EVENT DETAILS AND ABSTRACT

Title: The Dynamics of Incremental Learning in Multilayered Neural Networks
Speaker: Professor Sara A. Solla
Speaker Info: Northwestern University
Brief Description:
Special Note:
Abstract:

Neural network models consist of large arrays of highly interconnected nonlinear units that mimic the behavior of biological neurons. These models provide a prototype to investigate a system that interacts with the environment through the execution of an action that generates a feedback error signal from the outside world; the information provided by this signal is used by the system to modify its internal state. Neural networks thus exhibit learning and adaptation capabilities that amplify their importance both as computational devices and as neurobiological models. We present an analytic solution for the problem of incremental gradient-descent learning of smooth input-output maps in multilayer neural networks. The approach reveals universal features of learning from examples in a variety of realizable, unrealizable, and overrealizable learning scenarios.