Reproduce “Duala: Dual-Level Alignment of Subjects and Stimuli for Cross-Subject fMRI Decoding” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shumeng Li, Jintao Guo, Jian Zhang, Yulin Zhou, Luyang Cao, Yinghuan Shi
Duala adapts cross-subject fMRI visual decoders to new viewers using one-hour NSD data. It combines constraints that preserve category-level semantic relationships with distribution-based adjustment for individual differences. In averages across four NSD subjects, it achieves 84.5% image retrieval and 81.1% brain retrieval, compared with 83.1% and 76.0% for MindTuner. These results are limited to the tested dataset and subjects.
Key insights
- Explicitly enforcing semantic consistency at the stimulus level through triplet loss and relational alignment effectively preserves clear class boundaries and discriminability for new subjects, even with minimal adaptation data.
- Modeling and perturbing embedding distributions based on subject-specific and category-level variations robustly adapts the model to individual neural responses while maintaining overall semantic structure.
- Combining dual-level regularization with a parameter-efficient fine-tuning approach like LoRA allows for superior cross-subject fMRI decoding performance without requiring extensive new subject data or large model modifications.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.