Supervisor

Dr. David McQuaid

Programme

MSc in Data Analytics

Subject

Computer Science

Abstract

This study examines the effects of three Deep Neural Network compression techniques—Quantisation, Pruning, and Weight Sharing/Clustering—on CNN and ANN models trained for image classification tasks. The models were tested on the CIFAR-10 dataset for multiclass classification and a binary classification task using a dataset derived from COCO. The best validation accuracy achieved was 74.7% with a CNN on CIFAR-10 and 53% with the best ANN. On the COCO dataset, a modified CIFAR-10 CNN model achieved 75%. The models were compressed using the three techniques and benchmarked on a ThinkPad laptop and Raspberry Pi 3B+ based on metrics relevant for resource-constrained Internet of Things (IoT) applications, including accuracy, energy consumption, inference speed, and model file size.

The results revealed that weight sharing increased model file size and reduced throughput by up to 26x but significantly lowered energy consumption by 15.6%. Pruning and Quantisation preserved CNN accuracy while reducing model size by up to 81%, although quantisation increased inference time by an average of 3.5x on the Raspberry Pi. The study highlights how these compression techniques affect model performance and offers insights for deploying deep learning models in resource-limited environments.

Date of Award

2024

Full Publication Date

2024

Access Rights

open access

Document Type

Capstone Project

Resource Type

thesis

Share

COinS