Research map / Time series / Foundation models for time series
Unsupervised Representation Learning of Brain Activity via Bridging Voxel Activity and Functional Connectivity
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes BRAINMIXER, an unsupervised MLP-based framework that jointly encodes voxel-level activity time series and functional connectivity graphs and aligns them via mutual-information (NCE) pre-training for brain classification and anomaly detection.
Problem
Brain representation learning typically uses either voxel-level activity (aggregated into beta weights, losing temporal dynamics) or ROI-level functional connectivity (losing voxel detail), and often requires supervised labels.
Approach
Two MLP-Mixer-style encoders: a voxel activity encoder with dynamic self-attention across time and voxels plus functional-system patching, and a functional connectivity encoder using biased temporal random-walk patches and a permutation-invariant adaptive temporal pooling; the two views are aligned with an NCE mutual-information objective.
Key results
- BRAINMIXER reaches 67.24% accuracy on BVFC multi-class brain classification, beating the best baseline PTGB (55.89%) and reporting a 14.3% average improvement over the best baseline.
- On edge-level anomaly detection with 1% corrupted edges, BRAINMIXER gets 91.62 AUC-PR on BVFC versus 87.13 for GraphMixer.
- Ablation shows removing pre-training drops BVFC classification accuracy from 67.24% to 63.58%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BRAINMIXER instead of GMM | brain classification · BVFC | accuracy: 53.11±1.44 -> 67.24±1.47 | Helped |
| Use BRAINMIXER instead of GMM | brain classification · BVFC-MEG | accuracy: 53.04±1.73 -> 62.58±1.12 | Helped |
| Use BRAINMIXER instead of GRAPHMIXER | brain classification · HCP-Mental | accuracy: 91.13±1.44 -> 96.32±0.29 | Helped |
| Use BRAINMIXER instead of GRAPHMIXER | brain classification · HCP-Age | accuracy: 48.32±1.11 -> 57.83±1.03 | Helped |
| Add pre-training (ablation) | BRAINMIXER · BVFC edge AD | AUC-PR: 88.75±2.16 -> 91.62±1.36 | Helped |
| Add the VA Encoder (ablation) | BRAINMIXER · BVFC edge AD | AUC-PR: 87.99±2.04 -> 91.62±1.36 | Helped |
| Use the VA Encoder instead of TST (ablation) | BRAINMIXER · BVFC edge AD | AUC-PR: 89.02±1.18 -> 91.62±1.36 | Helped |
| Add the FC Encoder (ablation) | BRAINMIXER · BVFC edge AD | AUC-PR: 84.27±4.37 -> 91.62±1.36 | Helped |
| Use the FC Encoder instead of BNTRANSFORMER (ablation) | BRAINMIXER · BVFC edge AD | AUC-PR: 87.18±2.03 -> 91.62±1.36 | Helped |
| Add functional patching (ablation) | BRAINMIXER · BVFC edge AD | AUC-PR: 86.35±2.97 -> 91.62±1.36 | Helped |
| Use functional patching instead of partitioning (ablation) | BRAINMIXER · BVFC edge AD | AUC-PR: 88.56±1.42 -> 91.62±1.36 | Helped |
| Use TPMIXER instead of MEAN pooling (ablation) | BRAINMIXER · BVFC edge AD | AUC-PR: 88.51±1.03 -> 91.62±1.36 | Helped |
Limitations
- Experiments on BVFC use only three fMRI subjects.
- Relies on a fixed functional-system atlas (Schaefer) for patching.
Experimental setup
Datasets: BVFC BVFC-MEG ADHD-200 TUH-EEG ASD (ABIDE) HCP
Benchmarks: BVFC BVFC-MEG HCP-Mental HCP-Age ADHD TUH-EEG ASD
Metrics: accuracy AUC-PR
Baselines: GOutlier NetWalk HyperSAGCN Graph MLP-Mixer GraphMixer BrainGNN FBNetGen BrainNetCNN ADMIRE BNTransformer PTGB USAD
Builds on: MLP-Mixer GraphMixer TST BNTransformer
Closest papers by meaning
- Brain Harmony: A Multimodal Foundation Model Unifying Morphology and Function into 1D Tokens (NeurIPS 2025 · similarity 0.86)
- NeuralFlix: A Simple While Effective Framework for Semantic Decoding of Videos from Non-invasive Brain Recordings (AAAI 2025 · similarity 0.86)
- Meta-Learning an In-Context Transformer Model of Human Higher Visual Cortex (NeurIPS 2025 · similarity 0.86)
- BrainLM: A foundation model for brain activity recordings (ICLR 2024 · similarity 0.85)
- Brain-OF: An Omnifunctional Foundation Model for fMRI, EEG and MEG (NeurIPS 2026 · similarity 0.85)
- FlatClip: Reusing Image Foundation Models for fMRI Representation Learning via Cortical Flatmaps (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.