Darwin, Skinner, Turing and the Mind

So it is not because of their intellectual complexity that Darwinian ideas are important; nor is it because of their scope. For whereas Newton's and Einstein's ideas apply to the entire universe, so far, Darwin's only apply to a small portion of our own tiny planet. But as "importance" is a human criterion, and humans are all part of that small portion of the planet to which Darwin applies, Darwin's ideas are surely as important (to us) as Newton's and Einstein's are. It is for this reason that although the Darwinian picture was already there, in broad strokes, in the 19th century, its ramifications are still being unravelled in the 21st.

The resistance to Darwin was not an intellectual resistance. It was not that Darwinism was hard to understand. It was that some of us had reason to want it not to be true. The main resistance was religious: Religions typically had their own ideas about the origin of humankind, but that was not the real bone of contention. The real bone of contention was the mind, or rather the soul, which was supposed to be an immaterial something that was outside the sphere of the completely materialistic Darwinian forces that shape survival/reproduction machines.

We will not discuss this tension between religion and Darwinism here except to note that religion really did not have that much to worry about as regards the capacity of Darwinism to give a fuller or better account of the soul. As we shall see, Darwinism turns out to be as powerless to explain the existence and purpose and origin of mental states as any other physical theory is, including Newton's and Einstein's.

But before we get to what Darwinism cannot explain, we need to get a fair sense of what it can explain. There is little dispute that Darwinism is the best and fullest explanation of the existence, purpose and origin of biological structure: Body size and shape, wings, fins, feet -- all of these can be fully understood as having been shaped by the way they successfully improved survival and reproduction in certain creatures, evolving in certain environments under certain prior conditions.

Along with its successful explanation of biological structure, Darwinism also explained biological function; the two are inseparable: Wings, fins and feet help survival and reproduction because of the way they are used -- to fly, to swim, to walk. Organisms have brains, which are also biologically evolved structures, and those brains generate their capacity to fly, swim and walk. Some of this behavioral function is inborn, hard-wired in the structure of the brain, like the reflex. But in general it is not economical or even possible for evolution to second-guess in advance every movement an organism may need to make in its lifetime. So behavioral function, unlike body structure, relies on the environment of an organism to help to shape it through learning, wherever possible. (Actually, even the shaping of body structure tends to be off-loaded onto the environment as much as is feasible in development, through "epigenetic" processes.)

So-called "Baldwinian" evolution is a process whereby behavioral capacities, instead of being fully pre-coded in the genome the way anatomical characteristics are, are instead left to be learnt from real-time, real-life experience. The organism inherits only the capacity and the tendency to learn to do certain things early and quickly. Imprinting in ducks is a good example: A duckling is not born with the hard-wired tendency to recognize and follow its mother. It is born only with the tendency to follow the first moving shape it sees, and to learn that that is the shape of its mother and of its species, so that it not only follows her while a duckling, but attempts to mate with similar shapes when it is an adult.

Baldwinian evolution -- the shaping of a tendency to learn certain specific kinds of things -- is a special "prepared" kind of learning, usually specific to certain contexts and objects, and often strongest in a "critical period" in development. The human propensity to learn language is also thought to be an example of this kind of prepared learning. Maybe all learning of which a species is capable is "prepared" to some degree. The "tabula rasa" on which anything and everything can be written is perhaps only a philosopher's fantasy. Yet human beings do seem to have some generic, nonspecific, all-purpose learning capacity, and other species seem to have them to some degree . To this extent, Skinner enters the picture, alongside Darwin.

Skinner always likened the kind of operant learning with which his name has become associated to evolution. As in Darwinian evolution, in which there is ("natural") selection among random genetic variations on the basis of their adaptive "consequences" -- that is, success in surviving and reproducing -- in Skinnerian learning there is selection among random behavioral variations on the basis of their rewarding "consequences": that is, success in attaining some "reinforcement." The organism's brain is prepared by Darwinian evolution in such a way as to make what is adaptive for the organism feel rewarding.

