SSL-BN: Self-Supervised Learning Based on Structural Similarities in Brain Networks

ICLR 2026 Conference Submission21157 Authors

Keywords: Self-supervised learning, brain connectivity networks, fMRI-based diagnosis

TL;DR: We propose a self-supervised framework for brain networks, which pre-trains the encoder by first dispersing the embeddings of all samples, and then applying a specially calculated non-trainable matrix to linearly transform the embeddings.

Abstract: Functional magnetic resonance imaging (fMRI) data provide critical information for the diagnosis of neurological disorders, as correlations among features of different regions of interest (ROIs) capture functional characteristics of the brain. Brain networks are an effective modeling paradigm for fMRI data, and recent works have explored GNN-based and Transformer-based approaches for brain network analysis. However, the dense and weighted edge structure of brain networks poses challenges for GNN-based methods, while Transformer-based methods typically require large amounts of labeled data. To address these issues, we propose a **S**elf-**S**upervised **L**earning framework for **B**rain **N**etworks (SSL-BN). Our approach pretrains a Brain Network Transformer by dispersing sample embeddings and refining them with a fixed, non-trainable matrix derived from a novel structural similarity measure, enabling contrastive representation learning without data augmentation. To our knowledge, SSL-BN is the first self-supervised framework specifically designed for brain networks. It employs a simple loss function, eliminates the need for augmentation, and significantly improves model performance on limited labeled data. Extensive experiments on the publicly available ABIDE dataset demonstrate that SSL-BN achieves state-of-the-art performance compared to prior methods.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 21157