Revealing interpretable object representations from human behaviour
A major area of interest in cognitive science is the study of how object concepts are represented in the mind, and how these representations relate to human behaviour. There have been many attempts to represent object concepts in terms of semantic features, a vector of variables indicating the presence of different aspects of their meaning. These are usually obtained by interviewing subjects, or simply postulated by researchers. Given the near-infinite number of tasks or contexts of usage for concepts, however, these approaches are necessarily subjective and often limited in terms of how many objects are covered. In this talk, I will describe the ongoing effort by our collaborators to collect a behavioral dataset of millions of judgements on thousands of objects, and our approach to derive a sparse, non-negative embedding representation of those objects from the data. The dimensions of this embedding are interpretable, conveying degrees of taxonomic membership, functionality, and perceptual attributes, among other characteristics. Furthermore, they capture the latent similarity structure between objects, and can be used to build effective prediction models for a variety of human behavioural judgements, including categorization, typicality, and feature ratings.