2 An alternative account of moral skill

We have only recently begun to understand what that “something else” is. It has to do with the peculiar way the brain is wired up at the level of its many billions of neurons. It also has to do with the very different style of representation and computation that this peculiar pattern of connectivity makes possible. The basics are quite easily grasped, so without further ado, let us place them before you.

The first difference between a conventional digital computer and a biological brain is the way in which the brain represents the fleeting states of the world around it. The retinal surface at the back of your eye, for example, represents the scene currently before you with a pattern of simultaneous (repeat: simultaneous) activation or excitation levels across the entire population of rod- and cone-cells spread across that light-sensitive surface. Notice that this style of representation is entirely familiar to you. You confront an example of it every time you watch television. Your TV screen represents your nightly news anchor’s face, for example, by a specific pattern of brightness levels (“activation” levels) across the entire population of tiny pixels that make up the screen. Those pixels are always there. (Tiptoe up to the screen and take a closer look.) What changes from image to image is the pattern of brightness levels that those unmoving pixels collectively assume. Change the pattern and you change the image.

It is the same story with any specialized population of neurons, such as the retina in the eye, the visual cortex at the back of the brain, the cochlea of the inner ear, the auditory cortex, the olfactory cortex, the somatosensory cortex, and so on and so on. All of these neuronal areas, and many others, are specialized for the representation of some aspect or other of the reality around us: sights, sounds, odors, tactile and motor events, even features of social reality, such as facial expressions. These neuronal activation-patterns need not be literal pictures of reality, as they happen to be in the special case of the eye’s retinal neurons. But they are representations of the fine-grained structure of some aspect of reality even so, for each activation-pattern contains an enormous amount of information about the external feature of reality that, via the senses and internal brain pathways, ultimately produced it.

Just how much information is worth noting. The retina contains roughly 100 million light-sensitive rods and cones. (In modern electronic camera-speak, it has a rating of one hundred megapixels. In other words, your humble retina still has ten times the resolution of the best available commercial cameras.) Compare this to the paltry representational power of a typical computer’s CPU: it might represent at most a mere 8 bits at a time, if it is an old model, but more likely 16 bits or 32 bits for a current machine, or perhaps 64 or 128 bits for a really high-end machine. Pitiful! Even an old-fashioned TV screen simultaneously activates about 200,000 pixels, and an HDTV will have over two-thirds of a million (1,080 × 640 = 691,200 pixels). Much better. But the retina, and any other specialized population of neurons tucked away somewhere in the brain, will have roughly 100 million simultaneously activated pixels. Downright excellent. Moreover, these pixels—the individual neurons themselves—are not limited to being either on or off (i.e., to displaying a one or a zero), as with the elements in a computer’s CPU. Biological pixels can display a smooth variety of different excitation levels between the extremes of 0 percent and 100 percent activation. This smooth variation (as opposed to the discrete on/off coding of a computer’s bit-register) increases the information-carrying capacity of the overall population dramatically. In all, the representational technique deployed in biological brains—called population coding because it uses the entire population of neurons simultaneously—is an extraordinarily effective technique.

The brain’s computational technique, which dovetails sweetly with its representational style, is even more impressive. (As with any computer, a computational operation in the brain consists in its transforming some input representation into some output representation.) Recall that any given representation within the brain typically involves many millions of elements. This poses a prima facie problem, namely, how to deal, swiftly, with so many elements. Fortunately, what the brain cannot spread out over time—as we noted above, it is far too slow to use that strategy—it spreads out over space. It performs its distinct elementary computations, many trillions of them, each one at a distinct micro-place in the brain, but all of them at the same time. Let us explain with a picture so you can see the point at a glance.

At the bottom of figure 1 is a cartoon population of many neurons—retinal neurons, let us suppose. As you can see, they are currently representing a human face, evidently a happy one. But if the rest of the brain is to recognize the specific emotional state implicit in that sensory image, it must process the information therein contained so as to activate a specific pattern within the secondary patch of neurons just above it. That second population, let us further suppose, has the proprietary job of representing any one of a range of possible emotions, such as happiness, sadness, anger, fear, boredom, and so forth. The system achieves this aim by sending the entire retinal activation-pattern upward via a large number of signal-carrying axonal fibers, each one of which branches at its upper end to make fully eighty synaptic connections with the neurons at this second layer. (Only some of these axonal fibers are here displayed, so as to avoid an impenetrable clutter in the diagram. But every retinal neuron sends an axon upward.)

