ca, and 4cP). At other times, we lump layers together: when I mention the “superficial layers/’ I’m combining layers 1,2, and 3.
Part of the monkey superficial pyramidal
neuron reconstructed by McGuIre et al
(J. Comp. Neurol. 1991), showing axon
terminals to Immediate neighbors (thin
axons amidst dendritic tree) as well as
branches to cells a microcolumn away.
Furthermore, there are three functional groupings that have become apparent: on the analogy to the mail boxes stacked on many a desk, layer 4 could be said to be the IN box of neocortex, because most of the inputs from the thalamus terminate there. The deep layers could be called the OUT box, as pyramidal neurons of layers 5 and 6 send axons outside the cortex, back to thalamus or down to the spinal cord, and so forth. The neurons of the superficial layers seem to constitute the INTERNAL mailbox of the neocortex, specializing in the interoffice memos. Interactions among the superficial pyramidal neurons are what this book is mostly about, as these neurons seem capable of implementing a darwinian copying competition, one that can shape up quality from humble beginnings.
The axons of the superficial pyramidal cells are prominent in the corpus callosum and other long corticocortical paths, but also in the intrinsic horizontal connections: those axon branches that never leave the superficial layers because they run sideways. They preferentially terminate on other superficial pyramidal neu- rons — and in a patterned manner, too. Some axon branches go to near neighbors, but the ones that go further ignore a whole series of intermediate neurons before communicating with those about 0.5 mm distant.
Those sparsely populated gaps are something like the Sherlock Holmes story about the dog that didn’t bark in the night. It took a long time before anyone noticed this fact. In 1975 came the first hint of these gap patterns. In 1982, when Jennifer Lund and Kathleen Rockland first studied the gaps in the superficial layers’ intrinsic horizontal connections, it was in the visual cortex of the tree shrew. Though the gap distance varies, we now know that it is a common arrangement for many areas of neocortex, and for many animal species. Thanks to the detailed reconstructions of several HRP-injected superficial pyramidal neurons by Barbara McGuire and her colleagues, we also know that these synaptic connections are likely to be excitatory, probably using glutamate as their neurotransmitter, and that their predominant targets are other superficial pyramidal neurons.
Their axons have dozens of branches, going sideways in many radial directions, fanning out eventually into thousands of axon terminals. Although no single superficial pyramidal neuron has enough terminals to fill in a doughnut, we might expect a small minicolumn group of such neurons to produce a ring of excitation, as well as the central spot of excitation from the branches to immediate neighbors. Point-to-area is the more common arrangement for axon projections, such as those of the pyramidal neurons of the deep layers. Recurrent inhibition is also seen, but only the recurrent excitation of the superficial layers of neocortex has this Sherlock-Holmes feature of prominent silent gaps.
Optical imaging techniques that look down on the brain’s surface are now capable of resolving a spread of activity in cortex. Stimulation of a restricted area of retina, of a type that classically would be expected to concentrate cortical activity in only one area
branching axon activating cortical loops
of the exposed cortical surface, is now seen to
contribute to multiple hot spots of activity at
macrocolumnar separations, much as predicted.
The neocortical versions of long-term potentiat-
ion (LTP) are also concentrated in the superficial
layers. We know that N-methyl-D-aspartate
(NMDA) types of postsynaptic receptors, which
I have the unusual characteristic of augmenting their strength when inputs arrive in clusters (such as quasi-synchronouslyfrom different sources), are especially
common in the superficial layers.
All of this raises the possibility of self-reexciting loops, not un- like the reverberating circuits postulated for the spinal cord by Rafael Lorente de N6 in 1938, in the very first volume of the Journal of Neurophysiology. If the synaptic strengths are high enough, and the paths long enough to escape the refractory periods that would otherwise limit re-excitation,closed loops of activity ought to be possible, impulses chasing their tails. Moshe Abeles, whose Jerusalem lab often observes more than a dozen cortical neurons at a time, has seen some precise impulse timing
of one neuron, relative to another, in premotor and prefrontal cortex neuron ensembles. It is unknown whether or not these firing patterns represent reverberation, in Lorente’s original sense of recirculating loops. These long,precisely-timed firing patterns are important for the notion of spatiotemporal patterns that I will later develop.
