The Role of Evolutionary Biology and Computational Theories in Cognitive Neuroscience

The cognitive neuroscience of central pmccssca is currently a mystery. The brain u a vast and complex collection offitnctionally integrated circuits. Recognizing that natural selection engineen a fit between saucnur and kc t ion is the key to isolating thew circuits. Neural circuits were designed to solve adaptive problemr If one caa define an adaptive problem doody enough, one can see which circuits have a structural design that ia capable of solving that problan. Evdutiony bidagiur have developed a saia of sophiaticatcd mod& of adaptive probiems. Some d there models analyze coarrtrabta on the evolution of the cognitive procasa that gwcrn social behavior: cooperation, threat, courtship, kindirected asairtance, and so on. T h e fonns of social behavior are generated by complex computational machinery. To discover the functional architecture of this machinery, cognitive neuroscientists will need the powerful inferential tooh that evolutionary biology provides, induding its welldefined theonca of adaptive function. The cognitive scicnccs have bem conducted as if D d n never lived. Their goal k to isolate functionally integrated subunits of the brain a d determine how they work. Yet m a t cognitive scicntiats punue that goal without any clear ~mtion of what "function" means in biolagy. When a neural circuit is discovered, very few researchers ask what its adaptive function is. Even fewer use theories of adaptive function as tools for discovering heretofore unknown neural systems. InLEDA COSW~OU Department of Psychology, and JOHN TOOBY Department of Anthropology, Center for Evolutionary Psychology, University of California, Santa Barbara, Calif. deed, many people in our field think that theories of adaptive function are an explanatory luxury--f8na unEalsifiable speculations that one indulges in at the end of a project, after the hard work of figuring out the structure of a circuit has been done. In this chapter, we will ague that theories of adap tive hnction arc not a luxury. They an a necessity, crucial to the hturc development of cognitive neuroscience. Without them, cognitive neuroscientists will not know what to look for and will not know how to interpret their results. As a result, they will be unable to isolate functionally integrated subunits of the brain. Explanation and discooq in cognitive neuroscience [tlrying to understand perception by studying only neurons is like vying to understand bid fight by studying only feathm: it just cannot be done. In order to u n d d bird flight, we have to undentand aerodynamics; only then do the s a c turc of feathers and the diffetcnt shapes of birds' wings make sense. (Marr, 1982.27) David Marr developed a general explanatory system for cognitive science that is much cited but rarely applied. His three-level system applies to any device that processes information-a calculator, a cash register, a television, a computer, a brain. It is based on the following observations: 1. Information-processing devices are designed to solve problems. 2. They solve problems by virtue of their structure. COSMIDES AND TOOBY: EVOLUTIONARY BIOLOGY AND COGNITIVE NEUROSCIENCE 1 199 3. Hence, to explain the structure of a device, one fer in how they represent information, in the procleeds to know whereby they transform input into output, or both. So a. whd problem it was designed to solve, and knowing the goal of a computation does not uniquely b. why it was designed to solve that problem and not determine the design of the program that realizes that ome other one. goal in the device under consideration. [n other words, one must develop a task analysis of the 5. Many different physical systems-from neurons problem, or what Marr called a computoriod thmy. in a brain to silicon chips in a computer-can impleKnowing the physical structure of a cognitive device ment the same program.' So knowing the structure of and the information-processing program realized by a program does not uniquely determine the properties that structure is not enough. For human-made artifacts of the physical system that implements it. Moreover, and biological systems, form follows function. The the same physical system can implement many prophysical structure is there because it embodies a set of grams, so knowing the physical properties of a system programs; the programs are there because they solve a cannot tell one which programs it implements (table 79.1) $ particular problem. A computational theory specifies : what that problem is and why there is a device to solve A computational theory defines what problem the i it. It specifies the fwction of an information-processing device solves and why it solves it, but it does not specify 4 device. Marr felt that the computational theory was how this is accomplished; theories about programs and : the most important and the most neglected level of their physical substrate speclfy how the device solves explanation in the cognitive sciences. the problem. Each explanatory level addresses a differThis functional level of explanation has not been ent question. To understand in information-processing neglected in the biological science, however, becaw device completely, Marr argued, one needs explanzit is essential for understanding how natural selection tiona on all t h m levels: computational theory, pradesigns organisms (for background, see chapter 78). gramming, and hardware (set table 79.1). An organism's phenotypic swucture can be thought of A computational theory of function is more than an as a collection of design ftaturcs-of machinu, such as explanatory luxury, however. I t is an asential tool fbr the eye or liver. A dcsign feature can cause its own discovery in the cognitive and neural sciences. Whethspread by solving adaptive problems--problems, such er a mechanism was designed by natural selection or as detecting predators or deroxifying poisoas, that recur over many generations and whose solution tends to promote reproduction. Natural selection is a feedback TABU 79. I process that "chooses" among alternative designs on Thrae h i s at which any mclchinr ca@ng out an informationthe basis of how well they function. By selecting designs proces.siug tad must be &stood 1 il on the basis of how well they solve adaptive problems, 1. C tlirorl: * . thL proca en&an a mt ktwthe won of What is the god of the computation. why ia it appmpdate, and what is the logic of the strategy by which it can be a device and its structure. To understand this c a d oue relationship, biologists had to develop a theoretical 2. R*m vocabu1w that d i s t inmes saucnue and How computational theory be implemented? In function. Marr's computational theory is a functional puticul~, what k the nprucntation for the input and outlevel of explanation that corresponds roughly to what put, and what k the algorithm for the tradormation? biologists refer to aa the "ultimate" or "functional" 3. Hardware implmmmbn: explanation of a phenotypic structure. How can the representation and algorithm be realized Even though there is a closc causal relationship bephficallyT tween the function of an information-processing device In duh'0~9 b i o l o ~ : and its structure, a computational theory of a device Explanations at the level of the computational theory arc called ultimate-level explanations. does not its structure. This is because Explanntions at the level of mprr.rntation and algorithm, there are many ways to skin a cat. More precisely: or at the level of hardware implementation, a n called proxi4. Many different information-processing programs mote-level explanations. can solve the same problem. These programs may difFrom Marr, 1982,25. 1 200 EVOLUTIONARY PERSPECTIVES by the intentional actions of a human engineer, one can count on there being a close causal relationship between its structure and its Function. A theory of function may not determine a program's structure uniquely, but it reduces the number of m b i l i t i u to an empirically manageable number. Task demands ndically constrain the range of possible solutions; consequently, very few cognitive programs are capable of solving any given adaptive problem. By developing a careful task analysis of an information-processing problem, one can vastly simplify the empirid search for the cognitive program that solves it. And once that program has been identified, it is easy to develop clinical tests that will target its neural basis. It is currently fashionable to think that the findings of neuroscience will eventually place strong constraints on theory formation at the cognitive level. In this view, once we know enough about the properties of neurons, neurotransmitters, and cellular development, figuring out what cognitive programs the human mind contains will become a trivial task. This cannot be me. Thttt are millions of animal species on earth, each with a different set of cognitive programs. Ihr smnr basic mural tisstu mrbodks aU of these program. Facts about the properties of neurons, neurotransmitters, and cellular development cannot tell one which of these milliona of program the human mind contains. The cognitive structure of an information-processing device "depends more upon the computational problems that have to be solved than upon the particular hardware in which their solutions are implemented" (Marr, 1982, 27). In other words, knowing what and whg allows one to generate focused hypotheses about how. To figure out how the mind works, cognitive neuroscientists will need to know what problems our cognitive and neural mechanisms were designed to solve. Bdyond intuition: How to build a computational