FPGA implementation of a blind source separation system based on stochastic computing

Abstract

We have constructed a blind source separation system based on stochastic computing techniques, and have implemented it using an FPGA. In stochastic computing, analog quantities are represented by pulse sequences. The advantage of this method is its simple circuitry. For this reason, stochastic computing systems have been applied to massive circuits such as artificial neural networks. Blind source separation systems are a growing focus of interest. These are systems that infer source signals from mixed signals received by sensors. A blind source system using a neural network model has recently been proposed. However, it is difficult to implement this system in actual circuits, since there are cases in which the values of synaptic weights fall outside the range within which the hardware can process them correctly. Therefore, we propose a blind source separation system that can be implemented in hardware, namely a system in which the values of synaptic weights can be kept within the range that permits the hardware to process them. We then constructed this system based on stochastic computing and investigated it using functional simulations. Finally, we implemented our system on an FPGA board, where we succeeded in separating source signals from mixed signals.