Alumni
2024
Joy Wairimu Ndirangu — Honours, Computer Science — September 2024
Thesis: Enhancing Interpretability in Deep Learning Models for Breast Cancer Detection
Abstract: This thesis applies two CNN architectures — AlexNet and Houby & Yassin's CNN — to classify mammographic images from the KAU-BCMD dataset into BI-RAD categories, addressing the "black-box" problem in deep learning diagnostics using SHAP-based explainability techniques. The models achieved strong performance, with AUC scores of 0.925 and 0.932, and F1-scores of 76.7% and 79.4%, respectively. Meanwhile, SHAP analysis highlighted the key image regions driving each prediction, helping to bridge the trust gap between deep learning systems and radiologists.
The link to the paper: Click Here
Errol DaRocha — Honours, Computer Science — December 2024
Thesis: A Comparison of Autoencoders and Variational Autoencoders for Anomaly Detection in Dermoscopic Images
Abstract: This thesis compares Autoencoder (AE) and Variational Autoencoder (VAE) architectures for unsupervised anomaly detection in dermoscopic skin images using the HAM10000 dataset of 10,015 images. Both models were trained on normal (benign) data and evaluated on their ability to distinguish benign from malignant lesions via reconstruction error. Averaged across 30 training runs, the AE outperformed the VAE in accuracy (71.84% vs. 67.79%), F1-score (73.03% vs. 67.33%), and False Negative Rate (23.76% vs. 33.58%), suggesting AE architectures as a promising and scalable approach for automated early skin cancer detection.
The link to the paper: Click Here
Joseph O'Neill — Master of Science, Computer Science — Fall 2024
Thesis: A Fully Secure Approach to Privacy-Preserving Machine Learning for Satellite Image Classification
Abstract: This thesis proposes a fully secure privacy-preserving machine learning system for satellite image classification by combining the CKKS homomorphic encryption scheme with Support Vector Machines (SVMs). Unlike existing approaches that sacrifice full security for accuracy, this system encrypts both the images and the ML model using 128-bit AES equivalent encryption, enabling ciphertext-to-ciphertext predictions in the cloud without ever exposing raw data. Preliminary results demonstrate classification at roughly 30,000 images per hour with 87% accuracy — matching the unencrypted baseline — confirming that fully secure image classification is achievable without loss of performance.
The link to the paper: Click Here
2025
Chenuka Gamage — Honours, Computer Science — Winter 2025
Thesis: Early Diagnosis of Alzheimer's Disease Using 3D Convolutional Neural Networks
Abstract: This thesis presents a 3D Convolutional Neural Network (3D CNN) approach for the automated early detection of Alzheimer's disease using FDG-PET scans from the ADNI dataset. After preprocessing the scans into standardized 64×64×64 voxel volumes, the model was evaluated across ten independent test runs, achieving an average accuracy of 71.25% and a best run accuracy of 78.75%. These results demonstrate the potential of 3D CNNs as a foundation for deep learning-based diagnostic tools in resource-limited medical settings.
The link to the paper: Click Here
Hamza Ben Alla — Honours, Computer Science — April 2025
Thesis: Enhancing Early Alzheimer's Diagnosis with Interpretable 3D CNN Models Using MRI Scans
Abstract: This thesis presents a 3D CNN model built from scratch to classify stages of Alzheimer's disease using MRI scans from the ADNI dataset, in both binary and multi-class settings. Through a carefully designed preprocessing pipeline and optimized hyperparameter tuning, the model achieved 94.37% accuracy in multi-class classification (AD, MCI, and CN) and 91.79% in binary classification (AD vs. CN), outperforming several state-of-the-art models. Interpretability was added using an occlusion-based technique, which confirmed the model's focus on clinically relevant brain regions including the hippocampus, amygdala, and thalamus.
The link to the paper: Click Here
Excellence Sowunmi — Master of Science, Computer Science — Spring 2025
Thesis: OLS with Internal Subsymbolic Learning and Reasoning Using Cross-Modality Autoencoders
Abstract: This thesis introduces a cross-modal autoencoder architecture that bridges the gap between neural and symbolic AI by learning bidirectional transformations between images and text. Built on the neuro-symbolic framework of Silver and Mitchell, the model distinguishes between internal subsymbolic (conrep) and external symbolic (symrep) representations, enabling both perceptual understanding and abstract reasoning across modalities. Using a shared reasoning layer, selective input masking, and curriculum-based training, the system achieves stable and accurate cross-modal reconstruction even when one modality is masked — with applications in computer vision, natural language processing, and multimodal reasoning.
