Chapter 7 Chapter 6 The Operational Mechanism of the Brain "Darwin Machine"

Our intellect comprehends the phenomenal world by the principle of transcendental schemes... This is a skill so deep in the human mind that it is difficult for us to guess the secrets employed by nature.

"Why," said the Dodo, "it is best explained by demonstration." Is this chapter really required?If you can read the previous chapter without realizing what's missing, then, in that sense, this chapter is redundant. All of this depends on how comfortable you are with the group composition.Some people don't want to know more, we say: "ignore the details, just have an implementation overview".But this chapter is not about the details omitted in the previous chapter. It is written from a different perspective, in a "bottom-up" way, rather than starting from deduced principles. Unfortunately, the principle is much like a composition diagram - a convenient schematic fiction.The real group village is composed of a series of information and decision-making processes, and boxes and symbols alone cannot cover its connotation.Composition diagrams fail to take into account the factor of "people" and how they talk to each other, and also fail to take into account "institutional memory".Nor does it take into account how specialists can become generalists; how decisions made at one level interact with decisions made at another level.Any pictorial interpretation of the brain will suffer from these shortcomings of compositional diagrams. This account of intelligence has so far failed to give much thought to neurons and how these brain cells exchange information with each other; how they remember past events; how they work together to make decisions at local and regional scales .Although some of these processes are unknown, it is now possible to draw a plausible overview of the competition between different brain codes. A general rule that works well when discussing scientific issues is to give a special case, even if it is only a possible but not recognized mechanism. This is exactly what this chapter discusses.I will give an example of how our cerebral cortex can function as a kind of Darwinian machine; idea).This example shows how, offline, we can simulate actions that will occur in the real world, an ability we believe is key to enabling intelligence to make correct predictions. What is always going on in the minds of many ordinary people, and those who disagree with the idea that "computers have minds," is that it is impossible to imagine a mechanism capable of producing a mind.Using the building blocks described in this chapter, I can imagine how a thinking machine could be built.Your criteria can be different, but it's your choice.In just one chapter of text, you can see a "bottom-up" mechanical example of how our mental processes operate both consciously and subconsciously for the novel and the routine . Unless the brain is dead, the gray matter of the brain isn't actually gray.In a living brain, gray matter has a rich blood supply.Just imagine the reddish taupe color of a river after a thunderstorm, and you'll get the right impression of the color of active "grey matter". But the brain's white matter is actually porcelain white, the color of the fat it has that coats and insulates the long, slender protrusions of nerve cells.These protrusions, called "protrusions," are similar to wires that carry the neuron's output to nearby or distant targets. "Myelin" is the technical name for the fatty insulation, and white matter is actually a collection of nerve fibers that go to the office, much like you might find in the basement of a telecommunications center building.It is these insulating fibers that make up the bulk of the brain, interconnecting the much smaller parts of the brain that carry out important functions. At one end of the axon is the spherical, swollen cell body of the neuron, containing the nucleus.It contains the DNA templates for the daily functioning and maintenance of cells.There are many tree-like branches extending from the cell body, called dendrites.This part of the neurons does not have the white myelin, so they are gray when aggregated.The other end of a neuron's axon usually makes contact with the dendrite of a downstream neuron. (If you look closely with an electron microscope, you'll see a narrow gap between two neurons, called a synapse.) The upstream neuron releases tiny amounts of neurotransmitters to the synapse, a "no man's land" , and then diffuse to the membranes of downstream neurons, opening channels in some membranes. (Although some retrograde neurotransmitters are also present, a synapse is usually one-way traffic, hence the term "upstream" and "downstream" neurons.) Taken together, individual neurons look a lot like the roots of a shrub or herb such as ginger.It is a typical computational unit that combines the effects of thousands of inputs, mostly excitatory, some inhibitory—much like deposits and withdrawals in a bank account, and then responds with a voice to Its thousands of listeners (downstream neurons) speak.The information sent from this "checking account" is mainly about the state of its "account balance", and how quickly the balance is increasing.No message will be sent unless