EMERGENT SYNCHRONY is well known as a commonplace conseq- uence of recurrent excitation, one that ought to be seen with even weak connection strengths and short paths. In 1665, the Dutch physicist Christiaan Huygens noticed that two pendulum clocks hanging from a common support were synchronized. When he
disturbed the synchrony, it returned within a half hour. Harmonic oscillators are slower to entrain than nonlinear relaxation oscillators, which can take just a few cycles.
The most famous example of entrairunent is probably menstr- ual cycles in women’s dormitories. More dramatic in appearance is a whole tree filled with little lights, flashing in unison. No, I don’t mean a Christmas tree wired up, under the control of a single flasher — there’s a natural, wireless example based on hundreds of independent oscillators. The little lights are hundreds of fireflies, and they have no leader to set the pace.
Imagine a tree thirty-five to forty feet high, apparently with a firefly on every leaf, and all the fireflies flashing in perfect unison at a rate of about three times in two seconds, the tree being in complete darkness between flashes. Imagine a tenth of a mile of river front with an unbroken line of mangrove trees with fireflies on every leaf flashing in synchronization, the insects on the trees at the ends of the line acting in perfect unison with those between. Then, if one’s imagination is sufficiently vivid, he may form some conception of this amazing spectacle.
It doesn’t require any elaborate notions of mimicry to account for the firefly entrainment; even small tendencies to advance to the next flash when stimulated with light will suffice to create a “rush hour.” Furthermore, you usually do not see waves propagating through such a population, except perhaps when the flashing is just beginning or ending. Even in cortical simulations with prop- agation delays,near-synchrony is seen, in much the way (anomal- ous dispersion) that some velocities can exceed the speed of light.
Weak mutual re-excitation (a few percent of threshold) is quite sufficient to entrain; one need not postulate strong connection strengths in the manner needed for Lorente’s recirculating chains. So long as the neurons (or fireflies) already have enough input to fire repeatedly, there will be an entrainment tendency if they mutually re-excite one another. And that is exactly what super- ficial pyramidal neurons, 0.5 mm apart, seem so likely to do. The triple combination — mutualre-excitation, silent gaps that focus it, and the resulting entrainment tendencies — is what gives the
superficial layers of neocortex the potential of being a Darwin Machine.
LOOKING DOWN FROM ON HIGH at the superficial layers of neocortex, in what the neuroanatomists call “tangential slices/’ is like looking down on a forest from a balloon. Any one neuron is seen in a top-down perspective, orthogonal to that seen from the side in the usual surface-to-depth slice. Like the branches of any one tree, any one neuron has a dendritic tree, but also an axon tree, much as the foresf s tree has branching roots below ground.
The axon of a single superficial pyramidal neuron will be seen to spread in many directions. Though sensory neurons and motor neurons may vary, the average interneuron sends out as many synapses as it receives, usually between 2,000 and 10,000. Not enough radial plots have yet been done to know how symmetric the horizontal spread is, but it seems clear that the axon branches travel in many directions from the cell.
GIVEN standard length excitatory axons,
…recurrent excitation betweensome cell pairs producesentrained firing patterns.
An entrained pair tends to recruit additional cells
that are equidistant.
…and so create a
The distance from the cell body to the center of the axon term- inal cluster, studied mostly in the side views, is not the same in all cortical areas. That “0.5 mm” mentioned earlier is really as small as 0.4 mm (in primary visual cortex of monkeys) or as large as 0.85 mm (in sensorimotor cortex). It scales with the width of the basal dendritic tree. I’ll use 0.5 mm as my standard example of this local metric; it corresponds to a basal dendritic tree of about 0.25
mm spread, which is also about the spread of one cluster of axon terminals and the extent of one silent gap.
