Accelerating Deep Convolutional Neural Network base on stochastic computing

Abstract

Deep Convolutional Neural Networks (DCNNs) are highly computational, and low budget platforms face many restrictions due to their implementation. Recently, Stochastic Computing (SC) demonstrated satisfying solution with simple and low power arithmetic units. In this paper, we present a highly efficient SC-based inference framework of DCNNs. We first propose a new Approximate Parallel Counter (APC), which is the most computational part in SC-DCNN. Our design relies on multiplexers to reduce the total number of input data. Second, a new mechanism is proposed for SC-based Rectified Linear Unit (ReLU) to make it more accurate and more similar to the actual ReLU function. Eventually, a new way is proposed for implementing soft-max function in SC domain with utilizing minimum resources. The simulation result shows that our platform for implementing LeNeT-5 achieves better area overhead, critical path length and power consumption than state-of-the-art works by term of slightly sacrificing accuracy.