Whenever science operates at the edge of what is known, it runs into general issues about the nature of knowledge and reality. Mundane science can operate without much concern for methodological and ontological issues, but frontier science cannot avoid them. For example, innovative research in theoretical and experimental physics inevitably encounters fundamental problems about the nature of space and time, as well as methodological questions concerning how scientific investigation should proceed. Cognitive science has made substantial progress in investigating phenomena such as perception, memory, learning, problem solving, and language use, but clearly it is still a frontier enterprise.
Here are some central questions that arise in cutting-edge research in cognitive science: What is an explanation? What is a theory? How can competing theories be evaluated? What is the relation between different fields of cognitive science such as psychology and neuroscience? What is the role of computer modeling in cognitive science? I will not attempt to answer these questions here, but I will provide some pointers to philosophical work that addresses them in ways that I think are useful for the self-understanding of cognitive science research, encouraging progress rather than hindering it. This list of questions is not intended to be complete, but it provides a sample of the sort of philosophical issues encountered by cognitive science at its most interesting.
2.1. What are theories and explanations in cognitive science?
Cognitive science operates with a multitude of conferences, societies, journals, research groups, and educational programs, but it rarely stops to ask what is the point of all this activity. In my view, the main aims of interdisciplinary research on mind and intelligence are to understand how the human mind works and to use this understanding to develop ways of making humans and machines more intelligent. If you disagree with this opinion, then we need to have a very general and hence philosophical discussion about what cognitive science is for, a question that cannot be answered just by the experimental and theoretical methods of cognitive science.
Cognitive science provides understanding by giving accounts of the nature of key phenomena such as inference. Such accounts are provided by theories that can be used to explain the results of experiments. Here are some possible answers to the question of what constitutes a scientific theory:
A cognitive theory is a set of mathematical formulae used to make predictions about behavior.
A theory is a computer program that simulates thinking.
A cognitive theory is a description of mechanisms that explain observed mental phenomena.
I prefer the third view of theories, because it fits well with the most successful practices in psychology and neuroscience, as well as related areas in biology and medicine (see, for example, Bechtel, 2008; Bechtel & Richardson, 1993; Craver, 2007; Darden, 2006; Thagard, 1999, 2005, 2006). But the first two options have also been assumed by leading contributors to cognitive science, so there is a philosophical issue here that deserves to be debated by anyone reflective about how the mind is to be understood.
In physics, it is often claimed that the primary function of a theory is to generate predictions, but explanations are at least as important. There is a long history of philosophical discussion of the relation between prediction and explanation (e.g., Hempel, 1965). The making of precise predictions is much more difficult in the biological than in the physical sciences, but explanations are just as important. Here are some candidate answers to the question of what is an explanation:
An explanation is an answer to a question about why something happened.
An explanation is a deductive derivation of a description of a phenomenon from a set of principles.
An explanation is a description of how the operation of a mechanism produces a phenomenon.
Consonant with my preferred view of theories, I see explanations as primarily mechanistic, but this view is highly controversial in the philosophy of science, and I expect that many cognitive scientists would find it puzzling or possibly odious.
In reply, I would argue that the most successful theoretical explanations in cognitive science, for example, using rule-based and connectionist ideas, have been mechanistic in the sense elucidated by philosophers of science. A mechanism is a system of parts whose interactions produce regular changes. Rule-based systems such as GPS (Newell & Simon, 1972) and ACT (Anderson, 2007) are clearly mechanistic in this sense, as are neural networks models such as PDP models (Rumelhart & McClelland, 1986) and more recent models closer to actual brain mechanisms (e.g., Eliasmith & Anderson, 2003). Then the primary difference between conflicting theories is the postulation of different kinds of parts and interactions as responsible for the psychological phenomena that everyone wants to explain.