This is also the meeting point between the two "mechanisms" of Darwinian psychology, the "distal" or "ultimate" mechanism, which was shaped by whatever helped an organism's ancestors survive and reproduce in the environment in which they evolved, and the "proximal" mechanism, which shapes what they do during their lifetimes in particular environments today. So what makes me inclined to follow whatever moving object I see if I am a duckling, or to imitate whatever words I hear if I am a child, what makes sugar taste good to me, and quinine taste bad, was shaped by distal evolution. But what has taught me to follow my mother, if I am a duckling, or to speak the language of my mother if I am a human, what makes me ready to learn to work for food and obey laws to avoid fines and imprisonment, is shaped by proximal learning and experience. Distal mechanisms operate in a genotype's evolutionary time and proximal mechanisms operate in a phenotype's lifetime.

There are only two problems here. The first is that whereas Darwinian evolution (with the help of Mendel, and Watson, and Crick) has more or less identified its distal mechanism, the proximal mechanism of Skinnerian learning is still not known. Never mind. There are candidates, including neural nets, statistical and machine learning algorithms, and even genetic algorithms borrowed from the distal mechanism.

The second problem, however, is that we had said Darwinian evolution first makes certain things feel rewarding (because they have positive adaptive consequences); then they can go on to serve as the proximal reinforcers in Skinnerian learning. But, strictly speaking, Darwinian evolution can only explain why we would evolve a propensity to seek to eat sugar -- simplifying, it is in order to keep our blood glucose levels up -- but not why that should  feel (i.e., taste) rewarding. In fact, Darwin cannot begin to tell us why anything should feel or taste like anything at all.

Here is another way to put it: For Skinnerian reinforcement to work, it is not necessary that there should be feeling organisms. A reward is simply something that makes a system inclined to do again whatever it was that led to the reward. One can design a machine that is capable of Skinnerian learning. This is the point where Turing joins Darwin and Skinner in explaining the mind.

Turing did not approach the mind from a biological standpoint, but from a technological one: How can we make machines that are intelligent, machines that can think? This naturally raises the question of what intelligence and thinking are. And Turing proposed a very natural test, which has since come to be known as the "Turing Test." The test has two requirements. Let us call the system about which we are asking whether or not it is intelligent "the candidate." The candidate must, first of all, be able to do whatever an intelligent, thinking system like us can do. If it cannot, then it fails. But even if is passes, there is a second requirement: It must be able to do everything completely indistinguishably from the way we do it, so indistinguishably that we are unable to tell the difference.

This second requirement has come to be called "Turing Indistinguishability," and what it really means is behavioral indistinguishability. Turing's insight was that although we certainly cannot "define" in advance what it means to be intelligent, or to think, or to have a mind, we are very good -- because of our social "mind-reading" capacities -- at inferring the mental states of others, and we of course do this on the basis of their behavior: There is no real telepathy involved. We can easily tell that a stone or a river or a mountain or an automobile or a tape-recorder does not have a mind -- that it is not thinking, not behaving intelligently, as we do.

Notice that among the various candidates that I have named, about whom we are fairly certain that they do not have minds, I did not name any living creatures. Although the only candidates about which our mind-reading abilities confirm with full conviction that they have minds are the members of our own species, we are almost as convinced by the behavior of our nearest relatives; and in general we feel intuitively quite confident about all of our fellow-mammals and even our fellow-vertebrates. Serious uncertainty only sets in with the lower invertebrates, microorganisms and plants.

But Turing was not concerned with living intelligence; he was concerned with artificial intelligence: How can we know that a man-made candidate is thinking? If Turing had been a philosopher, he would have stated outright that the only way to know for sure is to be the candidate. And since the only candidate one can be is oneself, that is the only one we can be certain has a mind (as Descartes reminded us). But (as Descartes also reminded us), certainty is only to be had in mathematics. So in everything else we have to settle for something less than certainty. What is that criterion in the case of intelligence? Turing suggested that it should be exactly the same criterion we use with one another:

We infer that other people have minds from their behavior, from what they do. We are also influenced by their appearance, but we know that that is not critical. So, to exclude any bias from appearances, the original Turing Test was meant to be based only on language. Now language is a behavior, to be sure, and a very general and powerful behavior, but it is not our only behavior, nor perhaps our primary one. So Turing is probably right that if we could ever design a candidate that was indistinguishable from us in its verbal capacity -- one that we could interact with by email for a lifetime without ever detecting any hint that it was not a real human pen-pal like us -- then we would have no better or worse grounds for doubting that it had a mind than we do for a real pen-pal.

