ICML 2026 paper uses graph transformers to decode brain states unsupervised
Graph transformer autoencoder learns brain connectivity maps from fMRI data without supervision
Deep Dive
A geometry-guided graph transformer autoencoder learns compact representations of functional brain graphs in an unsupervised manner. By using aligned functional gradient geometry as an inductive bias, the model separates cognitive states and enables decoding of visual stimuli, with performance further improved by incorporating neural dynamics. It also generates synthetic brain graphs via a diffusion model, advancing network neuroscience without labeled data.
Key Points
- Graph transformer autoencoder learns compact graph-level embeddings of brain connectivity without supervised labels.
- Uses functional gradient geometry as an inductive bias to align topological and spectral properties across subjects.
- Decodes cognitive states and visual stimuli, and generates synthetic brain graphs via diffusion model.
Why It Matters
Enables brain decoding and synthetic network generation without costly labeled data, accelerating neuroscience research.