Deep learning model compression and acceleration for resource constrained devices

Abstract

In recent years, machine learning and deep learning have shown remarkable improvement in newlinecomputer vision, natural language processing, stock prediction, weather forecasting, and audio newlineprocessing, to name a few. The performance of a deep neural network (DNN) is dependent newlineupon a significant number of weight parameters that need to be trained, which is a computational newlinebottleneck. DNNs are also known for their high resource requirements, weight redundancy, newlineand large-scale parameters. Hence, the utilization of DNNs is restricted to devices where newlineadequate resources required to execute them are not available, especially resource-constrained newlinedevices such as mobile phones, wearables, and other edge devices. For various practical applications, newlinethe trained models should be deployed on resource-constrained devices. Hence, it becomes newlineimperative to compress and accelerate these models before deploying them on resourceconstrained newlinedevices while making the least compromise on the model accuracy. To address newlinethese challenges, in the last couple of years, many researchers have suggested different techniques newlinesuch as pruning, quantization, low-rank factorization, and knowledge distillation for newlinemodel compression and acceleration. newlinePruning has emerged as an essential technique to reduce unimportant parameters and improve newlinethe model performance. However, finding the best pruning candidates and an optimal newlinenumber of parameters that can be pruned without significantly affecting the model performance newlineis time-consuming and requires a lot of manual tuning. Therefore, this thesis proposes novel newlinemethods for identifying and pruning the less important parameters of the trained deep learning newlinemodel. newline