the balance exceeds a certain threshold.Large deposits generate large information, just like paying interest with dividends.But just as piano keys only sound when you strike them hard enough, cortical neurons are usually quiet unless the input signal is appropriate, and their output is proportional to how much they are stimulated by the "account balance." (Oversimplified binary models often treat a neuron as a harpsichord key, with a threshold but no graded change in volume when struck harder.) Although the message of neurons with short axons may be simpler, neurons with axons longer than about 0.5 mm always employ a signal amplification device; an impulse, a transient voltage change of a standard size (like a dial loudness of the clavichord keys).When the impulse is amplified and sent to a speaker, it sounds like a "wow", which we call a neuron "discharge".To avoid the limitations of standard size changes, impulses are often repeated at a rate proportional to the "account balance".Sometimes, especially in the cerebral cortex, it takes only a few inputs out of thousands to synergize to trigger an impulse. What's really interesting is the gray matter in the cortex, because that's where most of the novel associations are made, and it's there that the look of a comb matches the feel of a comb in your hand.The brain codes for vision and touch are different, but they are somehow linked in the cortex, as well as the brain codes for hearing the sound comb or hearing the sound made when you pluck the teeth of a comb.No matter which of the above forms, you can always be sure that it is a bum.Therefore, it has been hypothesized that there may be specialized sites in the cortex, which we call associative memory pools, where signals from different sensory modalities pool.On the signal side, you've also linked the brain codes for pronouncing the word "comb" and brushing your hair with a comb.Thus there is a connection between the sensations evoked by the word "comb" and various motor manifestations, and we can expect to discover a dozen different cortical codes associated with combs. The cortical regions that perform all these associated functions are the thin icing on the white matter cake.The cerebral cortex is only 2 mm thick, but it has deep folds.The neocortex has an extremely uniform neuronal density (except for one layer of primary visual cortex).Whether it's the language cortex or the motor cortex, if you make a grid on the surface of the cortex, there are about 148,000 neurons per square millimeter of neocortex.However, if viewed from the side, some regional differences are visible in the layers at a depth of 2 mm. It is not the cake itself that has the layered structure, but the icing on the cake, much like the crescent-shaped crust of a layered pie.The deepest layers of the neocortex act like a "sending box," with fibers exiting the cortex to distant subcortical structures such as the thalamus or spinal cord.The middle layer is a "receiving box", where fibers arrive from places such as the thalamus.The superficial layers act much like interdepartmental mailboxes ("internal mailboxes"), forming connections with adjacent and distant superficial layers.Their axons pass through the body to the other side of the brain, but most "inner letterbox" mail is distributed within a localized range of a few millimeters.Such axonal branches branch laterally, rather than back through the white matter as longer "U-shaped fibers" branch. Some areas have large "inboxes" and small "outboxes," like those found on desks in editorial offices where letters from readers are processed.In addition to this stacked horizontal structure, there is also a set of ingenious vertical connections, just like columns on a newspaper. If we try to carefully analyze the various neurons in the cerebral cortex, we will find that neurons with similar functions tend to be arranged vertically in the cortex, forming a columnar structure, which is called a cortical column, and runs through the cortex. most levels.It's almost like a club, spontaneously organizing like-minded people from the crowd at a party.We naturally put names on these cortical clubs.Some of these reflect their size, others their probable profession (as far as we know). These micropillars are about 30 microns in diameter (like a thin strand of hair, closer to the silk of a spider's web).The best-known example is the orientation of the visual cortex, where all neurons seem to favor objects with lines or edges slanted at a particular angle.Neurons in one microcolumn responded best to boundaries that were sloped at 35 degrees, while neurons in another microcolumn would favor horizontal or vertical boundaries, among other things. If you look under a microscope (neurodissection technology has advanced a lot in the past century, it's still a lot of work), you will see a group of cortical neurons bundled together, like a bunch of celery.They have elongated "apical dendrites" that extend from the cell body toward the surface of the cortex.The brain body is often triangular in shape, hence the name "pyramidal neuron".The apical dendrites of these pyramidal neurons appear to be clustered in fascicles, with adjacent fascicles separated by 30 micrometers.There are 100 neurons in a