Combining Machine Learning with Stochastic Computing

Abstract

Autonomous vehicles are limited by the size and power consumption of their onboard combinational circuits. This paper explores the growing research around novel computing techniques that reduce the cost of such circuits by fundamentally changing the way they compute. Specifically, this paper first gives a thorough, yet easy to understand, overview of neural networks. Neural networks are a common processing algorithm aboard such vehicles that are computationally expensive and thus take up a large amount of space and power. Next, stochastic computing (SC) is introduced as the novel computing technique that promises to substantially reduce the complexity of such circuits. Finally, a summary of the existing work about combining neural networks with SC is given with a few contributions such as using an analog random number generator.