PHIL 29903/39903 The Philosophy of AI: Induction in the age of Big Data
Recent developments in artificial intelligence have brought about a radical reconceptualization of our idea of knowledge work. The model of the laboratory scientist, whose task is to conduct elaborate experiments that probe, in minute detail, the correctness of a theoretical hypothesis, is gradually giving way to that of the data scientist, whose concern is to wrangle massive datasets in an effort to extract from them reliable predictions with only a minimal theoretical guidance. In this course, we will explore some of the epistemological implications of this AI-driven shift in our conception of knowledge and the work that goes into acquiring it. Focusing on applications of artificial intelligence that utilize feed-forward deep neural networks for statistical inference, we will investigate what the shift to "big data" means for our philosophical theories of induction. Are the learning algorithms employed in the training of deep neural networks really "theory free"? If so, why should we trust that their predictions are reliable? How do neural networks purport to solve the curve-fitting problem and Goodman's new riddle of induction, without giving weight to theoretical virtues such as simplicity? Without a background of causal knowledge to structure their inferences, how do neural networks distinguish between causation and mere correlation, and if they cannot, why should we allow their predictions to serve as inputs to a theory of rational decision making? (B) (II)