Cameron Buckner (University of Houston)
This presentation will focus on the way anthropomorphism/anthropocentrism interface in attempts to solve the "interpretation problem" facing deep learning neural networks. For example, researchers have begun using methods to correlate network performance with verbal justifications recorded from human subjects reaching a similar outcome in a similar situation. Worries of anthropomorphism and anthropocentrism should be considered here, however, given that the justifications are generated ad hoc using powerful statistical methods, rather than playing a causal role in the networks' "decisions". Much could be learned by reviewing methodological considerations from comparative psychology and the psychology of human introspection.
