Statistical Mechanics Approach To Interactive Learning

The Department of Physics and Astronomy Colloquium: Prof. Susanne Still, ICS, UH Manoa, to speak on "Statistical Mechanics Approach To Interactive Learning."

Abstract:
The principles of statistical mechanics and information theory play an important role in learning theory and have been widely studied. The new aspect in my work is a focus on integrating feedback from the observer. A quantitative approach to interactive learning and adaptive behavior is proposed, integrating model- and decision-making into one theoretical framework. I start by asking two simple questions: given a time series, what is the class of models that one can construct from the observed (past) data which are maximally predictive at fixed model complexity? Since the observed data depends on the actions of the observer, we also need to know what class of action policies will give rise to data sets that allow for constructing a maximally predictive model. Predictiveness is measured by the information captured about future data, while complexity is measured by the coding rate. This naturally leads to an objective function that reflects the trade-off between predictive power and complexity. As a family of solutions, one finds Gibbs distributions, in which the trade-off parameter plays the role of a temperature: at high temperatures, efficient compression is paramount and predictive information is lost. In the low temperature regime the optimal model reflects the underlying process's causal organization exactly: sufficient statistics are retrieved by discovering the causal state partition of the observed data. A fundamental consequence of the feedback between the observer and the environment is that the optimal action policy balances exploration and control. Time permitting, I will discuss some simple examples which can be solved analytically and how reward maximization can be integrated into this theory.

Ticket Information
Free

Event Sponsor
Department of Physics and Astronomy, Manoa

More Information
Dr. John G. Learned, 956-2964, jgl@phys.hawaii.edu

Ongoing Events