How can cognitive science adjudicate between competing theories about the mechanisms that make mind works? There are various philosophical answers to this question, ranging from hypothetico-deductive (Popper, 1959) to Bayesian (Sober, 2008) to explanatory coherence (Thagard, 1992). This is not the place to defend my own favorite answer, which is based on a cognitive model that has had many philosophical and psychological applications. My point is that the issue of theory evaluation supports my Santayana- and Keynes-inspired remarks in the first paragraph. When cognitive scientists do not have explicit views about the methodological issues concerning theories, explanations, and evaluations, it is not because they do not have any views, just that the views they hold are usually implicit and unreflective. Methodological issues inevitably arise in key disputes in cognitive science, for example, concerning mental imagery and the value of abstract Bayesian models that assume that optimality is a property of human information processing. Such issues need to be confronted by scientists with the assistance of philosophy, not buried as if the answers to the philosophical questions were obvious.
Explanations in cognitive science often help themselves to the concept of causality without addressing longstanding philosophical issues about its nature. Is causality a matter of constant conjunction, mental schemas, probability, special powers, manipulability, energy transfer, or nothing at all? My own view is that the basic human concept of causality is a complex multimodal, neural representation of changes that includes both sensorimotor encodings of manipulability and verbal encodings of regularity (Thagard & Litt, 2008). Regardless of whether this view is correct, cognitive science needs to blend its frequent use of causality in explanations of human thought and behavior with philosophical reflection about what causality amounts to.
Philosophers do not have some special, a priori ability to settle questions about the nature of theories and explanations. But they have (or should have) awareness of a wider range of answers to these questions, as well as familiarity with the main answers that have been proposed in the past. Moreover, philosophy has been accustomed to ask such questions in full generality, so that they cover other sciences such as physics and biology as well as the various disciplines that make up cognitive science.
2.2. What is the role of computer modeling in cognitive science?
The use of computer models has been an important feature of cognitive science since Newell, Shaw, and Simon (1958) produced the first computational account of human problem solving. The organizational beginnings of cognitive science in the late 1970s, heralded by formation of the journal Cognitive Science and the Cognitive Science Society, explicitly looked for research that combined psychology and artificial intelligence. Since then, computational models have flourished in many interdisciplinary branches of cognitive science, including computational neuroscience, computational linguistics, computational organization science, and even computational philosophy. But what are these computer models actually contributing to cognitive science?
It is sometimes said that a computer program can be a cognitive theory, but I think this is a mistake, because programs are always full of minor details tied to the particular kind of programming language they are written in. For example, a program written in the common AI language LISP will have lists as its most common data structure, but I know of no LISP modeler who claims psychological significance for this particular form of representation. Rather, it is crucial to distinguish between theories, models, and programs (Thagard, 2005).
In keeping with my preferred account in the previous section, I think a theory is a description of an explanatory mechanism, consisting of parts and their interactions producing regular changes. In a psychological theory, the parts are mental representations such as concepts, rules, and images; and the interactions are processes such as spreading activation and rule matching and firing. In a neural theory, the parts are neurons and neural groups, and the interactions include excitation, inhibition, and learning by weight changes. In an organizational theory, the parts are agents and groups of agents, and the interactions include communication and other kinds of influence. In all these cases, theories are used to make claims about the kinds of mechanisms that produce various kinds of intelligent behavior.
The mechanisms in all these aspects of human mental activity are extraordinarily complex, so cognitive science needs to study them by constructing models, just as occurs in all fields of modern science such as physics. Mathematics is an invaluable tool for describing the interactions between parts, providing general and simplified descriptions of the changes that can result from the interactions. Models differ from theories in providing idealized descriptions of how the mental world is supposed to work. Such idealizations and simplifications make possible detailed explanations and predictions of the results of experiments.
