The engine of reason, the seat of the soul

For the uninitiated, there are two major tendencies in the modeling of human cognition. The older, tradtional school believes, in essence, that full human cognition can be modeled by dividing the world up into distinct entities -called _symbol s_-such as “dog”, “cat”, “run”, “bite”, “happy”, “tumbleweed”, a nd so on, and then manipulating this vast set of symbols by a very complex and very subtle set of rules. The opposing school claims that this system, while it might be good at concluding that Paris is the capital of France or that there must be blood flowing in the left-rear leg of a cow, can never capture the full measure -indeed, the esse nce -of human cognition. For them, the esse ntial features of cognition emerge from the combined effects of myriad, tiny actions far below the surface of consciousness. This is the camp to which Paul Churchland belongs. Now, let us turn to Churchland’s book, _The Engine of Reason, the Seat of the Soul _. It is a clearly written, easily understood presentation of some of the most important ideas and impressive contributions of connectionism. He leads the reader s tep by step through various kinds of “connectionist” networks, from the simple backpropagation networks developed in the ear ly 1980’s through the recurrent networks that were developed in response to problems that the simpler networks could not handle. He extrapolates from these networks to vastly larger, vastly more powerful networks that he believes will ultimately lead to a full simulation of human cognition. He describes a number of fascinating case s tudies, including Charles Rosenberg and Terry Sejnowski’s NETtalk, an early connectionist network that learned to pronounce English words. His excellent discussion of NETtalk accurately captures the excitement that this seminal program generated around 1986 when it first forced many traditional artificial intelli gence researchers to sit up and take connectionism seriously. Perhaps more than any other program in the field, NETtalk was responsible for the tremendous surge of interest in connectionism and in emergent (“bottom up”) models of cognition. The book includes a detailed and extremely interesting chapter on connectionist approachs to stereoscopic vision, detection of mines by submarines, pronounciation and, even, crab movement! Churchland carefully explains why recurrent networks, as opposed to simple backpropagation networks, must be used to process seque nces of events. There are chapters on brain dysfunction, consciousness (including some ground-breaking work by Rodolfo Llinas on neo-cortical oscill ations and the Crick-Koch hypothesis that these oscill ations may be the seat of consciousness), and potential technical uses of neural networks, including medical diagnosis. It all makes for truly fascinating reading. There are, however, a number of important problems with this book that cannot be ignored. To begin with, the book all too frequently reads like an “infomercial” for connectionism and “prototype vectors”. Infomercials, as everyone knows, contain a certain amount of truth wrapped in hyperbole and sold with evangelistic fervor. This is emphatically not what the neural network research program needs. When enthusiasm for an idea causes its proponents to intentionally downplay, overlook or conceal major diff iculties with it, the inevitable result is not only bad science, but a disill usioned public. The first page of the book is almost certainly the worst of all. Churchland writes: