Proceedings of the Annual Meeting of the Cognitive Science Society

Automatic model generation based on user-task interactions is of great use for behavioral predictions and understandingof cognition. Mapping which environment features cause which actions seems like a classification problem suited forDeep Learning (DL). Unfortunately, DL does not create an observable model, and is more suitable to making predictionsfrom billions of examples than from limited observations. There are, however, many tasks that lend themselves to symbolicinput, allowing an alternative approach - Symolic Deep Learning (SDL). Symbolic hierarchical representations have a longhistory in Psychological literature, though some of these were integraged as models of memory without action-selection(e.g. EPAM/CHREST), and some have run into computational limitations (e.g. configural-cue). SDL stands to benefitfrom better model integration and modern growth in computational power and algorithmic efficiency, and promises to bethe right paradigm for automatic model generation from limited user observations.