Figure 1

When the original retinal activation-pattern reaches the second layer of emotion-coding neurons, you can see that it is forced to go through the intervening filter of (4,096 axons × 80 end-branches each = 327,680) almost a third of a million synapses, all at the same time. Each synaptic connection magnifies, or muffles, its own tiny part of the incoming retinal pattern, so as collectively to stimulate a new activation-pattern across the second layer of neurons. That new pattern is a representation of a specific emotion, in this case, happiness. The third and final layer of this neural network has the job of discriminating these new 80-element patterns, one from another, so as to activate a single cell that codes specifically for the emotion still opaquely represented at the second population. That is achieved by tuning a further population of synaptic connections from every cell in the middle layer to each of the five cells in the final layer. In all, what was only implicit in the original retinal activation-pattern (mostly in the mouth and eyebrows) is now represented explicitly in the top-most activation-pattern across the five cells there located.

This trick is swiftly turned by the special configuration of the various strengths of each of the intervening synaptic connections. Some of them are very large and have a major impact in exciting the upper-level neuron to which it is attached, even for a fairly weak signal arriving from its retinal cell. Other connections are quite small and have very little excitatory impact on the receiving cell, even if the arriving retinal signal is fairly strong. Collectively, those 327,680 synaptic connections have been carefully adjusted or tuned, by prior learning, to be maximally and selectively sensitive to just those aspects of any face image that convey information about the five emotions mentioned earlier, and to be “blind” to anything else. The complex “pattern transformation” they effect plainly loses an awful lot of information contained in the original (retinal) representation. Indeed, it loses most of it. But it does succeed in making explicit the specifically emotional information that this little three-layer “neural network” was designed to detect.

This style of computation is called Parallel Distributed Processing (PDP), and it is your brain’s principal mode of doing business on any topic. Even in this cartoon example, you can see some of the dramatic advantages it has over the “serial” processing used in a digital computer. A typical 8-bit CPU has a population of only eight representational cells at work at any given instant, compared to fully 4,096 just for the sensory layer of our little cartoon neural network. The CPU performs only eight elementary transformations at a time, as opposed to 327,680 for the neural network, one for each of its 327,680 synaptic connections. When we consider the human brain as a whole instead of the tiny cartoon network above, we are looking at a system that contains roughly a thousand distinct neuronal populations of the same size as the human retina, all of them interconnected in the same fashion as in the cartoon. This gives us (1,000 specialized populations × 100 million neurons per population = ) a total of 100 billion neurons in the brain as a whole. As well, the total number of synaptic connections there reaches more than 100 trillion, each one of which can perform its proprietary magnification or minification of its arriving axonal message at the very same time as every other. Accordingly, the brain doesn’t have to do these elementary computations in laborious temporal sequence in the fashion of a digital computer. As we saw, a PDP network is capable of pulling out subtle and sophisticated information from a gigantic sensory representation all in one fell swoop. That is the take-home lesson of our cartoon network. The digital/serial CPU is doomed to be a comparative dunce in that regard, however artful may be the rules that make up its computer program. They simply take too long to apply.

Enough of the numbers. What wants remembering in what follows is the holistic character of the brain’s representational and computational activities, a high-volume character that allows the brain to make penetrating interpretations of highly complex sensory situations in the twinkling of an eye. You are of course intimately familiar with this style of cognition: you use it all the time. Every time you recognize frustration in someone’s face, evasion in someone’s voice, hostility in someone’s gesture, sympathy in someone’s expression, or uncertainty in someone’s reply, a larger version of the neuronal mechanism in figure 1 has made that subtle information almost instantly available to you.

