Implementation of stochastic computing in activation functions using stochastic arithmetic components

Abstract

A new computing method using stochastic-based numbers is gaining importance as an approximate computing method to save area, energy, and computation time based on the accuracy required. This works uses stochastic computing, which is suitable for enhancing the efficiency of neural network. Herein we focus on developing activation functions that are essential parameters in the design of neural networks. The activation function in stochastic computing is typically a threshold function that maps the input bits to a binary output based on a probability distribution. This paper presents the development of modified activation functions tanh and COS using SC-based arithmetic components. Two different types of stochastic number generators (SNGs) have been used. Error analysis has been done based on the computation using two SNGs. Also, accuracy measurement is performed using error analysis for these complex functions mentioned above.