Inference and Prediction in Neocortical Circuits

September 21 to 24, 2003

at the

American Institute of Mathematics, Palo Alto, California

organized by

Jeff Hawkins, and Bruno Olshausen

This workshop, sponsored by AIM, NSF, and RNI will be devoted to working toward an understanding of inference and prediction in neocortical circuits.

The cerebral cortex is responsible for most of our conscious experience, yet we remain largely ignorant of the principles underlying its function despite progress on many fronts of neuroscience. The principal reason for this is not a lack of data, but rather the absence of a solid theoretical framework for motivating experiments and interpreting findings. The purpose of this workshop is to bring together mathematicians, statisticians, computer scientists, neuroscientists and psychologists in order to work towards a theoretical framework for neocortical function.

The beginnings of a framework were laid down more than 100 years ago by the German physicist and psychologist Hermann von Helmholtz, who characterized perception as a processes of "unconscious inference." Since then, the view that inferential processes are central to perception has been widely accepted by psychologists, as it has become clear that most of our percepts are filled-in or extrapolated far beyond what can be computed directly from available sensory data. Closely related to inference is the idea of prediction - the ability to forecast future outcomes based on information up to the present - which many have recognized as crucial to the survival of any organism. However, it is far less clear how inference and prediction are performed by neural circuits in the brain.

In recent years, a number of investigators have begun to focus on how the cortex does inference and prediction, with the goal of making concrete, testable theories of these perceptual and cognitive processes at the neuronal level. A central ingredient of both inference and prediction is the ability to store information about probabilistic relationships in the environment. Work on associative memory models has demonstrated how networks of neurons can accumulate such probabilistic information through changes in synaptic weights, but these models typically operate at a fairly abstract level of representation that does not deal with the complexities inherent in raw sensory information. Over the past decade, there has been increasing interest in the math and statistics communities in using graphical models to describe probabilistic relationships in complex datasets. These models make direct use of the principles of Bayesian inference, and since they can be instantiated as neural networks they provide a natural theoretical framework for thinking about inference and prediction in the brain. At the same time, in the realm of psychophysics and neurophysiology, experimentalists have begun to reveal direct evidence for Bayesian inference processes at work within the brain. A rich area for investigation is now opening up on both theoretical and experimental fronts.

The workshop will differ from typical conferences in some regards. Participants will be invited to suggest open problems and questions before the workshop begins, and these will be posted on the workshop website. These include specific problems on which there is hope of making some progress during the workshop, as well as more ambitious problems which may influence the future activity of the field. Lectures at the workshop will be focused on familiarizing the participants with the background material leading up to specific problems, and there will be ample time between talks for discussions and for work to be done.

Invited participants include A. Angelucci, D. Ballard, V. de Sa, K. Friston, S. Grossberg, J. Hawkins, J. Hirsch, D. Kersten, T.-S. Lee, Z. Li, K. Nakayama, B. Olshausen, R. von der Heydt, S. Zucker.

The deadline to apply for support for this workshop has passed.


Go to the American Institute of Mathematics.
Return to the AIM Research Conference Center.