Brave Neurocomputational World. Review of Paul Churchland, A Neurocomputational Perspective: The Nature of Mind and the Structure of Science. Cambridge MA: Bradford/MIT Press, 1988. Connection Science. 1991, vol 3, pp.91-3 Tim van Gelder Department of Philosophy Indiana University Bloomington IN 47405 [email protected] In the past decade connectionism came of age as a research paradigm in psychology and neuroscience. With the publication of Paul Churchland's collection A Neurocomputational Perspective, it comes of age as philosophy of mind as well. This book demands to be read by connectionists who wish to understand the philosophical context and ramifications of their work, and by philosophers who wish to understand connectionism and the nature of mind more generally. Of the fourteen chapters, ten were previously published elsewhere, with some having already acquired the status of classics in recent philosophy of mind. Each fills in an overlapping portion of Churchland's comprehensive and unified perspective on the nature of mind, knowledge and science; thus, while most can profitably be read independently, a full appreciation of his work only comes from reading the collection as a whole. Churchland describes himself as "a scientific realist, of unorthodox persuasion" (p.139). He is a realist because he interprets scientific theorizing literally - i.e., as telling us what the world is actually made up of, and as describing the regularities governing the world thus constituted. Churchland's realism is unorthodox, however, in acknowledging that current theories may well be radically false, that the entities to which they purportedly refer may turn out never to have existed, that we may never have a final theory which gets everything exactly right, and even that the basic medium of scientific theorizing, language itself, may have fundamental epistemic limitations. He is a scientific realist not just because he is a realist about science, but because he believes that all knowledge is basically scientific - i.e., "conjectural and theoretical" - in nature (p.282). In particular, he rejects the old idea that science is a speculative, theoretical superstructure constructed upon a basis of observations that are themselves neutral between theories. Rather, the observational claims are always framed in terms of a vocabulary drawn from some scientific framework or other, and hence automatically incorporate the theoretical commitments and the vulnerability of that framework. Likewise, perception is classification of the environment according to the categories provided by some conceptual framework or other, a framework that is only acquired by mastery of some (scientific) theory of the domain; since we can always go on to master different theories, perception is fundamentally plastic in nature. Importantly, our ordinary commonsense knowledge of the everyday world turns out to be theoretical in nature, since it is framed in terms of simple yet dependable "folk" theories of the world around us. In particular, our knowledge of how people behave is part and parcel of "folk psychology," a thoroughly familiar theory invoking internal states

2 such as beliefs, wants, moods, pains and so forth, and connecting them with each other and what people do by means of platitudinous generalizations. A central feature of folk psychology is the idea that our knowledge of the world comes in the form of endless stored propositions, a notion that has been theoretically refined into sentential epistemologies (theories of knowledge which take sentences as the basic units of knowledge and rationality) and "Language of Thought" models in cognitive psychology and AI (theories of cognition which take sentence-like entities as the basic units of cognitive operation). The trouble with folk psychology is that it is a poor theory, and therefore susceptible to reduction or elimination via more powerful conceptual frameworks. This is where connectionism enters the picture (and where, incidentally, A Neurocomputational Perspective really starts to extend beyond Churchland's 1979 book Scientific Realism and the Plasticity of Mind.) Connectionism, in partnership with neuroscience and portions of cognitive psychology, is developing new and deeper theories of cognition based on our understanding of the brain as a complex array of dedicated, layered neural networks. These theories promise to reduce certain aspects of our mental lives (such as sensations, which turn out to be patterns of activity in sensory networks, their qualitative character explained in terms of location in sensory state space) and eliminate other aspects (such as propositional attitudes, which no longer offer explanatory purchase in view of the different and more powerful network-based explanations). The core idea is that mental representations are high-dimensional activation vectors, computation is vector transformation from one layer of a network to the next, and long term knowledge takes the form of vast matrices of weighted synapses. Cognition essentially amounts to the activation, by stimulation at the sensory periphery in concert with context-setting internal feedback, of "prototypical" vectorial representations which feed in turn to networks