But it is also true that verbal capacities don't arise out of nothing: They are grounded in nonverbal capacities: the sensorimotor and learning capacities that we share with other species. So although Turing's original test was purely verbal, it has since become obvious that a realistic Turing Test would have to be robotic.

Turing thereby pointed out the methodology for the explanation of proximal mechanisms: Our proximal mechanisms are whatever give us the capacity to pass the Turing Test. Artificial intelligence, particularly robotics, is accordingly the study of proximal mechanisms. It is perhaps not so ironic that biology should converge with engineering. Richard Dawkins's metaphor of the "Blind Watchmaker" for the distal mechanism of evolution already points in that direction: We are survival/reproduction machines. Our genes vary, and are selected, automatically, and blindly, by their success in survival and reproduction. No one designed us with a purpose in mind. Surviving and reproducing are things that we do. So what is guiding the Blind Watchmaker is the same criterion as the Turing Test (TT), namely, performance capacity.

What is it that a candidate must be able to do, in order to pass the TT? It is too vague to reply "everything we can do." What can we do? To a first approximation, we "navigate" the world we live in, but although autonomous locomotion has proved to be a bigger task for robotics than perhaps expected, it is all the things we do with the things we encounter that has turned out to be the real challenge. Nor is it enough to say that what we do is survive and reproduce. Even if this were all reduced to feeding and mating, we would still be faced with the problem of categorizing: What kinds of things should we eat and not eat (I won't raise the analogous question about mates)?

William James picturesquely described the sensory input to the newborn baby as a "blooming, buzzing confusion." Let us leave aside for the moment what it feels like to sense that input, and simply ask Turing's question about the mechanism that is capable of processing that sensory input in such a way as to sort it into the kinds of things that are edible and the kinds of things that are not.

As noted earlier, we get some prior help from Darwin: some of our categories already come "prepared" in advance by evolution. The frog is already born with its fly-detectors tuned to the kinds of inputs towards which it is a good idea to flick out your tongue to try to catch them. But the distal origins of those detectors would need an explanation, and that explanation would probably be along lines similar to the explanation of how we acquire the much larger number of nonprepared categories that we must learn from experience.

So perhaps our most fundamental Turing capacity is the capacity to acquire categories -- the capacity to learn, by trial and error, guided by Skinnerian feedback from the consequences of categorizing correctly and incorrectly, what kinds of things there are that matter to our survival, and what to do with them. This requires somehow resolving the Jamesian confusion between what is in this category and what is in that one. In general, we have to learn to detect some features of the input and ignore others. Sometimes this is easy, as it would be if there were only two kinds of things, zebras and giraffes. Sometimes it is hard, as when we need to detect cancer cells, or the sex of baby chickens.

There are the beginnings of models for sensory category learning, but none that can do human-scale -- that is Turing-scale -- categorization. These are all still just toy models. Will we need to design a mechanism that can learn all of our categories by trial and error? Perhaps not, because there is evidence that we do not do it all that way either. I am not speaking of the "prepared" categories that we get "for free" from Darwin. There is a third way, and it not only sets us apart from other species, but it also reminds us of the original form of the TT, which was purely verbal. If Skinnerian trial-and-error learning, guided by corrective reinforcement, can be called acquiring categories robotically, then the other way to acquire categories is linguistically: Instead of having to learn them from experience, the hard, time-consuming and hence costly and often risky way, you can be told what's what by a verbal description from someone who already knows.

It is easy to forget what a huge evolutionary advantage such a capacity represents. Nonhuman species can learn categories the hard way, including the imitation of the performance of their fellow-creatures, but they cannot instruct one another explicitly, as our species can. We have done computer simulations that suggest that within just a few generations organisms that have the capacity to acquire categories verbally out-survive and out-reproduce organisms that can only acquire them robotically. "Language," if it can be called that, in these simplistic toy models, evolves quickly and quickly becomes a dominant way of acquiring categories.

But not the exclusive way. Not only do some, at least, of the older categories that we use in sentences that describe or define newer categories have to be grounded directly through the older robotic form of learning, but it is likely that linguistic categories need to be "refreshed" with direct sensorimotor experience at all levels. In principle, given a small, directly grounded start-up vocabulary of category names it should be possible to get all the rest from verbal definitions and explanations. But not even logic and mathematics are learned purely symbolically. We always need some direct experience with concrete examples.