microtome organized around a bundle of apical dendrites, and at any level, only a few dozen fibrous dendrites may be seen in each bundle.Neuronal fasciculation is common in cortical regions other than the visual cortex, and microcolumns appear to be a common unit of cortical organization on anatomical basis; Yuan has no idea what preference he has. Other "interest groups" are much larger and consist of many microcolumns.These so-called macrocolumns are about 0.4-1.0 mm in diameter (equivalent to a thin pencil lead), and are sometimes more like elongated curtain rolls than a cylinder.The large columns appear to be a result of how the input is organized, such that input from the left eye in the visual cortex tends to alternate with input from the right eye every 0.4 mm.The same is true for inputs from other cortical areas.Taking the cortical area just in front of the corpus callosum as an example, you can see that the input from the prefrontal cortex forms a large column, and on either side of that is a large column formed by the converging inputs from the two lobes. Cortical neurons interested in color tend to cluster (though not exclusively) in "blobs".Unlike the larger ones, the small plaques do not penetrate all layers of the cortex and are only found in the superficial layers involved in internal information exchange.They're not entirely made up of "color experts"; perhaps only 30 percent of the neurons in a small patch are color-sensitive.The distances between small patches are similar (if not identical) to those between large columns. What does the next organizational level look like?According to the change of layer thickness, there are 52 Braudmann's areas in each hemisphere of the human brain. The relative thickness varies, just as the relative volumes of incoming letters, outgoing letters, and letters exchanged between offices vary between adjacent "departments". Braudmann's area 17 is often referred to as the primary cortex.However, generally speaking, it is not yet ripe to label these areas by department on the organizational chart (for example, there are 5-6 functional areas in District 19).A Braudmann zone averages ZI square centimeters in terms of non-rugged area.If the ratio of the bark holds elsewhere, there are 10,000 macrocolumns and 1 million microcolumns in the average cortical area. The number 100 comes up over and over again.A microcell has 100 neurons, a large column has about 100 microcolumns, and a cortical area has 100 x 100 macrocolumns [which reminds me that we left out the intervening organization - "supercolumns" or "microcolumns" area", containing about 100 large columns], and the number of cortical areas in both hemispheres of the brain is just over 100. Can we scale this hundred-fold figure even further?It puts us in the yardstick of social organization: what does 100 heads make of?That reminds certain legislative bodies, such as the US Senate, that the United Nations represents more than a hundred countries. It is all well and good to understand the immutable units of the brain group, such as cortical areas or microcolumns, but we also need to understand those ephemeral workspaces of the brain—closer to automatic scratchpads and buffers, which may superimposed upon the more rigid forms of anatomical organization. To be able to work with new things, we're going to need certain empirical structures that we use temporarily and then disappear, like those little hexagonal honeycombs that appear when you forget to stir your oatmeal.Sometimes an aspect of these previously formed structures is so strong on the strength of interconnections in the brain that they revive again, in which case the experiential structure becomes a new memory or habit. In particular, we need to understand the brain's code (whose different coding patterns signify different meanings, such as each word in our vocabulary), and how it is built.At first glance, it appears that we are dealing with a four-dimensional pattern: active neurons are scattered throughout the three-dimensional cortex, and the patterns of neuron activity are functions of time.But mostly because the micropillars organize all the layers of the cortex with similar functions, most people who work with the cortex think of it as a two-dimensional sheet, much like a retina.True, the retina is about 0.3 millimeters thick and divided into several layers, but what it paints is clearly a two-dimensional image. So it is possible for us to think about the possibility of "two-dimensional ten time", in fact, that's how we understand movie screens or computer terminals.Perhaps these transparent, patterned membranes appear to overlap when different layers of the cortex do different things.Imagine flattening the cortex like a pie crust on four sheets of typed paper, with the little shiny dots like pixels on a display board, then when that cortex sees a comb, hears or speaks a word, or sends out Instruct the hand to comb the hair, what kind of pattern do we observe? Recalling may be the establishment of a spatiotemporal sequence of neuronal firing similar to that produced by memory input, except that some of the vines that are insignificant in memory formation have been pruned.The space-time pattern established during this memory is very