event
Tuesday 08 May 2012: Dynamical singularities in online learning and adaptive delayed-feedback control
Asaki Saito - Future University Hakodate and JST PRESTO, Japan
Harrison 209 14:00-15:00
The ability to adapt, as by learning, is indispensable for the survival of living things, and can also be critical in other real systems, e.g., ecological, economic, and social systems. To understand highly dynamic phenomena observed in the systems, it will be necessary to analyze the complex dynamics of adaptive processes themselves, instead of focusing on efficiency, stability, or final results. In this talk, I will introduce two studies on characterization of adaptive processes as dynamical system processes. The first one deals with an online learning system of a recurrent neural network, and the second one treats an adaptive delayed-feedback control system, where these systems are chosen as two of the simplest artificial adaptive systems. I will show that these systems exhibit singular behavior (``neutral behavior'') different from ordinary chaos, such as that exemplified by zero Lyapunov exponents. Also, I will discuss some mechanisms explaining the neutral behavior observed in these systems.
References:
A. Saito, M. Taiji, and T. Ikegami, Phys. Rev. Lett. 93, 168101 (2004).
A. Saito and K. Konishi, Phys. Rev. E 84, 031902 (2011).