It is tempting to nod here and say, based on one's own experience, that it is obvious that abstract linguistic descriptions alone, even if every word is fully understood, are not enough. That somehow our understanding is enriched in an essential way by direct experience: A picture we actually see is always worth more than a thousand -- or ten thousand -- words of description.

That is true, but remember that that we are speaking of here is learning about categories, about kinds of things, rather than about individual things in their infinite detail. Learning about kinds involves ignoring most of the irrelevant details of individual examples.

You may reply that even in order to recognise and understand and remember kinds it is important not just to hear an abstract description of their features and the rules for identifying them: that categorization is a sensorimotor skill, just as tennis is, and neither can be learned or used on the basis of verbal instruction alone.

That too is true, but behind it is yet another intuition that may well be wrong, although it is the most compelling intuition of all: It's not enough to hear an abstract description of what something looks like: You actually have to see it to really know what it looks like. After all, even the faint picture of it that you get from a verbal description is grounded in the real sensory experience.

Now we are back, however, to the same problem that made -- and still makes -- religious believers resist Darwinism. For the Darwin/Skinner/Turing explanation is all a blind, behavioral explanation. It explains the underlying causal mechanisms of what we can do. But it does not and cannot explain what we feel. A Skinnerian reward need not feel good (it need not feel like anything at all) in order to perform the mechanistic function of reinforcing behavior. The fact that a picture is worth a thousands words does not depend in any way on the fact that the picture is seen. The input need merely be transduced and processed by the machine. We can already demonstrate with simple, mindless, blind computers today how a picture is more than thousands of words. The same is true of the reduced information in a verbal description, compared to a direct sensory input. All of this is just as true of blind, mindless zombies as it is of our seeing, conscious selves. And neither Darwin's Blind Watchmaker nor the Turing Test are capable of discerning the difference.

So although I hardly think it is grounds for religious rejoicing, or for claims of empirical vindication, it can truly be said that the Darwin/Turing world is fully compatible with the existence of an immaterial soul.

BIBLIOGRAPHY

Cangelosi, A. & Harnad, S. (2001) The Adaptive Advantage of Symbolic Theft Over Sensorimotor Toil: Grounding Language in Perceptual Categories. Evolution of Communication 4(1)
http://cogprints.soton.ac.uk/documents/disk0/00/00/20/36/index.html

Dawkins, R. (1976) The selfish gene. Oxford : Oxford University Press

Dawkins, R. (1986) The blind watchmaker. New York : Norton

Harnad, S. (1987) (Ed.) Categorical Perception: The Groundwork of Cognition . New York: Cambridge University Press. http://cogprints.soton.ac.uk/documents/disk0/00/00/15/71/index.html

Harnad, S. (1990) The Symbol Grounding Problem Physica D 42: 335-346. "A Szimbolum-Lehorgonyzas Problemaja." Magyar Pszichologiai Szemle XLVIII-XLIX (32-33) 5-6: 365-383. http://cogprints.soton.ac.uk/documents/disk0/00/00/06/15/index.html

Harnad, S. (2000) Minds, Machines, and Turing: The Indistinguishability of Indistinguishables. Journal of Logic, Language, and Information 9(4): 425-445. (special issue on "Alan Turing and Artificial Intelligence") http://cogprints.soton.ac.uk/documents/disk0/00/00/16/16/index.html

Harnad, S. (2002) Turing Indistinguishability and the Blind Watchmaker. In: J. Fetzer (ed.) Evolving Consciousness Amsterdam: John Benjamins. Pp 3-18. http://cogprints.soton.ac.uk/documents/disk0/00/00/16/15/index.html

Harnad, S, Steklis, HD and Lancaster, JB. (1976) (Eds.) Origins and Evolution of Language and Speech. Annals of the New York Academy of Sciences 280. http://cogprints.soton.ac.uk/documents/disk0/00/00/08/66/index.html

Skinner, B. F. (1984) Selection by consequences. Behavioral & Brain Sciences 4: 477-510.

Turing, A.M. (1950) Computing Machinery and Intelligence. Mind 49 433-460 http://cogprints.ecs.soton.ac.uk/archive/00000499/