Probabilistic topic models and associative memory

Mark Steyvers
University of California at Irvine
Dept of Cognitive Sciences

Collaborative work with Tom Griffiths (Stanford University)

The topics model is a probabilistic approach to semantic cognition in which topics are represented as probability distributions over words. From a corpus of educational text documents, we were able to estimate in an unsupervised manner a large number of interpretable topics. We show how these topics can be used in a simple gist-based model for free recall to explain variability in eliciting false memories across study lists. The model assumes that study words activate a distribution of topics (“the gist”) which concentrates on one or more topics depending on the number of categories/themes in the study list. At retrieval, free recall is modeled as a reconstructive process using the stored topic distribution as well as verbatim memory traces as cues. We compare this probabilistic approach to the spatial framework of latent semantic analysis where words are represented as points in a high-dimensional semantic space.

Presentation (PowerPoint File)

Back to Probabilistic Models of Cognition: The Mathematics of Mind