similar to the display board in the stadium. Many small light spots are on and off, and what is established is an image.A more general form of Hebbian cell clustering should avoid pinning down that spatio-temporal pattern to specific cells, as if a display board could be rolled: the image always says the same thing, although it is Formed by different lamps. While we tend to focus on the lights that are on, those constant lights also contribute to the formation of the image.If these dark lights are suddenly turned on indiscriminately, the image will instantly become a lake.This seems similar to what happens with a concussion.When a soccer player is carried off the field with a concussion due to injury, he can tell you what game they were playing, but after 10 minutes, he can no longer remember what just happened.The damage slowly causes many neurons to "light up," so that the image becomes blurred, as is the case in bright fog.Climbers call this whitish bright fog "milk white sky". (Remember: Darkness is sometimes the result of whiteness.) What is the most basic pattern for expressing a certain meaning?A major thread seems to me to be the need for pattern replication, and long before deoxyribonucleic acid (DNA) rose to prominence, geneticists and molecular biologists were searching for a pattern that could be reliably replicated in cell division. molecular structure. When the double helix structure was "discovered by Crick and Watson in 1953 (I was living in Cambridge University at the time of writing, facing the building where they worked across the yard), people were extremely satisfied. One of the reasons is that it provides a way to make copies by borrowing DNA base complementary pairing (C and G combine, A and T pair). Untie the "zipper" of the double helix into two halves, and each "zipper" on the half "zipper" A DNA position will immediately pair with its opposite half in all the loose DNA floating in the nucleic acid "soup". amino acid sequence of a protein) paved the way. Are there similar replication mechanisms of brain activity patterns that help us identify the most relevant Hebbian cell clusters?That's what we might properly call the brain's code, because it's the most basic way of representing something (a specific meaning of a word or an imaginary object, etc.). We have not yet directly observed the replication process, and we lack tools with sufficient spatial and temporal resolution (although we are getting closer).But there are 3 reasons why I feel comfortable placing this bet. The strongest argument for the existence of a replication process is the Darwinian process itself, which is essentially a replication competition that is influenced by many environmental factors.This is really a fundamental way of turning things from chaos to order, and it would be amazing if the brain didn't use it. The replication process is also necessary for fine ballistic movements, such as throwing, which require hundreds of idiosyncratic motion command patterns in order to make a throw within the "launch time". In the last chapter we talked about the argument for artificial facsimiles: communication in the brain requires remote replication of patterns. Since 1991, my favorite candidate for a local seed circuit that produces copies of spatiotemporal patterns has been mutually reinforcing circuits in the "inner letterbox" layer.The wiring of these surface layers of the cerebral cortex is very special, for a neurophysiologist.It's almost surprising.I examine these circuits to understand how faulty activity is harnessed and why seizures and hallucinations are uncommon.These same circuits also have a certain tendency to crystallize, and they must be particularly good at replicating patterns in space and time. Of the hundreds of neurons in a microcolumn, approximately 39 are superficial pyramidal neurons (ie, whose cell bodies are located in the superficial orthogonal and solitary).It is their circuits that are unique. Like all other pyramidal neurons, they release an excitatory neurotransmitter, usually glutamate.There is nothing unique about glutamate itself, it is an amino acid more typically used as a building block of peptides and proteins.Glutamate diffuses across the synapse, opening several types of ion channels across the membrane of the downstream neuron, one of which is specialized for the passage of sodium ions, thus increasing the internal voltage of the downstream neuron. A second downstream channel activated by glutamate, called NMDA channels, allows both calcium ions and some sodium ions to enter downstream neurons. The NMDA channel is of particular interest because it is associated with long-term potentiation (LTP), a change in synaptic strength that lasts longer, lasting minutes in the neocortex. (A few minutes is actually close to the neurophysiological "short-term", but LTP sometimes lasts for several days in the hippocampus, an evolutionarily older cortical area, and this is where the "long-term" name comes from. ) LTP arises when several inputs arrive at a downstream neuron nearly simultaneously (within tens to hundreds of milliseconds), and it simply turns the "volume control" switch up for a few minutes for those input signals.Those signals are "imprints" that make it easier