Accelerating a Machine Learning Algorithm on a Graphics Processing Unit

Abstract

Life long learning from zero(LL0) is a lifelong learning algorithm that has a dynamic neural network architecture. Many machine learning tools perform poorly on dynamic structures due to the overhead of growing computational maps with expanding networks. This thesis explores the possibility of delivering higher performance for the LL0 algorithm compared to the existing PyTorch implementation by developing a custom solution. This developed solution has a strongly coupled mapping of the LL0 algorithm with the GPU to achieve hardware acceleration. A set of benchmarks are defined to compare the performance of the between implementations. Furthermore, the thesis develops a methodology to investigate potential bottlenecks and parallelism with the implementation mapped to a GPU. The thesis achieves a significant speedup of ×19.48 on the number of feedforward per unit of time, compared with the similar PyTorch implementation, on an MNIST dataset.