High Performance Multilayer Perceptron on a CustomComputing

Multilayer perceptrons (MLPs) are one of the most popular neural network models for solving pattern classiication and image classiication problems. Because of their ability to learn complex decision boundaries, MLPs are used in many practical computer vision applications involving classiication (or supervised segmentation). Once the connection weights in a MLP have been learnt, the network can be used repeatedly for classiication of input test 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, a high performance architecture for MLPs is presented using a novel class of hardware known as \custom computing machines". The main features of the proposed architecture are: (i) the number of nodes in a layer is not xed; (ii) the number of layers in the network is not xed; (iii) it is based on a set of reprogrammable FPGAs and a programmable crossbar; and (iv) it has a signiicant speedup over a uniprocessor. The system has been used for a page segmentation application with an appreciable speedup of approximately 150 over a SPARCstation 20 for one million pattern vectors with 20 features per pattern.