Visualizing processes in neural networks

Abstract

A real-time visualization toolkit has been designed to study processes in neural network learning. To date, relatively little attention has been given to visualizing these complex, nonlinear systems. Two new visualization methods are introduced and then applied. One represents synaptic weight data as “bonds” of varying length embedded in the geometrical structure of a network. The other maps the temporal trajectory of the system in a multidimensional configuration space as a two-dimensional diagram. Two-dimensional graphics were found to be sufficient for representing dynamic neural processes. As an application, the visualization tools are linked to simulations of networks learning various Boolean functions. A multiwindow environment allows different aspects of the simulation to be viewed simultaneously using real-time animations. The visualization toolkit can be used in a number of ways: to see how solutions to a particular problem are obtained; to observe how different parameters affect learning dynamics; and to identify the decision stages of learning. A demonstration videotape is provided.