Concepts, Experts, and Deep Learning

Deep learning has recently emerged as a new powerful source of AI algorithms, applied with amazing success for games (chess, Go), classification, face recognition, natural language translation, and administrative decision making. In this paper I analyze the somewhat surprising philosophical significance of deep learning by three themes. First, concept formation by means of massive amounts of training examples in unsupervised learning seems to reinforce a crude form of Lockean concept empiricism, in contrast to its Kantian alternatives. Secondly, deep learning algorithms seem to achieve the highest status of intuitive experts in Hubert L. Dreyfus’ account of skill acquisition, so that his criticism of AI has to be reconsidered. Thirdly, deep learning with neural networks is a bottom-up approach, in contrast to the top-down approach of traditional AI, yet as sub-conceptual it is not associated with conceptual or propositional content.Peer reviewe