If two superficial pyramidal neurons, about 0.5 mm apart, are interested in the same features because of similar inputs and thresholds, their spike trains ought to start exhibiting occasional spike synchrony. It need not be all the spikes from each neuron for the following analysis to be relevant; only some of their spikes need synchronize via the recurrent excitation.
There should also be some minor tendency for two such cells, already firing repeatedly, to recruit another cell 0.5 mm away that is almost active. If that third superficial pyramidal neuron becomes active, we should see threeoften-synchronized neurons forming an equilateral triangle. But that is not the end of it: there is a second site receiving synchronous input from the
parent pair (this is exactly like that elementary exercise in plane geometry where a compass is used to bisect a line or drop a perpendicular). So a fourth neuron might join the chorus.
And because the third and fourth cells provide new annuli of excitation, either can combine with one of the first pair to bring a fifth point into synchrony. What we have, it is apparent, is a mechanism for forming up a triangular array of some size, nodes of synchronized activity 0.5 mm
from corresponding cells of this chorus. It could work either by synchronizing preexisting activity or by recruiting otherwise sub- threshold neurons at the nodes. Once a potential node is surround- ed by a few synchronous nodes exciting it, there ought to be a hot spot, an unusually effective con- vergence of simultaneous inputs.
This triangular array annexat-
ion tendency is not unlimited. (Regions with insufficiently excited neurons, as I discuss in the latter part of chapter 6, provide barriersto any further empire-building.) And the triangular array is surely ephemeral, here now and gone in seconds. When it is shut
down by enough inhibition (or reduction of excitation), it will be as if a blackboard had been erased.
Traces will linger, however, in much the way that blackboards retain ghostly images of former patterns. The synaptic strengths should remain changed for a while; indeed, the synchrony- sensitive long-term potentiation of the superficial neocortical layers suggests that synaptic strength can remain augmented for many minutes. This will make it easier to recreate the active form of the triangular array — perhaps not all of its spatial extent, but part of it.
THE LATTICE CONNEcnviTY seen in the anatomy, it should be said, does not fall into neat triangular arrays, measured by distance in the tangential plane of section. Though the neuroanatomists speak of”polka-dot” patterns and ‘lattices” for the axon terminal clusters in the superficial layers, the spacing of the clusters is only roughly triangular. Of course, adjusting conduction velocity or synaptic delay during atune-up period could make a triangular array, when replotted as “driving time” rather than distance.
But not even an equal conduction time, for converging simultaneously on a potential recruit, is actually required for the present theory. Though exact synchrony has been convenient for introducing the principles, all that triangular arrays require in the long run is a prenatal tune-up period that results in a good- enough self-organization, so that most of the six surrounding
nodes produce axon clusters that mutually overlap in a manner that aids entrainment. It may not matter to thisself-organizing principle what an external observer would find “regular.” I’ll stick to triangular array terminology for the theory, but don’t expect to find exact triangles in either the anatomy or physiology, only good-enoughapproximations.
FROM A PAIR OF LIKE-MINDED CELLS, we see the possibility of a large
chorus, all singing in synchrony. Furthermore, it’s a chorus that can recruit additional members out on its edges. Like a choir standing on risers, these singers tend to space themselves so that each is standing in between two singers on the row below. The choir isn’t as perfect a triangular array as the fruit displays at your corner grocery, but it’s a good enough approximation to the familiar packing principle.
So far, this choir only chants in unison. It’s monomaniacal, perhaps only interested in one feature of the stimulus. It’s surely not the true Hebbian cell-assembly The choir corresponding to a concept representation would surely sing in parts, just as sopranos carry one melody and the altos another, each group having different interests. We will need polyphony for harmonious categories, not just chants.
