Rajesh Godasu

Date of Award

Spring 3-2024

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Information Systems (PhDIS)

First Advisor

David Zeng

Second Advisor

Ronghua Shan

Third Advisor

Renae Spohn

Fourth Advisor

Julie Wulf Plimpton

Abstract

Transfer learning has emerged as a pivotal technique in deep learning, allowing pre-trained models to be fine-tuned for novel tasks. This has often led to enhanced performance and reduced training time. Lightweight architectures known for their efficiency and speed, without compromising accuracy complement this technique. These two techniques combined address the issues of limited availability of training data and the requirement of high computational resources. Numerous researchers delved into these methods to address the challenges posed by the Covid-19 pandemic. Within this framework, our study sought to evaluate the multi-stage transfer learning method across several dataset sizes and truncated versions of the lightweight MobileNetV2 architecture for medical image classification. Initially, an exhaustive literature review was conducted to present the latest advancements in transfer learning and lightweight architecture applications for COVID-19 image classification. Additionally, this study explores the trade-offs between finetuning and dataset size in a multi-stage transfer learning system using a lightweight architecture. Lastly, in this study, we aim to analyze the performance of truncated lightweight models against different training dataset sizes in a multi-stage transfer learning framework. The results of this study significantly enhance the knowledge in this specialized research area. The most popular lightweight architectures were identified, and we found that standard CNNs can be truncated without losing performance. The study also established that mid-sized datasets and freezing 90 to 95 layers yield the best performance on the target task. Furthermore, the study highlights a key trade-off in stage one transfer learning: smaller datasets require more extensive layer training, while larger datasets need fewer re-trained layers for optimal performance. Also, beyond a certain dataset size, deeper fine-tuning does not lead to improved accuracy. Finally, we established that complex models tolerate more frozen layers, maintaining adequate learning capacity. The findings support the idea that while more complex models demonstrate higher accuracy, simpler models perform competitively. Moreover, the impact of model complexity is reduced on target performance with small-size datasets. These insights underscore the efficiency of multi-stage transfer learning with lightweight models, especially when pre-trained models are applied to new datasets. The present research adds to the improvement of the effectiveness of medical expert systems while also reducing the burden on healthcare professionals.

Download

Share

COinS