"Never Draw to an Inside Straight": On Everyday Knowledge

I

What would the structure of everyday reason look like if we tried to teach it to an intelligent machine? Something like this: Richard Powers's Galatea 2.2 tells the story of a novelist, coincidentally also called "Richard Powers," who returns to the United States after the breakdown of his marriage. He takes a fellowship at a midwestern university where he works as a kind of participant observer in the Center, a massively funded institute for the study of artificial intelligence. There he meets Lentz, a recognizable type of the mad scientist, with "freakish frontal lobes," a "monstrous beak," ¹ and an aversion both to natural light and to human contact. Lentz works with neural networks which mimic and perhaps reenact associative learning, and is training his current net to recognize beauty by playing it Mozart. With Powers as his research assistant he develps a project to build a neural network that will learn to interpret and comment on any text in the examination taken by Masters students in English for admission to Ph.D. candidacy; the test (a classic Turing test) will be taken in a blind competition with a real person.

The catch is that, in order to "understand" literary texts at this level of complexity, the network must understand everything: it must know not only the canonical texts of English literature, with their relevant contexts in other literatures, but all of the cultural encyclopaedia available to highly trained North American postgraduate students, as well as the encyclopaedia of human experience that underlies it, and it must have a detailed familiarity with the language, syntax, idioms, and conceptual and metaphorical logics in which these knowledges are formulated. One of the central genres in which the novel works is, accordingly, that of the treatise on learning, and specifically the study of child language acquisition. As the neural network develops through eight successive implementations, A to H, each of which incorporates the previous stage as a subsystem of an increasingly distributed and interdependent emergent system, the novel charts a progressive hierarchy of human knowledges from the formal to the very informal.

Thus Implementation A, which is fed massive vocabulary lists, learning to recognize words and to parse rudimentary syntax, fails just because of the power of its ability to recognize fixed verbal patterns: it learns and retains too much and, like Jorge Luis Borges's Funes with his photographic memory of particulars, is therefore unable to move from data to generalization. Its successor, Implementation B, is taught how to forget, and moves from A's pattern recognition to a form of computational linguistics which allows it to answer a riddling nursery rhyme (the one about the man going to St. Ives with seven wives), but to do so only because of its "unfailing literalmindedness" (G 95). B can generalize to cases but cannot move from cases to rules; it is replaced by Implementation C, set up as a structure of parallel processing between distributed subsystems, which can generalize about its own generalizations but which opens up, by its shortcomings, a further dimension of knowledge acquisition. For it to understand a word like "ball" it must not only process an almost infinite number of predicates and exceptions, but it must compensate for its lack of referential and affective knowledge: its incomprehension of what a ball feels like in the hands and in relation to a human body. C cannot match its verbal knowledge with visual, haptic, and kinetic knowledges, nor can it understand objects as part of physical and social interaction with others, grounded in time and in complex structures of exchange.

One of the problems for any rule-based learning process is that, beyond the level of formal systems, interpretation is contextual and contexts are almost infinitely extensible. Implementations G and H try to come to terms with this problem by building in recursive structures that will allow them to train themselves, to develop rules for encountering...