metadata
title: SAE Brain Semantic Interface
emoji: 🧠
colorFrom: blue
colorTo: purple
sdk: docker
pinned: false
app_port: 7860
SAE Brain Explorer
Interactive explorer for Sparse Autoencoder (SAE) features with brain alignment.
Features
- UMAP scatter of 10,000 SAE features (DINOv2 L11 Spatial)
- Click any feature to see top-activating images with heatmap overlays
- CLIP text search across features
- NSD brain dataset: same features viewed through 34K natural scene images
- Phi-based brain alignment: which features are most predictive of fMRI activity
- Cortical map visualization
Data
- SAE: DINOv2 L11 Spatial, d=10,000, k=100
- Images: COCO val2017
- Brain: NSD sub01 fMRI aligned to SAE features via Phi regression
Environment Variables (set in Space settings)
| Variable | Required | Description |
|---|---|---|
HF_DATASET_REPO |
Yes | HF dataset repo ID, e.g. username/sae-nsd-data |
HF_TOKEN |
If private | HF read token for the dataset repo |
GOOGLE_API_KEY |
No | Enables Gemini-powered auto-labeling |