GATSBY COMPUTATIONAL NEUROSCIENCE UNIT
UCL Logo

A factorial hidden markov model of perception and working memory

David E. Huber
Department of Psychology, University of California , San Diego , USA

In recent years, generative Bayesian belief networks have successfully characterized many information processing systems. Such models assume that conceptual representations are responsible for generating observations and that the goal of cognition is one of inference. Huber, Shiffrin, Lyle, and Ruys (2001) introduced a generative model of perceptual identification termed Responding Optimally with Unknown Sources of Evidence (ROUSE). The key mechanism within ROUSE is "explaining away" in which inference occurs between competing sources of an observation so as to reduce source confusion between primes and subsequent targets. Extending the model in time with a hidden markov structure allows demarcation of temporal onsets and offsets as well as determination of the temporal source (i.e., a new perceptual object versus one from the immediate past). By including multiple temporal identification buffers, the new model includes the original ROUSE model, but also incorporates the ability to account for presentation dynamics, capacity limitations, and sequential effects in working memory.