Algorithmic mapping of neural networks with multi-activation product units onto SIMD machines

Abstract

A modification to the algorithmic mapping algorithm for neural network models proposed by W. Lin et al. (1991) is presented. The modified algorithm can accommodate a larger class of network models recently proposed. The new neural network model uses vectorial interconnections between neurons and multiactivation product units. The generalized delta rule for the Rumelhart-Hinton-Williams neural networks can still be used with appropriate enhancement. The implementation of the model is targeted for fine-grain mesh-connected SIMD machines. The basic routing procedures are similar to those in the algorithmic mapping algorithm but with more flexibility in specifying the size of the data to be shifted between processors.<>