to re-establish a specific space-time pattern for a short period of time.As far as we know, LTP is the best basis for short-term memory, which is not destroyed by distraction, and it is also believed that LTP provides the backbone for the formation of truly long-lasting changes in synaptic structure, which are permanent. "Imprinting" helps to re-establish long-term unused space-time patterns. Those "inner letterbox" layers are where most of the NMDA channels are located, and where most of the LTP in the neocortex takes place.These superficial layers had two other unique features, both related to the connections that superficial pyramidal neurons formed with each other.On average, a cortical neuron contacts less than 10% of the total number of neurons within a radius of 0.3 mm, but about 70% of the excitatory synapses of any superficial pyramidal neuron come from other neurons within 0.3 mm. Pyramidal neurons, and therefore superficial pyramidal neurons, can be said to have an unusually strong tendency to excite each other.For neurophysiologists, this is a situation that raises eyebrows because it is an extremely unstable device prone to strong oscillations unless carefully regulated. There are also specific patterns of "regression excitatory" connections at these superficial levels that are not seen at other deeper levels.That is, the axon of a superficial pyramidal neuron extends sideways for some distance without forming any synapses with other neurons, and then this axon produces a dense cluster of endings.Like an express train, it skips intermediate stations.The distance from the cell body to the center of the peripheral plexus was approximately 0.43 mm in the monkey's primary visual cortex; 0.65 mm in the adjacent secondary visual area; 0.73 mm in the sensory area; and 0.85 mm in the motor cortex.For convenience, let me refer to them collectively as "0.5mm".This axon can continue to extend the same distance to form a terminal plexus.This pattern of children's "build a house" game can extend for several millimeters. Throughout the annals of cortical neuroanatomical research, this pattern of skipping spatial configurations is unique.We don't know much about its function, but it certainly makes you think that cortical areas spaced 0.5 mm apart might sometimes do the same thing, with patterns of activity that repeat, like the repeating patterns of wallpaper. You may have noticed that this jumpy spatial configuration is the same as the distance between the large columns is about 0.5mm.The small patches of color are separated from each other by about the same distance, but there are differences. The second superficial pyramidal neurons, 0.2 mm away from the first, also have axons themselves, but have different "stops" for express trains, and they still form terminal plexuses at 0.5 mm intervals, but each plexus is far from the first The terminal plexus of each neuron is 0.2 mm.When I was in college, the Chicago Transit Authority had such a train system of trains A and B. One train stopped at even-numbered stations, and the other stopped at odd-numbered stations. For the convenience of changing trains, several stations are common to the two trains. .Of course, any one subway station sometimes stretches over a block; likewise, our superficial pyramidal neurons are not co-located, with dendritic clusters that radiate laterally from the cell body, often stretching 0.2 mm or more. Try contrasting this with a large column.Up to this point, macrocolumns have been viewed as a territory with a common input source, as if you could build a fence around a group of microcolumns with a common address list.The color blobs have a common output target (the secondary cortical area specialized in color).We will not speak here of large columns with laterally directed excitatory axonal branches, although jump space configuration may be a cause (or consequence) of large columns at adjacent organizational levels.Imagine this: In a forest of jagged branches, each tree has a telephone wire leading out to touch a tree in the distance, not only bypassing the intervening tree, but also jumping over the tree that divides the forest have common input fences. Sideways "recurrent" connections are common in real neural networks.Case suppression was the subject of research by the 1961 Nobel Prize winner Geory von Fkekesy and the 1967 Nobel Prize winner H Keffer Hartline.It tends to sharpen the blurred boundaries of a spatial pattern (although they may compensate for blurry optics, with some side effects such as certain gauge illusions).But the surface pyramidal neurons were excitatory with respect to each other, suggesting that unless inhibitory neurons counteract this tendency, activity can spread among each other like a bush on fire.What's going on here?Is this why the cerebral cortex is so prone to seizures when inhibitory neurons fatigue? Furthermore, the standard hop spacing implies the possibility of round-trip transmission, the kind of reverberating loop that early neurophysiologists postulated.Two neurons separated by 0.5 millimeters can maintain the continuous activity of each other.Neurons have a dead zone called the refractory period after a firing, which lasts 1-2 milliseconds, during which time it is almost