Published in the Proceedings of the SIPAR Workshop on Parallel and Distributed Computing Gen eve Switzerland October Connectionist Quantization Functions

One of the main strengths of connectionist systems also known as neural networks is their massive parallelism However most neural networks are simulated on serial computers where the advantage of massive parallelism is lost For large and real world applications parallel hardware implementations are therefore essential Since a discretization or quantization of the neural network parameters is of great bene t for both analog and digital hardware implementa tions they are the focus of study in this paper In a successful weight discretization method was developed which is exible and produces networks with few discretization levels and without signi cant loss of performance However recent studies have shown that the chosen quantization function is not optimal In this paper new quantization functions are introduced and evaluated for improving the performance of this exible weight discretization method