Divide2Conquer (D2C) & Dynamic Uncertainty Aware Divide2Conquer (DUA-D2C): Overfitting Remediation in Deep Learning Using a Decentralized Approach

Supervisor: Prof. Dr. Golam Rabiul Alam

Part 1 (Divide2Conquer): Introduced the Divide2Conquer (D2C) method, which partitions training data to train identical models independently on each subset, aggregating parameters periodically to limit the impact of individual outliers/noise and mitigate overfitting. Published at IEEE BigData 2024.

Part 2 (DUA-D2C): Extended D2C to develop Dynamic Uncertainty-Aware D2C (DUA-D2C), an enhancement that dynamically weights edge models based on performance and prediction uncertainty, further improving robustness. Currently under review.

Medical Image Segmentation & Classification based on a Dual-Domain Approach of Spatial and Spectral Feature Fusion

Supervisor: Prof. Dr. Mohammed Imamul Hassan Bhuiyan

Part 1 (S³F-Net): Proposed S³F-Net, a dual-branch framework that learns from both spatial (CNN) and spectral (SpectraNet) domains simultaneously, achieving performance improvements over unimodal baselines & state-of-the-art competitive accuracy on multiple medical imaging datasets.

Part 2 (Spatio-Spectral Medical Image Segmentation): Currently extending the S³F-Net architecture to perform robust and accurate segmentation tasks on medical images, leveraging its multi-domain feature learning capabilities.

Bioradiolocation-Based Multi-Class Sleep Stage Classification

Supervisor: Prof. Dr. Mohammed Imamul Hassan Bhuiyan

This undergraduate thesis focused on a non-invasive sleep stage classification method using bioradiolocation signals. By extensively extracting time & frequency domain features and employing a Random Forest classifier, the system outperformed state-of-the-art methods in terms of several metrics. Published at ICECE 2022.

Machine Learning & Statistical Modeling Driven Repression and Mobilization Analysis of Bangladesh's Social Movements

This project applies ML and statistical modeling to analyze the dynamics of state repression and citizen mobilization during Bangladesh's recent social movements. One paper from this work focusing on July Revolution is already under review.

Multi-Modal Fusion of Different Representations of the Signal for Biomedical Audio Analysis

This project explores fusion of features learned from different representations of biomedical audio signals, to exploit the complementary strengths of each representation. We explored different representations including waveforms, spectrograms, and scalograms for robust diagnostic models, and one paper from this research is currently under review.

Antimicrobial Resistance (AMR) Prediction from Genomic Data (SNPs) Leveraging Deep Learning & Explainable AI

This research exploits different perspectives provided by different types of models. We explored the utility of the sequential information available in SNPs through sequence-aware models and currently investigating fusion mechanisms to combine them with "bag-of features" models. We also incorporate Explainable-AI methods to biologically validate our models, by mapping the SNP positions to the corresponding genes and investigating their relevance in AMR. One paper from this research has been accepted at SCA/HPCAsia 2026.