A Compressed. Code Emerges
Tkese self-re-exciting systems [cell-assemblies] could not consist of one circuit of two or three neurons. but must have a number of circuits. . . . I could assume
that when a number of neurons in the cortex are
excited hy a given sensory input they tend to hecome interconnected, some of them at least forming a multicircuit closed system. . . . The idea then  was that a perceptconsists of assemblies excited sensorily, a concept of assemblies excited centrally, hy other assemblies.
D. O. HEBB, 1980
POLYPHONIC MUSIC elaborated on chants by combining a number of independent but harmonizing melodies. The task of this chapter is considerably easier: we only have to
combine notes, each from a different triangular array, into a simple melody (polyphony, as chapter 7 will show, is a useful analogy to whafs going on in category representations). While this chapter starts with some issues regarding the cortical landscape from which the choir sings, it soon progresses to an abstraction much like written music.
Happily, by the end of this chapter, we will see the choir coalesce into sections, each of which sings the complete song. Unlike the placements favored by choirmasters, the sopranos are not grouped together; if s more like each section has one soprano, one alto, one bass, and so forth, each singing a different part. Each
section is surrounded by neighbors, sections that are similarly diverse. You might think that this would make it difficult for a choirmaster — if one exists — to conduct, but remember that string quartets get along nicely without a conductor, and what I will describe here is a chorus of string quartets.
In cortex, it looks as if one string-quartet section occupies a space that is hexagonal in shape and about 0.5 mm across. It could constitute the most elementary version of Hebb’s cell- assembly, one that could represent a word, a face, or a pronunciation. Cloning indeed clues us in, suggesting what the relevant code might be — that characteristic pattern needed for the first darwinian essential. To get there, however, we first need to consider a few more aspects of the geometry and its relevant neurophysiology.
THE “HOT SPOT” COULD BE SIZEABLE, because of the width of those
0.25 mm clusters of terminals. But the history of neurophysiology suggests that, functionally speaking, the hot spot might be far smaller, perhaps as small as a minicolumn (0.03 mm diameter, and a small percent of the area). Before returning to the spatial extent of a triangular array, let us consider the size of its nodes (I’ll use node as a punctate theoretical term, with hot spot referring to physiologists measure, and axon terminal clusters referring to what anatomists see).
Anatomical connectivity (the fanout of the axon terminals, the width of dendritic trees) is usually far more widespread than physiological responsiveness (such as receptive field centers). Indeed, at a few removes, every neuron in the brain can potent- ially connect to every other neuron — but such extensive funneling rarely happens. Antagonistic surrounds serve to concentrate things. In the retina, for example, wide areas of the photoreceptor mosaic would seem to have paths to a second-order cell, but a bipolar cell usually has a far smaller receptive field center, thanks to flanking inhibition (or it has an inhibited center with flanking excitation, the other type of antagonistic center- surround arrangement commonly seen) except during dark adaptation.
In addition to antagonistic arrangements, some axon terminals seem to have very weak synaptic strengths; indeed, we sometimes talk of “silent synapses.” Anatomically, they’re there; physio- logically, they’re undetectable most of the time. An example of this second type of physiological focusing is the projection from a thalamic neuron, one specializing in just one finger tip, to the hand map in cerebral cortex. Its axon terminal branches seem to span much of the hand’s map, but, when you look at the cortical neurons they’re feeding, you find that they typically have small receptive fields, little larger than those of thalamic neurons. Another indicator of size: visual cortex cells at millimeter separat- ions with similar orientation preference are interconnected, suggesting the possibility of hot spots that are as small as those 0.03 mm minicolumns.
The superficial pyramidal neurons are not the only cells contacted by the intrinsic axon collaterals; about 20 percent of the axon terminals are onto smooth stellate neurons (that themselves produce GABAergic inhibition), presumably contributing to forms of flanking inhibition that reduce the size of the hot spot.
THERE SHOULD BE A MARKED
STABILITY of the triangular arrays formed by the hot spots, even under various perturbations. Let us suppose that a triangular array is firing in a repeated cycle. And that one point in the midst of a tri- angular array tries to fire out of sync, later than its neighbors.
