Using On-Line Arithmetic and Reconfiguration for Neuroprocessor Implementations

Abs t rac t . Artificial neural networks can solve complex problems such as time Series prediction, handwritten pattern recognition or speech processing. Though software simulations are essential when one sets about to study a new algorithm, they cannot always fulfill real-time criteria required by some practical applications. Consequently, hardware implementations are of crucial import. The appearance of fast reconfigurable FPGA circuits brings about new paths for the design of neuroprocessors. A learning algorithm is divided into different steps that are associated with specific FPGA configurations. The training process then consists of alternating computing and reconfiguration stages. Such a method leads to an optimal use of hardware resources. This paradigm is applied to the design of a neuroprocessor implementing multilayer perceptrons with on-chip training and pruning. All arithmetic operations are carried out with on-line operators. We also describe the principles of the hardware architecture, focusing in particular on the pruning mechanisms.