FPGA Architecture To Enhance Hardware Acceleration for Machine Learning Applications

Abstract

Many algorithms have been developed in the field of Machine learning and its sub-fields such as neural networks, Deep learning and so on, for applications such as pattern classification, image and video processing, statistical data mining and so forth. These algorithms perform such tasks with remarkable accuracy. However, when implemented in traditional processor core software, many of these algorithms get choked when the size of the network or computational demand of the algorithm scales up. FPGAs and ASICs offer higher computation capability, throughput and bandwidth. Features such as massive parallel processing, high performance and reliability, are strong arguments for FPGAs for Deep Neural Network applications. FPGAs offer reconfigurable, flexible architectures where even the most popular GPUs fall short. This paper proposes a robust, re-configurable architecture for deploying Machine learning algorithms and presents its advantages by implementing a Neural Network in an FPGA and comparing its results with an implementation in a Raspberry Pi.