Multilayer perceptrons on Splash 2

Abstract

Multilayer perceptrons (MLPs) are one of the most popular neural network models for solving pattern classification and image classification problems. Because of their ability to learn complex decision boundaries, MLPs are used in many practical computer vision applications involving classification (or supervised segmentation). Once the connection weights in a MLP have been learnt, the network can be used repeatedly for classification of new input patterns. Several special-purpose architectures have been described in the literature for neural networks as they are slow on a conventional uniprocessor. In this paper, we describe mapping of MLPs onto Splash 2 - a 'custom computing machine'. The main features of the proposed mapping are: (i) the number of nodes in a layer is not fixed; (ii) the number of layers in the network is not fixed; (ii) it is based on a set of reprogrammable FPGAs and a programmable crossbar; and (iv) it has a significant speedup over a uniprocessor. The mapping has been used for implementing a 3-layer MLP for page segmentation application with an appreciable speedup of approximately 150 over a SPARCstation 20 for one million pattern vectors with 20 features per pattern.