A hardware learning scheme for feedforward neural networks

Abstract

A hardware learning scheme is proposed for implementation of FNN's (feedforward neural networks) with on-chip learning. The proposed learning scheme is quite different from the conventional hardware implementation but quite similar to the conventional software approach. The proposed learning scheme is implemented by inserting switching circuits between multi-layered feedforward circuitry and learning circuitry. The feedforward circuitry is implemented using nonlinear synapse circuits. And the learning circuitry is implemented by employing MEBP (Modified Error Back-Propagation) learning rule. The proposed scheme has been implemented using conventional CMOS technology and its operation has been verified using HSPICE circuit simulator The proposed learning scheme is very suitable for the future implementation of the large-scale neural networks with learning.