ABSTRACT
We describe an experimental neural network learning system that is designed to rapidly prototype telecommunications applications. It functions as a co-processor in a Sun4 workstation through the VME bus interface and is controlled by a friendly, graphical user interface. The experimental prototype system uses the cascadable analog learning chips that have been described in a previous publication [1]. A four-chip network consisting of 128 neurons and about 2000 adaptive synapses has been successfully tested. It can be used for 1) classification problems by use of feedback or feed-forward neural networks, 2) optimization problems by use of mean-field and Boltzmann settling, or 3) problems that require high speed learning. As an example of a problem that requires continuous, high speed adaptation, the system has been used to prototype a simple channel equalization problem. The experimental system is functioning as a general purpose test bed for running neural network applications.