impossible to trigger another firing. The conduction time required for 0.5 mm is about 1 millisecond, and then synaptic detachment slows transmission by another half millisecond.Thus, if the connection between the two neurons is strong enough, you can imagine that the second neuron's impulse will recede to the first neuron at about the same time as it recovers to produce another impulse.However, the connections between neurons are usually not strong enough so that such rapid firing (even if initiated) cannot be sustained (but in the case of the heart, the connection strength between adjacent cells is indeed strong enough, when damage slows the conduction time cyclic repetitive excitation is an important pathological phenomenon). If the meaning of the cortex's standard jump spacing is not an impulse to chase its tail, then what does this mean?It seems that its significance is likely to lead to synchronization. If you sing in a chorus, you're getting in sync with everyone else by listening to others—if you're listening to your own voice, it's either too late or too early.Of course you also influence others.Even though it's a bit of a struggle to hear everyone sing, you're in sync with the others very quickly because of that feedback. Your position in that chorus is much like the position of a surface pyramidal neuron in the neocortex, which receives excitatory input from neighboring neurons around it.Networks like this have been studied extensively, even if not enough has been done on the superficial neocortex itself.This kind of network even with a small amount of feedback, syncs up (which is why I'm assuming you're a little hard of hearing, that's all) Two identical pendulums tend to sync up if they're close together, precisely because they The resulting air and shelf vibrations. Menstrual cycles are also said to tend to synchronize in female dormitories. While harmonic oscillators like pendulums take a little time to get in sync, non-linear systems like the pulses in neurons generated) can be quickly synchronized, even if the interconnection strength is relatively weak. Must this synchronizing tendency have something to do with the replication of spatiotemporal patterns?It's all a matter of simple geometry, the kind that the ancient Greeks discovered when they gazed at the tile mosaics of their bathroom floors, and that many of us rediscover in wallpaper patterns. Let's imagine a "banana committee" forming, various neurons scattered throughout the primary visual cortex, responding to this or that feature of the banana you see.The outline of the banana shape is a particularly effective stimulus for certain neurons that specialize in detecting boundaries and their orientation.There are yellow-loving neurons in the color patches. Given their tendency to excite each other (assuming a 0.5 mm jump distance between their axon terminal clumps), they would have a tendency to synchronize - not in that neuron (which I'll call "yellow 1") All impulses in Yellow 2 will be synchronized with those in Yellow 2, but a certain percentage will occur within a few milliseconds. Now imagine another superficial pyramidal neuron equidistant (0.5 mm) from "Yellow 1" and "Yellow 2".Maybe it only receives a faint yellow input, so its yellow discharge isn't active.But this "Yellow 3" also receives input from "Yellow 1" and "Yellow 2".Furthermore, certain inputs from the two synchronous neurons of "Yellow 1" and "Yellow 2" will reach the dendrite of "Yellow 3" together due to the same transmission distance.This is what hi-fi fans call "sitting at the hotspot", i.e. sitting at a point equidistant from the two speakers placed at the vertices of an equilateral triangle, moving slightly to one side, the stereo illusion is destroyed, and what you hear It is the monaural sound of the closer speaker.At hotspots in the cortex close to "yellow 3", the two synaptic inputs are additive, i.e. 2+2=4 (approximately).But since Yellow 3's impulse firing threshold might be 10, it remains quiet. It doesn't mean much.But these synapses are glutamate synapses in the superficial layer of the cortex, which have NMDA channels that allow sodium and potassium to enter downstream neurons.That itself isn't that important either. I have not yet told you why neurophysiologists find NMDA channels so fascinating compared to all other synaptic channels; they are not only sensitive to arriving glutamate, but also to pre-existing voltage across the postsynaptic membrane sensitive.If the voltage is raised, the next arriving glutamate will cause a larger effect, sometimes twice the standard amount.This is because under normal circumstances, there is a magnesium ion embedded in the center of many transmembrane NMDA channels, which acts as a plug, and the increase in voltage will flush out the plug, so that the sodium that was originally blocked will be released when glutamate opens the gate the next time. and calcium influx into dendrites. The consequence of all this is important: it means that impulses arriving simultaneously are more efficient than expected by 2+2, whose sum can be 6 or 8 (non-linear).Repeated approximate synchronization of the two inputs is even more effective because it happens to clear the magnesium plugs in each other's