It has, however, six neighbors that are all sending it synchronous
inputs at the standard time in the cycle, thus tending to correct it (actually, it may have a dozen because the axons tend to have several terminal clusters 0.5 mm apart). The same argument applies if the idiosyncratic neuron attempts to omit a impulse. Similarly, early firings tend to be corrected the next time around if the neuron has any tendency to produce longer-than-average interimpulse intervals following an earlier-than-average impulse.
Spatially, there is the aforementioned tendency to focus synch- ronous excitation on the center of the hot spot. What both tenden- cies mean is that, like a crystal forming, we might expect to see some standardization. One might almost think of it as error correction.
Another force for standardization may be the minicolumn of association cortex (represented as a raised bump in many of my “tangential slice” illustrations), those hundred cells organized around a dendritic bundle. The well-studiedorientation columns of primary visual cortex seem to prefer similar stimuli, and the superficial pyramidal neurons have many close-in axon collaterals before the silent gap that often excite near neighbors. These suggest that a number of the minicolumn’s 39 superficial pyra- midal neurons may be synchronously activated. Because of this, I have not found it useful to distinguish between the individual superficial pyramidal neuron and all the superficial pyramidal neurons within a given minicolumn. The six neighbors could be as many as 39×6=234 superficial pyramidal neurons, speaking together to exact conformity.
One may thus think of the “cells” and “nodes” as really work- alike minicolumns. This tendency to act as a group could eliminate the “holes” in the lattice that might otherwise result from the incomplete “polka-dot” annuli of an individual superfic- ial pyramidal neuron, and give rise to thepoint-to-annulus prop- erty that I infer.
DOES A HOT SPOT FORM at precisely the location suggested by the node of a triangular array? It need not, of course, if other wiring principles override the triangular tendency; for example, making connections with other orientation columns of the same orientation angle might obscure triangular tendencies in primary visual cortex. But this is a theory for association cortex, not for the most specialized of cortical areas. Clustered recurrent excitatory terminals have indeed been found in many neocortical areas of many animal species.
Any one superficial pyramidal neuron’s annulus isn’t perfect, of course, because its terminal clusters do not provide full coverage. But when six minicolumn’s worth of them overlap at
the same node (even more, actually, because of the tendency of the axon to continue across another silent gap to produce another cluster), there will be one point that will have more input than others, and this ought to help define the node more narrowly. Furthermore, the cells implementing surround inhibition in the superficial layers of neocortex, the large stellate cells, have axons that reach far enough (except in rats) — so six inhibitory point-to-area circles also help define a node via subtraction.
As we shall see in the next chapter, some synaptic augment- ation mechanisms, such as those at the NMDA synapses, are available for rewarding such convergence. The NMDA synapses have a remarkably imprecise notion of synchrony, so aug- mentation per se might not be sensitive to the equal conduction distances that define the triangle. Exact synchrony of synaptic potentials depends on identical conduction times, all else being equal. Yet ordinary spatial summation — one bump standing on the shoulder of another postsynaptic potential — can define synchrony with considerable precision if the threshold for impulse production can be exceeded only by the optimal overlap, peak atop peak. Higher thresholds will shrink the size of hot spots and, if the conduction speeds are equal, center them equidistant from their inputs.
LET US NOW CONSIDER THE ENSEMBLE PROBLEM in the context of this
tendency to recruit a triangular chorus. We don’t have just one triangular array, but multiple ones that interdigitate.
When looking at a banana, various types of feature detectors ought to be interested; let us say that the parents of one triangular array (A) are fans of the yellow color of the banana. Other superficial pyramidal neurons will likely be interested in one or another of the tangents to the banana’s profile, and so one might get another triangular array (B) forming up to specialize in horizonal line representation. This second horizontal array need not be synchronized with the yellow array (as a common form of the binding theory assumes) for present purposes.
Furthermore, the horizontal-tangenf s array might start several millimeters away from where the yellow array starts. One can easily imagine a half-dozen separate features, each with its own