producing the appropriate motor response. Churchland does not hesitate to explore the many radical consequences of these views. They suggest, for example, the possibility that we might eventually possess vastly enriched introspective capacities once we develop the ability to perceive our own internal lives in the detailed terms provided by this new, more penetrating conceptual framework. They also have startling ramifications for the philosophy of science, and in particular for our understanding of what theories and explanatory understanding are. Traditional philosophy of science, beholden to sentential epistemology, sees a scientific theory as a body of sentences, knowledge of that theory as storing those sentences, and explanatory understanding as a matter of deduction from universally quantified conditionals. After stressing the numerous shortcomings of this approach, Churchland presents a new picture of an individual's theory of the world as his current configuration of synaptic weights, a configuration which serves to partition the space of possible activity patterns into a vast number of prototypical regions; explanatory understanding then consists in the activation of the prototype most appropriate to the situation currently confronted. In Churchland's brave neuroscientific world, language has been displaced both as the medium of thought and as the form of scientific theorizing. The persistent difficulties encountered by sentential models in psychology, epistemology and philosophy of science can be traced to fact that the fundamental mode of human cognition is non-linguistic, permitting at best superficial approximation in linguistic terms. Consequently we must reject the idea that truth (an inherently linguistic notion) is the ultimate epistemic virtue, and adopt instead pragmatic criteria such as efficacy in our assessments of the virtue of

3 conceptual frameworks. Our vastly improved understanding of people, including ourselves, will be an understanding of evolving systems of layered networks, whose countless synaptic interconnections embody practical know-how for dealing with the world, supporting and transforming high-dimensional, prototypical patterns of activation. The dizzying vastness of the space of possible theories of the world, conceived of as points in synaptic weight space, underscores the both the plasticity of perception and the incredible diversity of conceptual schemes, and casts doubt on any naive idea that we will eventually uncover any one definitive theory of the world. This is philosophy of mind at its best. Not for the intellectually timid, the ideas presented threaten to thoroughly revolutionize our understanding of mind and its place in the world, if only we are sufficiently daring to explore them. Churchland writes in a clear, compelling and entertaining style; his theses fit together to form an elegant overall perspective, and are always carefully argued. Well-informed about the relevant empirical research, he also has a confident command over the deep and complex philosophical issues involved. A Neurocomputational Perspective will provoke vigorous debate and no doubt some violent reactions. Alternatively, for those sympathetic with its basic aims and perspective, it opens up vast areas of research. Thus, for example, much more remains to be said about how the kind of fundamentally non-linguistic cognitive systems Churchland describes can be responsible for our own enormously subtle and complex linguistic behavior. This is a fascinating and difficult topic which he barely touches on (see, e.g., p.195). As connectionists probe these kinds of problems, we can expect powerful new kinds of neurocomputational architectures to emerge, architectures which will no doubt inspire even more sophisticated conceptions of knowledge representation and cognition than can be drawn from the kinds of early backpropagation networks which dominate the discussion in these essays. In particular, one drawback of these simple layered networks is that they present a rather static picture, with primary emphasis placed on points in state space and how they fall into complex regions of cognitive significance. This downplays the temporal dimension of cognitive activity: some more recent connectionist work suggests that the actual basic unit of cognitive significance may well turn out to be something like the activation trajectory. There is also plenty still to be done fleshing out Churchland's pioneering but often somewhat sketchy applications of neurocomputational conceptions of cognition to traditional philosophical concerns such as the influence of social and cultural factors on a person's world view, the nature of our understanding of others, of morality, and so forth. Vaguely connectionist metaphors often find easy application in all kinds of areas; if such metaphorical redescription is to amount to genuine philosophical illumination, we will eventually want more systematic detail than is provided in many of Churchland's tantalizing suggestions. Pointing out shortcomings of this kind is merely a different way of recommending the book. The philosophical setting of connectionism, Churchland demonstrates, is at least as exciting and fertile as connectionist work itself.