channels.Soon, repeated synchronized inputs from Yellow 1 and Yellow 2 could trigger an impulse in Yellow 3. The reciprocal reexcitation of the standard spacing and the increase in NMDA synapse strength work together seamlessly, both because of the synchronous tendency.Emerging features often emerge from this combination of seemingly unrelated events. Now we have 3 active neurons forming the three corners of an equilateral triangle.But there can also be a fourth one, located 0.5 mm away from the other side of "Yellow 1" and "Yellow 2".There is not much data on how many axonal branches a single superficial pyramidal neuron has, but a top-down view on a perfused superficial pyramidal neuron stained with a dye followed by a careful computer reconstruction shows that in many directions There are branches on it.Therefore, there must be a doughnut-like excitatory ring about 0.5 mm away from the neuron.Two such rings, the centers of which are 0.5 mm apart, are also 0.5 mm away from "Yellow 1" and "Yellow 2", and they have two points of intersection, just like in the exercise of bisecting lines in plane geometry. Therefore, it would not be surprising if "Yellow 1" and "Yellow 2" would recruit both "Yellow 3" and "Yellow 4" once they acted synergistically and synchronized with each other.There are other neurons on the equidistant "hotspot" of the pair of Yellow 1 and Yellow 3: perhaps a Yellow 5 will also join the chorus, if it has enough other inputs to make Pairs of input words that reach their threshold range.As you can see, there is a tendency to form a triangular array of often synchronized neurons that can extend several millimeters along the cortical surface. Because a neuron can be surrounded by 6 other neurons, all of which will fire at some point, we have error correction: even if a neuron wants to do something else, it will be fired. Forced back into the synergy established for its recalcitrant neighbour.At its core, this is an error-correcting procedure, just what artificial facsimiles need—if only the slender intercortical axon terminals did what the local terminals did: Small pieces spaced 0.5mm apart rather than terminating at a point. In the right hemisphere, they do fan out in slices. The problem raised by the idea of ​​a "pool" of associative memory is the need to keep a spatiotemporal code consistent over long distances within the cortex (eg, from left to right brain via the corpus callosum).Distortion of spatiotemporal patterns due to lack of fine topological mapping (axon terminals do not always fan out, do not terminate at a single point), or jerky in time (non-uniform conduction velocity), for unidirectional信息流也许并不重要，在那种情况下，只是在通路中一种任意的密码为另一种任意的密码所取代而已。 但是，因为在相隔较远的皮层区域之间的连接常常（7条通路中有6条）是交互性的，在前向传输中初始时空放电模式的任何畸变都需要在反向通路上得到补偿，从而使特有的时空模式始终能作为一种感觉或运动图式的局部密码。你可以用一种逆变换校正这种畸变，就像把一个压缩的卷宗松开；也可以用前述误差校正机制来作出修正。你也可以与木同的密码相安无事，只要它们在局部范围内具有相同的涵义（像真名和浑名）。这被称为退化密码，6个不同的DNA三联体均编码为亮氨酸便是一例。我以前曾想，哪一种都比一种误差校正模式更有可能，但当时我并没有意识到，从与回归性兴奋和同步敏感的NMDA通道必然相伴的结晶化所产生误差校正会有多么简单。 现在想象一下，有一个光导纤维阵列把一个皮层区与另一侧的相应区连接起来。真实的光纤把一幅象分解成许多点，然后忠实地将点作长距离传输；要是在光纤的那头观察，你会看到亮点组成的与前端相同的输入图案。 轴突并非光纤，因为在每一端都有许多细芽。真实的轴突并不终止于一点，而是展开于大型柱那么宽。真实的轴突束也不像一根相干的光纤束，与相邻的互不干扰，它们可以彼此交混起来从而使一个点走入叉路，其在另一端的终止位置发生偏移。真实的轴突的传导速度也会改变，一起开始的冲动可以在不同时间到达，使时空模式发生畸变。 但是局部误差校正特性提示，所有这些在一束皮层间纤维的远端都无甚干系。由于那些三角形阵列的存在，所送出的是一种有冗余度的时空模式。在远端的每一点所接收的不仅可以有来自信在目标上的轴突的输入，还叠加有来自离该点0.5毫米多达6根轴突的返回输入。是的，其中有些迷路了，有些到达迟了，但是接收神经元倾向于注意那些反复同步的输入，也许为了复制与起源点相同的放电模式，有几个输入就可以了，对“散兵游勇”则弃之不顾。 一旦时空模式的一个小区域在远端重又形成，就像我已解释过的那样，它能扩展为一个较大的区域。就这样，同步的三角形阵列使零乱的布线有可能在皮层内把时空模式作长距离传送——倘若在开始时的时空模式在空间上有10余个拷贝，而在远端终结时有足够范围的相同的模式。 一个阵列会变得多大？如果跳跃间距在边界处发生改变，这阵列可能局限在其原来的布劳德曼区。例如，在猴的初级视皮层，跳跃间距为0.43毫米；在其邻近的次级视区为065毫米，这就不大可能发生跨越界线的神经元的募集，但这是一个经验性问题，我们将必须对此进行考察。募集更多神经元进入该三角形阵列需要有对该香蕉已有一定兴趣的候选者。 因此，“黄色”的三角形阵列可能不比接收黄香蕉象的视皮层大很多。对线段朝向敏感的神经元平的可能也是同一回事：几个神经元进入同步，参加已定调的合唱，从而形成其中心在另处的一个0.5毫米的三角形阵列；对于香蕉的每一种独立地被察觉的特征，可能会存在一个不同的三角形阵列，它不一定在皮层上跨越相同的距离。俯视被展平的皮层，假设当一个冲动发放时一个微型柱会发亮，我们将会看到一群闪烁的光点。 如果我们把视野局限于05毫米的一个圆圈，我们不大会看到多少同步活动，有一个对“黄色”敏感的神经元每秒放电几次，另一个对“线”敏感的神经元每秒放电十几次，等等。但是如果把我们的视野增宽至几个毫米，那我们一忽儿看到几个点发亮，过一忽儿又是另一些点发亮。每一群发亮的点本身会形成一个三角形阵列。总起来看，各种阵列组成一个“香蕉委员会”。 请注意，在神经元的募集开始满员之前，原先的“黄色和线段委员会”的范围可能会大于外5毫米。即使原先的“委员会”散见于凡是米的范围内，三角形阵列也会尽力建立一个小得多的单元模式（这种模式在需要恢复时可能更易重建）。我们已经把密码压缩到比最初所占据的更小的空间之内，也复制了多余的拷贝。这有一些有趣的涵义。 这是一种与香蕉的表象有关的时空模式，但是，它是香蕉的皮层密码吗？我将把这种不忽略任何重要信息的最小的模式称为基本模式，“线段”、“黄色”三角形阵列能通过这些模式得以重建。 如果我们缩小视野来看闪烁的微型柱，那么在什么范围内我们便不再能找到同步化的微型柱呢？是的，大约是0.5毫米，但不是0.5毫米的圆圈，而是，个其平行面间距为0.5毫米的六角形。这是一个简单的几何学问题：六角形瓷砖的相应点（如有上角）形成三角形阵列。任何大于该六角形的将开始把一些多余的点包括进来，而这些点已经由其三角形阵列的另一些点来表示。因此在我们局限的视野中有时会看到两个同步的点。 基本模式通常不会充满该六角形（我想象在此六角形中，百多个微型柱中有十余个是活动的，但是其余的必定要保持安静，否则会使图象模糊）。我们无法看到边界被勾划出来，以致当在复制一片区域过程中俯视皮层表面时，我们不会看到一种蜂窝状结构。确实，当墙纸设计者构建一种重复图案时，他们经常要注意使图案单元的边界不易看出，从而使它总体上看起来天衣无缝、虽然是三角形阵列进行募集和建立密集的图案，但看上去好像六角形在不断地被复制着”。 这种三角形同步活动不一定持续很久，它是组构的一种短生形式，可能在伴有皮层兴奋性降低的EEG（脑电）节律的某个相位被擦洗掉。如果我们想要重建一种已经消失的时空模式，我们能从两个相邻的六角形小片开始。当然，可以从扩展的香蕉形镶嵌原先覆盖的任何两个相邻的六角形开始，它不必一定是原先的那一对。记忆痕迹——对重新唤起该时空模式至关紧要的“印记”——可能只有两个相邻六角形中的回路那么小。这种极小模式的重复复制可能控制一个区域，就像是一块晶体生长起来，或者墙纸重复一种基本图案一样。如果这种“旋律”在其终止之前“重奏”足够多次，LTP有可能以某种方式滞留下去，使那种时空模式易于在这个或那个位置重新产生。 如果这种空间模式较稀疏，几个大脑密码（如“苹果”和“柑橘”的密码）能够重叠起来使你形成一个范畴（如“水果”）。如果你试图把点矩阵打印机打印的几个字母重叠起来，你所得到的是一个墨团。但是，如果矩阵点子稀疏，你有可能把一个一个字母复原起来，因为它们每一个都产生十分清晰可辨的时空模式。因此，这类密码也能方便地用来形成能分解成若干单元的各种范畴，正像叠加的旋律常常能单个听出来一样。由于这种远程复制的特点，你能形成多模范畴，如“梳子”的所有内涵。 我的朋友唐·迈克尔（Don Michael）认为，默念可能相应于通过诗文一建立一种无意义的密码的镶嵌，这种密码并无明显的共鸣或关联。