EmoBrain studies fine-grained emotion decoding from short-window, low-data task-fMRI using pretrained brain and vision representations.
The canonical system uses Qwen/Qwen3-VL-4B-Instruct with two swappable brain
encoder families:
- E1 ViT: an image-pretrained ViT adapted to an fMRI ROI grid.
- E2 BFM: Brain-JEPA or SwiFT pretrained brain representations.
It predicts 34 independent emotion scores. A multimodal teacher receives brain, V-JEPA2 video, and human-written MindCaptioning descriptions; a brain-only student learns from both hard labels and cached teacher outputs. The scientific scope includes transfer under short temporal windows and limited task-fMRI, contextual supervision, future cross-dataset generalization, and cortical interpretability.
The only supported implementation is project/code/. The earlier Qwen2.5
pipeline is preserved under project/legacy/qwen25/ for provenance and must not
be used for new experiments.
See README_KR.md and
docs/notes/implementation_spec_20260702.md.