Block-Based Neural Network High Speed Optimization

Abstract

Block-Based Neural Network (BBNN) consists of a 2-D array of memory-based modular component NNs with flexible structures and internal configuration that can be implemented in reconfigurable hardware such as a field programmable gate array (FPGA). The network structure and the weights are encoded in bit strings and globally optimized using the genetic operators. An asynchronous BBNN (ABBNN), which is a new model of BBNN, enables higher performance for BBNN by utilizing the parallel computation and the pipeline architecture. An ABBNN’s operating frequency is kept stable for all scales of the network, while conventional BBNN’s decreases accordingly. The architecture of ABBNN provides the capabilities to process and analyze high sample rate data at the same time. However, optimization by the genetic algorithm is a high-cost task, and the memory access is one of the causes which degrade the training performance. In this paper, we introduce a new algorithm to reduce the memory access in BBNN optimization. ABBNN, optimized with the proposed evolutionary algorithm, is applied to general classifiers to verify the effectiveness with regards to the reduction of memory access.