PhD Oral Exam - Elie Neghawi, Electrical and Computer Engineering

When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.

Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.

Abstract

The rapid evolution of Convolutional Neural Networks (CNNs) has produced increasingly efficient and versatile algorithms, but the factors driving their superior performance remain underexplored. While previous research has primarily focused on explaining Semi-Supervised Machine Learning (SSML) algorithms in a model-specific manner, this thesis aims to generalize those findings, making them applicable across a wider range of CNNs. The challenge lies in achieving a method that can both enhance performance and improve interpretability while remaining adaptable to various models.

This thesis introduces a post-hoc Explainable Artificial Intelligence (XAI) method, called Unified Model Agnostic Computation (UMAC), designed to generalize common components of CNN models by drawing insights from SSML and Self-Supervised Learning (SSL) algorithms. Our research begins by focusing on two primary aspects: (1) the effect of parameter updates during training on both labeled and unlabeled data in SSML and SSL, and (2) the transition from model-specific SSML frameworks to a more generalized, model-agnostic approach using SSL.

In the first phase, we used SSML as a foundation, breaking down their components into preprocess-centric and classifier-centric elements, which led to the creation of Semi-Supervised Computation Processes (SSCPs). These processes were tested across five state-of-the-art SSML algorithms and three SSL algorithms, using various Deep Neural Networks (DNNs). Although this phase acted as a testing ground to understand the mechanics of SSML, it allowed us to identify key drivers of performance, especially in relation to parameter updates and data handling.

Through 45 rigorous experiments, we observed an 8% reduction in training loss and a 6.75% increase in learning precision using the Shake-Shake26 classifier with the RemixMatch SSML algorithm. A key observation was the positive correlation between labeled data and training time, showcasing the importance of label quantity in enhancing model efficiency.

In the second phase, we transitioned from SSML to SSL to prove that the methodology could be generalized to a model-agnostic approach. By integrating SSL components, we aimed to develop a unified framework that worked across various DNN architectures. Building upon this analysis, we developed a UMAC process for SSL, tailored to complement modern self-supervised learning algorithms. UMAC serves as a model-agnostic XAI methodology that explains models by composition, systematically integrating and enhancing state-of-the-art algorithms. Through UMAC, we identified key computational mechanisms and crafted a unified framework for self-supervised learning evaluation. Our systematic approach yielded a 17.12% improvement in training time complexity and a 13.1% boost in testing time complexity, with notable improvements observed in augmentation, encoder architecture, and auxiliary components within the network classifier. This phase demonstrated that UMAC could enhance accuracy and reduce training loss under different data conditions, showing its adaptability to different models and datasets.

In the third phase, we applied the UMAC framework to the field of medical image classification. Medical imaging tasks often suffer from data scarcity, making it challenging to achieve both high performance and model interpretability. By leveraging the UMAC methodology, we integrated it into CNNs and Transformers to generate high-quality representations, even with limited data. Experiments across five 2D medical image datasets showed that UMAC outperformed traditional augmentation methods by 1.89% in classification accuracy. Additionally, incorporating XAI techniques ensured that the models provided transparent and reliable decision-making processes, enhancing their interpretability in critical medical applications.

Throughout this process, UMAC served as an XAI method based on explaining models by composition, systematically breaking down computational processes to reveal how model components contribute to overall performance. This approach enabled us to create a unified, model-agnostic framework that enhanced both transparency and efficiency in CNNs.

Ultimately, this thesis contributes a structured and generalizable approach for machine learning (ML) developers, offering step-by-step guidelines to improve model performance and interpretability. By generalizing the computation processes of SSML and SSL through the UMAC framework, we provide developers with the tools needed to optimize their models across various domains, particularly in fields where transparency and accuracy are critical.