However, to draw out the consequences of a model, it is often crucial to produce a simulation by writing a computer program that implements the mathematical assumptions of the model and the main mechanistic claims of the theory. Generation and testing of a computer program written in LISP, C++, Matlab, or some other higher-level programming language is invaluable in determining whether the idealized parts and interactions actually behave in ways that produce the results expected from observations. I do not mean to suggest that cognitive scientists always proceed from theory to model to program, because the process of discovery often leads in the opposite direction. Thinking about how a program might work can suggest a model that grows into a full-fledged mechanistic theory. But theory, model, and program remain distinct.
I do not expect all cognitive scientists to accept this account of the relation of programs, models, and theories, although I think it gives a good account of decades of practice by myself and others in developing psychological, neural, and organizational theories and programs. More widespread is David Marr’s (1982) distinction among computational, algorithm, and implementation levels, which I think is misleading in various respects (e.g., I think that the theories are about mechanisms, not computation). But I will not try to settle the issue here. My crucial point is that the self-understanding of cognitive science cannot simply presuppose any given account of the relations among theories, models, programs, and experiments. These relations have been extensively discussed by philosophers of science, with insights about common practices in physics, biology, and other sciences that need to be applied (or, if necessary) withheld from cognitive science. Ignoring such issues goes hand in hand with simply adopting a philosophical view that may be deeply flawed.
2.3. What are the relations among cognitive science disciplines?
Cognitive science involves at least six integral disciplines: psychology, neuroscience, linguistics, philosophy, anthropology, and artificial intelligence. A major question of the highest generality concerns the relations among and between these disciplines. When cognitive science was officially organized in the late 1970s, the diagram in Fig. 1 served to illustrate actual and possible connections among the six disciplines. An important philosophical problem that should interest all participants in cognitive science concerns the nature of such connections.
Figure 1. Connections among the cognitive sciences, based on Gardner, 1985, p. 37, from a 1978 Sloan Foundation report. Unbroken lines indicate strong interdisciplinary ties, and broken lines indicate weak ones. The ties between philosophy and both neuroscience and artificial intelligence are much stronger today.
Download figure to PowerPoint
One very useful way of thinking about the relations among these disciplines is to think of them as operating at different levels (Churchland & Sejnowski, 1992; Craver, 2007; Newell, 1990). Anthropology operates at the social level, dealing with the interactions among individuals as part of a culture. Psychology operates at the individual level, concerned with the mental representations and processes of individual thinkers, drawing on ideas from linguistics and artificial intelligence. Psychology, linguistics, and AI can also consider interactions between individuals, as in social psychology, sociolinguistics, and multiagent systems. Neuroscience operates below the psychological level, concerning itself with neural networks. Understanding of neurons often draws also on molecular processes, for example, how genes produce proteins within cells enabling the operations of neurotransmitters such as dopamine and serotonin. What are the relations among operations at these four levels?
Fig. 2 displays four commonly advocated views of such relations. The most familiar is (A), the classical reductionist view that changes at lower levels cause changes at higher levels. It could obviously be taken down to still lower levels involving atoms, subatomic processes, and quantum mechanical effects; but these do not seem relevant to cognitive science as it is currently practiced, so I shall ignore it (see Litt, Eliasmith, Kroon, Weinstein, & Thagard, 2006 for an argument that the brain is not a quantum computer). On this view, causality runs upward and so should explanation: social changes are explained as the result of psychological changes, which are the result of neural changes, all the way down to subatomic changes. This view is far from universally accepted and indeed there are intellectual circles where “reductionist” is an epithet almost as vitriolic as “idiot” or “bigot.”Bickle (2003) unabashedly defends a position he calls “ruthless reductionism.”
Figure 2. Four views of the relations between levels of explanation in cognitive science. Arrows indicate causality.
Download figure to PowerPoint
In the social sciences, some writers go far in the other direction, suggesting that the social level is the key source of causality. For example, Karl Marx said that the ideas of the ruling class are in every epoch the ruling ideas, suggesting that psychology is determined to a large extent by economics and sociology. There are many dozens of books written by sociologists and historians with titles of the form “the social construction of X” (Hacking, 1999). Notoriously, Latour and Woolgar (1986) argued that science is a social construction and argued for a moratorium on cognitive explanations of science. On this view, causality and explanation run only downward, from the social to the psychological; the neural and molecular levels are largely ignored.
