This paper describes the authors' recent development work involving the use of EPROM-based microcontrollers for implementation of artificial neural networks. The microcontrollers used are selected from the PIC family of devices, which are 8-bit devices employing a reduced instruction set computer (RISC) and Harvard architectures. The primary motivation for this work is to develop implementations of small neural networks which are simple to understand and experiment with, enabling them to be used as aids in the undergraduate teaching of neural networks and in demonstrations of their basic principles. Practical issues are addressed and results are presented for implementations of a single neuron and a small feedforward neural network. In each case on chip training is incorporated using the delta rule and the error backpropagation algorithm respectively. Proposals for hardware implementations of larger networks are included.
展开▼
