Class of Research · Cohort 1
Active Cohort · Wrapping UpFrom Research Questions
to Submission-Ready Papers.
A 10-week elite research cohort (started August 1, 2026, and wrapping up now) led by Dr. Majid Riaz. Focused on EEG-based Brain-Computer Interfaces & Neural Decoding, Wearable Sensing, and Multimodal Biosignal Machine Learning — spanning biomedical signal processing, self-supervised representation learning, and explainable AI.
From problem framing and literature mapping to full manuscript defense
Split from Week 3 into dedicated teams of 6–7 researchers each
Direct weekly mentorship on methodology, signal analysis, and publication
Research Domains
Three Active Research Tracks
After two weeks of foundational cohort training, researchers divided into three specialized groups, tackling distinct frontiers in neural decoding and physiological computing.
EEG-Based Covert Speech Decoding
Non-Invasive Brain-Computer Interfaces (BCI)
Decoding covert and imagined speech directly from high-density non-invasive electroencephalography (EEG) signals using self-supervised representation learning and explainable AI.
Key Methodologies
Multimodal Wearable Biosignal Sensing
Physiological Computing & Digital Health
Wearable multimodal biosignal sensing across synchronized EEG, ECG, PPG, GSR/EDA, EMG, and respiratory sensors for real-time activity recognition, neuro-rehabilitation, and continuous health monitoring.
Key Methodologies
Cross-Domain Contrastive Learning for Time-Series
Self-Supervised Physiological AI
Self-supervised and contrastive representation learning for cross-domain physiological time-series, bridging heterogeneous EEG, cardiac, and wearable dynamics into shared latent representations.
Key Methodologies
Week-by-Week Execution
The 10-Week Research Progression
From zero to a submission-ready paper. Weeks 1–2 operated as a unified cohort, followed by 8 weeks of intensive group-level model development, neurophysiological validation, and manuscript writing.
Introduction: Research Grounding & Problem Formulation
What research actually is, formulating a real research question, live 'anatomy of a top-tier paper' walkthrough, and hands-on live EEG and wearable signal demonstration.
Grouping & Topic Selection: Datasets & Environments
Deep-dives into the three research tracks with real benchmark datasets, formal group formation, and environment standardization across Python, PyTorch, MNE-Python, and NeuroKit2.
Literature Review & Hypothesis Gap Mapping
Each group mapped 3–4 closely related anchor papers, systematically identified open experimental gaps, drafted formal mathematical hypotheses, and set baseline comparison metrics.
Data Preprocessing & Biosignal Pipeline Engineering
Artifact rejection (ICA, notch, and bandpass filtering), time-frequency spectral representations (wavelets, STFT), spatial windowing, and building reproducible PyTorch biosignal data loaders.
Baseline Modeling & Benchmark Architecture
Implemented classical signal processing baselines (Common Spatial Patterns, SVM, Random Forest) alongside modern deep learning backbones (EEGNet, Spatial-Temporal Conformer, 1D ResNets).
Advanced Representations & Self-Supervised Learning
Formulated domain-specific pretext tasks and contrastive learning objectives tailored for non-stationary physiological time-series, addressing inter-subject variance.
Explainable AI & Neurophysiological Attribution
Applied Integrated Gradients, Grad-CAM, and Layer-wise Relevance Propagation (LRP) to verify that learned neural activation patterns align with established neurophysiological brain regions.
Cross-Subject Synthesis & Statistical Rigor
Rigorous leave-one-subject-out cross-validation, exhaustive ablation experiments, paired statistical significance testing, and generating publication-quality figures and tables.
Manuscript Authoring & LaTeX Structuring
Drafted the full conference manuscript in LaTeX following IEEE / NeurIPS formatting: Abstract, Introduction, Related Work, Methodology, Empirical Results, and Discussion.
Paper Defense, Camera-Ready Prep & Open-Source Artifacts
Internal peer review defenses, finalizing camera-ready manuscripts, open-sourcing code repositories on GitHub, and uploading model weights and curated dataset cards to Hugging Face.
Manuscript & Artifacts
Submission-Ready Research Outputs
Self-Supervised Spatial-Temporal Contrastive Learning for Non-Invasive Covert Speech Decoding in EEG-BCI
Afrium Class of Research Cohort 1 · Mentored & Advised by Dr. Majid Riaz (2026)
Abstract Summary
We present a novel self-supervised representation learning framework designed to decode imagined (covert) phonemes from high-density non-invasive EEG signals across subjects. By leveraging spatial-temporal contrastive pretext objectives and domain-invariant latent projections, our model surpasses supervised baselines while demonstrating robust neurophysiological alignment verified through Layer-wise Relevance Propagation. All benchmark scripts, weights, and preprocessed datasets are released for open reproducibility.
Over 10 weeks, three research groups took a question
from zero to a submission-ready paper.
Selected undergraduate researchers trained under top faculty and PhD scholars, producing defensible empirical work and advancing non-invasive brain-computer interfaces and wearable machine learning.