A more moderate, less imperialistic form of anti-reductionism is the autonomy view, (C) in Fig. 2, where the dotted lines indicate that explanations at each level can proceed independently. This view is popular among sociologists, economists, and anthropologists who want to maintain their independence from psychology without making strong claims of social constructivism. Similarly, some psychologists and philosophers of mind have wanted to defend psychology from the rapidly increasing incursion of neuroscience. One standard argument for defending the autonomy of psychology from neuroscience is the argument from multiple realisability. On this argument, mental states and processes can be instantiated in many different kinds of functional architectures such as robotic ones, not just neurons, so there is an autonomous level of psychological explanation. There have been various philosophical responses to this argument (e.g., Bechtel, 2008; Thagard, 1986) but I think that the most powerful response is just observation of current research trends. Not only cognitive psychology but also social, clinical, and developmental psychology are being increasingly tied to neural processes. Similarly, at the social level, economics is becoming increasingly influenced by behavioral and neural approaches. Hence, the autonomy view is becoming increasingly obsolete.
My own preferred view is the highly interactive one (D), in which there are causal interactions and hence explanatory relations among all levels. This view is not reductionist, because it rejects the one-way causal connections shown in (A), nor is it anti-reductionist, because it recognizes that molecular processes are part of the explanation of neural events, neural processes are part of the explanation of psychological events, and psychological processes are part of the explanation of social events. Elsewhere I defend this multilevel view in detail with respect to explanation of human emotions (Thagard, 2006) and consciousness (Thagard, forthcoming; Thagard & Aubie, 2008).
Many philosophers and scientists are suspicious of the idea of downward causation as somehow spooky or mystical, but it seems to me unproblematic. Here are just a few examples of cases of the most extreme kind of downward causation, where I think it is legitimate to say that social interactions cause molecular changes:
Having to give a presentation increases levels of the stress hormone cortisol.
Seeing a beloved causes increased activity of dopamine neurons.
Men whose favorite sports team has won a game enjoy increased levels of testosterone.
Male chimpanzees who become dominated have lowered levels of testosterone.
Women who room together tend to have their menstrual cycles coordinated, altering patterns of estrogen levels.
In short, social changes cause molecular changes.
Craver and Bechtel (2007) argue against the idea of downward causation between levels, but I will not try to respond to their arguments here. The point of my discussion is not to settle the issue, but to show that there is an important general question about levels of explanation in cognitive science that requires philosophical investigation. Ignoring these questions amounts to quiescently adopting one of the four views, usually (A) or (C), without reflection or justification. Progress in cognitive science across the full scope of its ambitions requires assessment of what view of the relations between levels of explanation contributes most to innovative and successful theories and experiments.
I have presented only four of the philosophical questions that are crucial to the successful operation of cognitive science, concerning the nature of theories, explanations, computer models, and relations among contributing disciplines. My point has not been simply to assert my own preferred answers to these questions, but to show that there are important controversies that must be addressed as part of a complete, high-level understanding of what cognitive science can accomplish. Ignoring such issues usually amounts to repeating philosophical positions about the nature of scientific knowledge that have proved inadequate in the past, in both the natural and social sciences. Behaviorism, for example, flourished in part because it meshed with the philosophy of positivism, which unduly restricted science to what is observable. The official styles of writing in journals in experimental psychology encode a hypothetic-deductive picture of science that does not fit either with actual practice in psychological research or with currently available philosophical discussions of the relations between science and evidence. In such cases, ignoring philosophy just leads to assumption of persistent but inadequate philosophical views about the general nature of investigation. The best science is highly philosophical because it pays attention to general issues as well as normative ones.