FPGA Based Accelerators of Deep Learning Networks for Learning and Classification

Abstract

A recent trend is to execute computationally intensive algorithms (or work flows) for business analytics using cloud environments which provide machine learning hardware support in the form of GPUs and TPUs. Businesses obtain their data at the sensor level and then perform algorithmic operations on the data via these cloud services. As a result, there can be high input/output data latency which tends to slow down productivity. This thesis work will explore the topic of executing computationally complex algorithms, such as the Convolutional Neural Network (CNN), Recurrent Neural Network(RNN) and Spike Neural Network (SNN), at the sensor level through the use of FPGAs (Field Programmable Gate Arrays) as an alternative to cloud-bound GPU and TPU services.