Now, however, you know how: massively parallel processing in a massively parallel neural network. Or, to put it more cautiously, almost three decades of exploring the computational properties of artificial neural networks, and almost three decades of experimenting on the activities of biological neural networks have left us with the hypothesis on display above as the best hypothesis currently available for how the brain both represents and processes information about the world. No doubt, the special network processor inside you, the one that is responsible for filtering out specifically emotional information, has more than the mere two layers depicted in our cartoon. In fact, anatomical data suggests that your version of that network has the retinal information climb through four or five distinct neuronal layers before reaching the relevant layer(s), deep in the brain, that explicitly registers the emotional information at issue. The original retinal information will thus have to go through four or five distinct layers of synaptic filters/transformers before the emotional information is successfully isolated and identified. But that still gives us the capacity for recognizing emotions in less than a few tenths of a second. (The several neuronal layers involved are only ten milliseconds apart.) On matters like this, we are fast, at least when our myriad synaptic connections have been appropriately tuned up.

The PDP hypothesis also gives us the best available account of how that synaptic tuning takes place, that is, of how the brain learns. Specifically, the size or “weight” of the brain’s many transforming synapses changes over time in response to the external patterns that it repeatedly encounters in experience. The overall configuration of those synaptic connections and their adjustable weights is gradually shaped by the recurring themes, properties, structures, behaviors, dilemmas, and rewards that the world throws at them. The resulting configuration of synaptic weights is thus made selectively sensitive to—one might indeed say tuned to—the important features of the typical environment in which the creature lives. In our case, that environment includes other people, and the pre-existing structure of mutual interaction and social commerce—the moral order—in which they live. Learning the general structure of that pre-existing social space, learning to recognize the current position of oneself and others within it, and learning to navigate that abstract space without personal or social disasters are among the most important things a normal human will ever learn.

It takes time, of course. An infant, before his first birthday, can distinguish between sadness and happiness, but little else. A grade-school child can pick up on most of the more subtle emotional flavors listed three paragraphs ago, though probably only in the behavior of young children like themselves. But a normal adult can detect all of those flavors, and more, quickly and reliably, as displayed by almost any person she may encounter. (Only psychopaths defeat us, and that’s because they have deviant or truncated emotional profiles.)

Withal, learning to read emotions is only a part of the perceptual and interpretational skills that normal humans acquire. People also learn to pick up on people’s background desires and their current practical purposes. We learn to divine people’s background beliefs and the current palette of factual information that is (or isn’t) available to them. We learn to recognize who is bright and who is dull, who is kind and who is mean, and who has real social skills and who is a fumbling jerk. Finally, we learn to do things. We learn how to win the trust of others, and how to maintain it through thick and thin. We learn how to engage in cooperative endeavors and to do what others rightfully expect of us. We learn to see social trouble coming and to head it off artfully. And we learn to apologize for and to recover from our own inevitable social mistakes.

These skills of moral cognitive output (i.e., our moral behavior) are embodied in the same sorts of many-layered neural networks that sustain moral cognitive input (i.e., our moral perception). The diverse cognitive interpretations produced by our capacity for moral perceptions are swiftly and smoothly transformed—again by a sequence of well-trained synaptic filters/transformers—into patterns of excitation across our motor neurons (which project to and activate the body’s muscles) and thereby into overt social behaviors, behaviors that are appropriate in light of the moral interpretations that produced them. Or at least, they will be appropriate if our moral education has been effective.

This weave of perceptual, cognitive, and executive skills is all rooted in, and managed by, the intricately tuned synaptic connections that intervene between hundreds of distinct neuronal populations, each of which has the job of representing some proprietary aspect of human psychological and social reality. That precious and hard-won configuration of synaptic weights literally constitutes the social and moral wisdom that one has managed to acquire. It embodies the unique profile of one’s moral character: it dictates how we see the social world around us, and it dictates our every move within it. It is not an exaggeration to say that it dictates who we are. If our character needs changing or correcting, it is our myriad synapses that need to be reconfigured, at least in minor and perhaps in major ways. In all of these matters, then, don’t think rules. Think information-transforming configurations of synaptic weights, for it is they that are doing the real work.