Mate Lengyel
Gatsby Computational Neuroscience Unit, University College London

Bayesian model learning in human visual perception

Wednesday 15th of March 2006 at 10:00am
3105 Tolman Hall (Beach Room)

Humans make optimal perceptual decisions in noisy and ambiguous conditions. Such percepts have been shown to be the result of Bayesian inference integrating prior expectations about the external world with currently available sensory evidence. Key to this process is the generative model that specifies how these different sources of information are integrated. Thus, in order to understand the computational principles of perception, it is important to characterize the forms of generative models that are available for perceptual inference.

We modeled human behavior in a set of completely unsupervised visual learning experiments in which subjects performed surprisingly well despite being required to infer complex generative models from highly ambiguous data. We used Bayesian inference to learn a distribution of generative models that sought to explain sensory experience by independent hidden causes, and automatically preferred those models that were 'as simple as possible, but no simpler'. We validated our theory by quantitatively fitting behavioral data in a range of different experiments. We find that this framework also accounts for paradoxical aspects of human behavior in these tasks. These results demonstrate that humans can infer complex models from experience and implicate Bayesian model learning as a powerful computation underlying such basic cognitive phenomena as the decomposition of visual scenes into meaningful chunks.

Work in collaboration with Gergo Orban (Collegium Budapest), Jozsef Fiser (Volen Center for Complex Systems, Brandeis University), and Richard N Aslin (Department of Brain and Cognitive Sciences, University of Rochester).

(video)

Join Email List

You can subscribe to our weekly seminar email list by sending an email to majordomo@lists.berkeley.edu that contains the words subscribe redwood in the body of the message.
(Note: The subject line can be arbitrary and will be ignored)