WEIGHT QUANTIZATION IN BOLTZMANN MACHINES

Hardware implementations of neural networks promise high computing power at moderate system complexity as compared to software simulations. Their design, however, must consider constraints imposed by currently available technology. Quantization of interconnection weights is the single most important such constraint. In this work, we describe simulations carried out for neural networks based on the Boltzmann Machine paradigm, on the impact of weight discretization, and choice of network architecture on performance. Our results show that this type of network is well-suited for operation with only a small number of weight levels.