The link to the paper: Click Here
Hamza Blagui — Mitacs Intern, Summer 2025
His study built an embedded system that performs image classification entirely on hardware, using a NeuroShield neuromorphic chip mounted on a PYNQ-Z2 board to train and run inference on the MNIST dataset without relying on a separate host computer. The design achieved 95.17% accuracy while using roughly half the hardware neurons of the reference benchmark, and processed images in real time at 134 images per second with power consumption under 2 watts. These results demonstrate that compact neuromorphic hardware can deliver fast, accurate, and energy-efficient intelligence suited for edge computing applications.
Hassen Gadacha — Mitacs Intern, Summer 2025
His work presents an explainable deep learning approach for detecting early-stage Alzheimer's disease from MRI scans, combining 3D convolutional neural network classification with SHAP-based interpretability methods. By pairing these two approaches, the study demonstrates that MRI analysis can be both accurate and transparent, with model predictions clearly explainable rather than functioning as a black box. The findings also emphasize that consistent preprocessing is essential for generating reliable explanations in medical AI applications.
Kylie DeViller — Honours, Computer Science — December 2025
Thesis: From Seed to Algorithm: A Comparative Analysis of Genome-wide Association Study and Random Forest for Genomic Trait Prediction in Apple
Abstract: This thesis applies two computational approaches — a multi-locus mixed model GWAS and a Random Forest algorithm — to analyze approximately 19 million SNPs from 504 apple accessions in Canada's Apple Biodiversity Collection, with the goal of identifying genetic markers linked to physical traits such as shape and size. GWAS successfully identified one SNP significantly associated with circularity, while Random Forest generalized poorly and yielded no significant markers. The findings highlight the current limitations of machine learning for complex genomic trait discovery and point toward future exploration of alternative algorithms to better support apple breeders in incorporating genomic insights into cultivar development.
The link to the paper: Click Here
2026
Mohammod Ehsanul Haque — Honours, Computer Science — April 2026
Thesis: A SMOTE-Based Approach for Minority Attack Detection in Intrusion Detection Systems
Abstract: This thesis tackles the challenge of class imbalance in intrusion detection by proposing a CNN–BiLSTM hybrid model trained on the UNSW-NB15 dataset, which contains over 2.5 million network flow records with an imbalance ratio exceeding 12,000:1. While a baseline model achieved 98% overall accuracy, its Macro F1-score of only 0.5081 revealed near-zero detection of rare attack types. Applying categorical-aware SMOTE-NC oversampling with 20,000 samples per minority class improved the Macro F1-score to 0.5307, and further hyperparameter tuning boosted Worms detection from 8.11% to 75% — the largest single-class gain in the study — while maintaining 99.14% binary classification accuracy.
The link to the paper: Click Here
Ibrahim Harunur Rashid — Honours, Computer Science — April 2026
Thesis: Deep Learning Approaches for Sensor-Based Human Activity Recognition
Abstract: This thesis investigates deep learning techniques for smartphone accelerometer-based Human Activity Recognition (HAR), comparing MLP, CNN, and CNN-LSTM architectures across the UCI-HAR, WISDM, and HHAR datasets. The CNN achieved the highest in-domain accuracy at 96.2%, while the CNN-LSTM showed the most consistent cross-validation results. However, when trained on UCI-HAR and tested on HHAR, performance dropped sharply to 17%, highlighting significant challenges in cross-dataset generalization due to device and domain mismatch. Data augmentation provided modest improvements but did not resolve the gap. SHAP interpretability analysis revealed that the X-axis accelerometer channel and transient motion segments contribute most to predictions, confirming the models learned physically meaningful patterns.
The link to the paper: Click Here
Brendan Toscano — Honours, Computer Science — April 2026
Thesis: Machine Learning Evaluation of Generative AI Debugging Feedback Plausibility for Introductory Python Programming
Abstract: This thesis investigates whether the plausibility of LLM-generated debugging feedback can be estimated using computable features derived from the IntroPyNUS dataset. After cleaning and processing 10,548 records of buggy student submissions, the study defined features for repair correctness, feedback relevance, hint concealment, and repair minimality, then trained binary and multi-class ML evaluators on four Python problems and tested them on a held-out fifth. The best binary evaluator — a five-feature Linear SVM — achieved 92% accuracy and an F1 score of 0.76, while a three-feature Random Forest reached a macro F1 of 0.876 without access to correctness labels. Applying these evaluators to four modern LLMs (GPT-5.2, GPT-4.1, GPT-4-Turbo, and Claude Opus 4.6) revealed that passing unit tests is not enough to judge tutoring quality: GPT-5.2 achieved the highest repair pass rate at 99% but the lowest plausibility rate, while Claude Opus 4.6 produced the most minimal and educationally grounded repairs.
The link to the paper: Click Here