Implementation of Neural Network in OpenCL using Intel FPGA SDK

Abstract

Today artificial intelligence techniques, such as deep learning and machine learning are becoming more popular, and has wide usage in many industrial as well as research domains. Special interest among Deep Learning architecture is in Convolutional Neural Networks (CNN). CNNs have been broadly utilized in numerous applications, for example, image classification, video examination and voice recognition. In CNN, many convolution and fully-connected layers requesting extensive measure of correspondence for parallel calculation, incredible amounts of data and computation power, which isn’t easily met by conventional computing platforms. As well as in data mining applications kNN is one of the useful algorithms for classification. The kNN utilized in applications like 3D object rendering, content-based image recovery, statistics, gene classification and so on. Sequential approach for these applications cost large computation time, especially in high dimensional spaces. This bottleneck has made the need of the parallel kNN on commodity hardware. GPU/GPGPU introduces parallel processing to industry long ago. However it is more energy eater and not affordable for a large server farm due to its high cost. The utilization of Field Programmable Gate Arrays(FPGA) gives a fascinating alternative. Newly developed design tools for FPGAs have made devlopers more compatible with the high-level software practices and making FPGAs more accessible to them. Altera Xilinx and many more of them have adopted OpenCL co-design framework for FPGA designs. OpenCL architecture is the ease elite arrangements for parallelising. In proposed work, CNN and kNN have been implemented using OpenCL. Prepared CNN model for working with 3D large sized CT-scan images, is been tested using Arria10 board. A kNN model is functioning as a classifier of Iris database. Performance of kNN model is been evaluated using CycloneV SoC board, which gave 40-60 percent improvement in execution time. In this thesis, CNN and kNN models and their development in OpenCL, temporal performance and hardware resources are discussed.