如果你维持默念足够久，从而把烦恼和执念洗擦干净，让那些短时程的印记消退，它可能给予你一个新的起点来走近长期的记忆印记，而不再系于短期的兴趣之上。 （默念的）沉浸于自身的无忧状态近乎完美，但遗憾的是并不长久。它易为内心所放动。犹如无根之木，情绪、感觉、渴念、烦化，甚至思想，都是以一种无意义杂乱的方式油然而生的，无法自制。它们越是牵强，越是荒谬，它们与人的意识集中的关系也就越少，它们也越顽强地挥之不去……使这种扰动恫效的唯一途径是保持安静和漠然的呼吸状态，沉浸于与环境中出现的任何东西的友善的关系之中，习惯于它，平静地看它，最终生倦，不再看它。 通过对表层锥体神经元的这种分析所产生的想法有一些吸引人的特征。已故的赫布会垂青于此，因为这显示了短期和长期记忆的某些最使人困惑的特征，可以怎样用细胞集群来加以解释。这些特征包括；记忆痕迹是以分布的方式存贮的，并没有一个位点对于它的复苏是关键的等等。格式塔心理学家”也会喜欢这种分析方式，因为这样就有可能借助于会超出物体界线的三角形阵列来对图形和背景加以比较，而形成这种比较的时空模式所代表的并不是单独的图形或背景，而是两者的综合。 精神活动包含有多侧面环境所影响的复制竞争，我想，达尔文和詹姆斯会欣赏这种精神活动展示的前景。西格蒙特·弗洛伊德（Sigmund Freud）”可能会被下意识的联想如何不时突现在意识的前景之中的机制所吸引。 虽然我认为发散式思维是新皮层的达尔文机最重要的应用，但让我先来解释一下它可以怎样应用于收敛式思维问题。假设有某种东西喷地一下子从你身旁穿过，并立即消失在椅子下。你猜想它是圆的，可能是橙色或黄色的，但是它运动的速度很快，已经超出你的视界，你不可能再看第二眼。what is that?如果答案不是显而易见的，那你作怎样的猜测？你首先需要列出几种可能性，然后加以比较，看哪一种可能性更大。 幸好复制竞争能够做到这一点。对于那个物体假定有一种大脑密码，它是由所有被激活的特征检测器所形成的：颜色、形状、运动，可能还有碰击地板时发出的声音。不妨说，这种时空模式开始召募其同类（见图65）。 它是否得在其毗邻处建立它的翻版取决于毗邻处是否发生共鸣，这种共鸣的基础是由毗邻皮层的突触强度的模式以及其他可能的活动状态所确定的。如果你在以前已多次看到过这样的物体，那么可能会有完全的共鸣，但是你并没有。不过，假定的大脑密码有“圆”、“黄”、“快”等成分，网球就有这些属性，你由此引起了共鸣，两相邻的皮层区也和着“网球”的旋律（混炖吸引子的一个很起作用的特征是，它能抓住附近相吻合的东西，将之转换为特有的模式）。如果共鸣不佳，就会丢失某些成分，因此，也许你的“柑橘”共鸣在皮层的另一区俘获一个不同的模式，尽管颜色并不完全对。 复制竞争又怎么样呢？在这里已经谈到了我们有“未知”、“网球”、“柑橘'等大脑密码的翻版。也许“苹果”也会突然蹦出来：如唱歌人在几分钟之前看到一个人在吃苹果，通过为那种模式增强的NMDA突触形成“苹果”这样暂时的印记。但是，“苹果” 模式即为“柑橘”模式所超越。在“未知”模式眼下占据领地的另一侧，“网球”模式正干得不错，最终征服并取代厂'未知'很式，甚至侵入了“柑橘”的领地。正是在这个时候，你会说：“我想我看到的是一只网球。”这是因为在“网球”这一“合唱”中，最终已经有了足够的“和声”，从而把一种连贯一致的信息经过皮层间通路从枕叶传送到领叶，再传送到你的左侧的语言皮层。 现在又有什么发生了：一种新的时空模式开始在工作空间复制拷贝；这一回你看到了很熟识的东西（椅子），很快就建立起一种关于“椅子”的有决定性意义的“合唱”，并没有任何真正的竞争，因为那种感觉时空模式赶在任何其他模式之前已立即激起了共鸣。然而，在“网球”和“柑橘”模式中所使用的NMDA突触仍然相当活泼，在之后5分钟左右的时间里，在它们原来所占据的那部分工作空间要重建其中任一时空模式将比通常更加容易。也许，“柑橘”不断复制拷贝，错误地激起“橙色水果”的共鸣，以致一分钟后，你会怀疑你对干网球的判断是否错了。 那就是复制竞争可能是如何发生的，以及我是如何想象我们的卞意识过程有时会晚半小时后才想起的缘由。模式共鸣有点像我们想象在脊髓中“运动”是怎样的一种过程：在各种神经元之间存在着不同的突触连接强度，在一定的初始条件下，你能突然发生与实施“行走”的时空模式的共鸣。当初始条件不同时，你则可能与其他的吸引子发生共鸣，如“慢跑”、“跨大步”、“跑步”或“造房子（式跳跃运动）”等。 在感觉皮层中，你可能突然闯入“橙”或“柑橘”的范畴，即使你所看到的水果并不是“橙”或一柑橘”。如我在第四章中已谈到的，那就是为什么对英语L和R两个发育日本人会有如此麻烦，因为他们对一个特定的日语音素的思维范畴会俘获这两个声音。现实很快为思维模式所取代，正如亨利·梭罗（Henry David Thoreau）所说：“我们仅仅听到和理解我们已经知道一半的东西。 " 皮层能够很快学会新的模式，不管是感觉的还是运动的，也能使之产生变异。这些变异使竞争成为可能，它决定什么模式能最佳地与连接特性发生共鸣，而皮层的连接特性常常为许多感觉输入和情感上的驱使所影响。 关系也能用时空模式来编码，就如感觉或运动的图式能被编码一样好。把密码组合起来产生一种新的任意模式，就如左手的节律能被叠加在右手的旋律上一样。 第四章的语言机提供了某些特例，以说明在一个句子中可以包涵多么精巧的关系：所有强制和非强制性的角色。一个动词（如“给”）的强制性语义是与关系相关的，当一种强制性角色末被充填时便引起认知上的别扭。嘿，这正如广告代理商已经发现的，“给他”这样的广告迫使你再去读广告牌，以发现你所疏漏的，你因此把该广告记得更牢”。 那么，一个句子是否就是一种在与其他的句子密码的竞争中复制出来的大型时空模式呢？那并不一定。为了作出一个决定，我们并不需要复制竞争，如果没有包含特别新的东西，简单的评估系统应该是足够了。请回忆一下第一章中关于鸳鸯的论述：评估系统在其作决定时将起作用，因为这些选择游水、潜水、晾干翅膀、飞走、再看一会儿周围等，在世代进化中已经定型了。一旦你对其标准的涵义有很深的理解之后，你会发现可复制图式并不是包揽一切的。 许多灵长类在其皮层的浅表层中有标准跳跃间距的接线，这种接线预示存在短生的三角形陈列。人们不知道任何动物有多么经常地用它来复制墙纸似的六角形图案；也许它仅仅短暂地发生在出生前的发育过程中，作为一件测试模式来引导那些依赖于使用的连接，之后再也不发生。也许，皮层的某些区域完全用于实施专门化功能，决不复制短生的模式，而另一些区域常常支持旁路复制，变成为以达尔文成型过程所用的可擦拭的工作空间。鉴于运动指令的拷贝对投掷动作特别有用（因为这些指令能减少定时上的颤抖），因而也许在人类投掷准确性的进化过程中存在某种自然选择，从而使之具有较大的工作空间。所有这些都是经验性问题；一旦我们的记录技术的分辨能力得到改进，我们将一定能看到六角形拷贝位于这一系列可能性的某一位置。 但是，要满足达尔文机的必要条件就需要某种与这类复制竞争很相近的东西——那就是我为什么引导读者穿过这个大脑迷津的真实理由。在这里，我们至少有了一个清晰的模式；有了复制、变异；有了为工作空间发生的可能的竞争；有了影响竞争的多侧面环境（现时的和记忆中的），以及有了下一代更可能具有由最大领地的拷贝所建立的模式异体（大的领地具有更长的周界线，正是这些界线上模式异体能摆脱误差校正倾向，并开始复制新的模式）。 在一本篇幅更长的关于新皮层达尔大机的书《大脑的密码》中，我将解释所有你从性、孤岛和气候变化的大脑同源物中所得到的趣味和速度。如果大脑中的达尔文过程快到能向我们提供进行正确猜测的智力，那么速度正是我们所需要的。 我们一直试图把大脑皮层分解为一些专门化的“专家”模块。对于探索专门化功能，这是一种上佳的研究策略。但是我并不很认真地把它看作是有关联合皮层是如何工作的一种概述。我们需要某些可擦拭的工作空间，需要能募集帮手来实施困难的作业。这提示，任何专家模块也应是通才，就像在紧急情况下，一位神经外科大夫也能取代家庭医生一样。我之所以偏爱短生性六角形镶嵌的理由之一，是它对专家一通才作谬提出了一种解答，即甚至一片具有专门性长期印记的皮层区，也能用作工作空间，以覆盖其上的短期印记来影响竞争。 这样一种镶嵌也提示了一种下意识思想可能进行的方式，它有时会把往昔的某种相关事实推入你的意识之流。特别重要的是，因为模式的异体本身能进行复制达到短时的成功，因此这条“拼花被褥”是有创造性的——一它能将一些不起眼的原始素材塑造成某种像样的东西。甚至更高形式的关系，如隐喻，似乎也有可能产生，这是因为大脑密码是任意的，能够形成新的组合。Who knows?也许现在你甚至已经习得了埃科的关于苹果PC计算机类比的大脑密码。 同步化三角形阵列对达尔文复制竞争有一些使人感兴趣的意义，这些阵列对复杂的语言也有其意义，这有可能从另一方向有力地推进智力的发展。 从原始语到完全成熟的有句法的语言之间存在相当大的飞跃，语言学研究者们怀疑两者之间存在着某种中间形式。原始语即使有丰富的词汇，但只有很少的结构，它主要是依赖于在几个词之间的简单的语境的关联来传递信息。结构的加入则大不相同了。 循环性嵌入结构（如句中句：I think i saw him leave to go home．我想我看到了他离开回家了）的脑机制被认为对于“通用语法”是至关重要的。语言学家还需要的是相隔较远的两个词之间的依从关系，包括代词和它所指代的对象间的关系。这种结合需要比局部范围更大的联系；此外，循环性嵌入结构需要构建这些联系的等级阶梯。在我们知道“梳子”的视觉涵意存贮在视皮层附近，其听觉涵义存贮在听皮层附近等等的情况下，大脑皮层不相邻的区域可能参与许多尝试性的关联之中。 皮层间轴突束要比非相干光纤束精得多，它们不存在邻居关系。随着每根轴突终未的分支展开，可能会失去点对点的映射，这有点像手电筒光束的辐射。尽管存在杂乱和污迹所致的非相干性，某些发生畸变的图形，通过经验还是可能在远端辨认出来，此时所采用的是与范畴性感知相似的类簇状分析（cluster-analysis）机制”。这必然使那些具有良好实践的特殊情景的传送成为可能，这相似于海员使用的信号旗——也许一次只能发几种旗语，因而限制了在皮层区之间可能传送的新的关联。嵌入结构可能局限于常用的句子。这种皮层间非相干性的能力一定能胜任原始语的处理。 但是，误差校正机制提供了将任意时空模式沿皮层间轴突束传送并一次成功的可能性，因此传送便不再限于某些特殊情况下的图形，这些图形虽在空间和时间上发生“畸变”，但已为目标皮层识别为有意义的信息。这种皮层间相干性意味着新的关联是可能得到传送的；目标皮层能以相似的误差校正把它送回，让它在起源皮层中被自动识别，而毋需对一种发生两次畸变的图形进行调整，然后构建与原始时空放电模式等价的模式。 采用相同密码的返回性投射意味着，你能有一种分布式的和声，远处的合唱队员以此使群体保持在临界大小之上。返回投射的歌声并不需要有充分的特征来帮助整个合唱，它可能更像那种跟唱技术，即一个人单调地教唱一句，听众在音乐上加以揣摩，重复地跟唱。返回投射也提供了能分辨模糊不清的检查跟踪系统。如有了能保持句子结构的关联，嵌入结构就有可能成立，即不再存在这样的危险： “the tall blond man with one black shoe”（穿着一只黑鞋的高个子金发男人几个词混合的思维模式被打乱为“a blond blackman with one tall shoe”（穿着一只高高的鞋的金发黑人）。 因此，皮层间的精细性本身是从原始语言向真正语言飞跃的一个候选者（虽然你仍然需要语义结构层次上的许多小规则）。诚然，向任意密码传送的转换可能同时使用“通用语法”的两个主要创新点——嵌入结构和远程联系。这样，我们现在有了几个候选对象，即达尔文机和相干的皮层间投射等，它们可能已经推进了智力和语言的发展，使不经常作出创新的直立人文化，在约25万年前进化为人类不断变化的文化。 在我们所有的研究终结时，我们必须再一次试图把人的灵魂视作灵魂，而不是一群营营作响的生物电信号；人有所欲，而不只是激素的涌动；人的心脏并非是一种纤维性粘滞的泵浦，而是隐喻的知性的器官。我们并不需要把它们视作是超自然的实体，它们是活生生的，有血有肉的。但是，我们必须相信它”1确是实体，不是被分解的断片，而是完整的。之所以完整，是因为我们通过对它们的思索；通过我们在谈论它们时所用的词语；通过我们把它们转化为言语的方式；已经使之成为真实。即使它们已在我们的眼前被剖析，但我们还是对它们的无懈可击而敬畏不已。