Research · B.Sc. thesis · Oct 2025 – Jun 2026

Graph-conditioned attention for task-fMRI classification: from supervised learning to self-supervised GC-JEPA

Authors

Shakib Shadman Shoumik, MD Imtiaj Ahmed, A. K. M Foyshal Sarkar, Mahir Ashhab Adeeb

Supervision

Prof. Chowdhury Mofizur Rahman (supervisor), Zayed Humayun (co-supervisor), Department of CSE, BRAC University

My part

Adding the connectivity bias to attention, analysing relationships between brain regions, hyperparameter tuning and fine-tuning experiments, and typesetting the thesis in LaTeX

Status

Thesis accepted, Summer 2026

Code

github.com/shoumikarc/jepa-research (team repository, maintained by the first author)

In one paragraph

We built a model that reads brain-scan data and recognises which of seven mental tasks a person was doing, such as language, movement or working memory. The key idea is a small change that lets the transformer use the known pattern of which brain regions work together when it decides what to focus on. On the NeuroGraph benchmark it reached 97.7% accuracy, above the 96.25% reported by a 2025 preprint on the same benchmark (which we did not re-run on our split). A version that first learns from unlabelled data reached 96.7%.

Method

Attention = softmax(QKᵀ / √d_k + α_ℓ · G) · V

• G is each subject's 400 × 400 functional-connectivity matrix: Pearson correlations between the BOLD time series of 400 Schaefer brain regions. Its rows are also the token features.
• α_ℓ is one learnable scalar per layer, shared across heads. The whole mechanism adds 4 parameters to a 4-layer, 128-dimensional pre-LayerNorm transformer.
• GC-JEPA adds a self-supervised stage before fine-tuning: the encoder learns to predict connectivity values between sampled region pairs, then is fine-tuned for classification.

Data

Human Connectome Project task-fMRI through the NeuroGraph benchmark: 7,443 graphs, 7 task classes (emotion, gambling, language, motor, relational, social, working memory), stratified 70/10/20 split into 5,210 / 744 / 1,489 graphs.

Results

• All three Graph-only seeds (97.58, 97.78, 97.65) are above the reference, a margin of about 21 of 1,489 test graphs.
• Learned α stays positive in every GC-JEPA layer, while JEPA-only keeps it near zero.
• Label efficiency: all three variants reach about 91–92% with 10% of the labels and about 95% with 25%. They are statistically tied, so on this clean benchmark pretraining gave no significant label-efficiency gain.
• Checkpoints were chosen on validation accuracy, with seeded runs in cuDNN deterministic mode on a single RTX 4070 Ti SUPER.

Limitations we reported

• The reference number is taken from its paper, not re-run on our split; there are only 3 seeds on one graph-level split, so generalisation to unseen subjects is untested.
• There is no supervised control with the graph bias switched off under the same schedule, so the Graph-only gain is not fully isolated.
• The self-supervised margin over JEPA-only is small and did not replicate in an independent re-run.
• One dataset of healthy young adults; no clinical validation.

What would come next

A complete 2×2 ablation, same-split baseline re-runs, more seeds and splits, a masked pretext task that uses the target encoder, and a second dataset such as ABIDE.

Data were provided by the Human Connectome Project, WU-Minn Consortium (Principal Investigators: David Van Essen and Kamil Ugurbil; 1U54MH091657) funded by the 16 NIH Institutes and Centers that support the NIH Blueprint for Neuroscience Research; and by the McDonnell Center for Systems Neuroscience at Washington University.