Research outputs centered on efficient models and rigorous evaluation.

Larger Models Yield Better Results? Streamlined Severity Classification of ADHD-Related Concerns Using BERT-Based Knowledge Distillation

Ahmed Akib Jawad Karim, Kazi Hafiz Md. Asad, and Md. Golam Rabiul Alam

Introduced LastBERT, a 29M-parameter student distilled from BERT-base with a 3.8× reduction in parameter count. Evaluated across GLUE tasks and ADHD concern-severity classification, with formal analysis supporting the argument that scale is not always a prerequisite for task competence.

Strengthening False Information Propagation Detection: Leveraging SVM and Sophisticated Text Vectorization Techniques in Comparison to BERT

Ahmed Akib Jawad Karim, Kazi Hafiz Md. Asad, and Aznur Azam

Benchmarked BERT against SVM classifiers using TF-IDF, Word2Vec, and bag-of-words representations under resource-aware evaluation, demonstrating that carefully specified classical pipelines can approach transformer-level performance with substantially less computation.

When Does Higher-Order Representation Improve Protein-Complex Recovery? A Leakage-Controlled, Matched-k Comparison of Graph and Penalized Hypergraph Spectral Clustering

Kazi Hafiz Md. Asad, Rafi Majid, Md. Tanjeelur Rahman Labib, and Ahsanur Rahman

Designed a leakage-free benchmark across nine IntAct MITAB datasets with held-out evaluation, matched-k comparison, significance tests, null models, and Gene Ontology enrichment to characterize when higher-order incidence structure improves recovery over pairwise projection.

Beyond Clean Accuracy: A Test-Blind Benchmark for Optimizer Reliability, Robustness, Calibration, and Computational Efficiency

Kazi Hafiz Md. Asad

Develops a research-engineered, test-blind benchmark for evaluating neural network optimization algorithms beyond conventional clean-test accuracy. Across 12 tabular classification datasets, the study systematically examines optimizer stability, robustness to feature corruption and label noise, hyperparameter sensitivity, predictive calibration, loss-landscape geometry, and computational efficiency. The benchmark is designed to reveal when apparent optimizer gains remain reliable under realistic perturbations, controlled experimental variation, and resource constraints.

ARGUS-KD: Attribution-Rationale Grounded Unified Student Distillation for Explainable Code Vulnerability Detection

Kazi Hafiz Md. Asad

Introduces ARGUS-KD, an explainable knowledge-distillation framework for code vulnerability detection that transfers not only predictive behavior but also the teacher model's underlying rationale and security evidence. Commit patch diffs are leveraged to construct ground-truth vulnerability rationales, while a compact student model is trained with a dedicated rationale head and a security concept bottleneck. The framework is designed to deliver efficient vulnerability predictions together with interpretable token-level explanations during inference without requiring additional explanation-time computation.

Sustainable Intelligence: Environmental Trade-offs of Cloud-Based Machine Learning

Examines energy, carbon, water, and hardware-lifecycle impacts across training, inference, and data-center operation, with attention to inconsistent measurement and reporting practices and mitigation strategies such as carbon-aware scheduling and algorithmic efficiency.

Naturalistic Behavioral Prognosis in Autistic Children: Longitudinal Evidence, Digital Phenotyping, and a Roadmap for Predicting Language and Adaptive Functioning

Synthesizes longitudinal evidence from naturalistic audio, video, wearable, and behavioral data, identifying limitations in external validation, uncertainty quantification, fairness, clinical utility, and cross-population generalization.