Reproduce “Task-guided cross-subject latent alignment: a multi-encoder-decoder VAE” 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
Angeliki Papathanasiou, Jascha Achterberg, Thomas E. Nichols, Rui Ponte Costa
A multi-encoder-decoder variational autoencoder (MED-VAE) aligns neural activity from different subjects into a shared latent space without requiring shared stimuli. This framework, guided by a pre-trained artificial neural network, yields superior semantic organization and enhances cross-subject neural and image prediction capabilities.
Key insights
- A pre-trained ANN can serve as an effective, task-relevant scaffold to guide the implicit alignment of heterogeneous neural data into a semantically organized common latent space without shared stimuli.
- The multi-objective training loss, particularly the fMRI-to-ANN reconstruction pathway, enables robust cross-subject generalization by aligning neural representations to a common, semantically rich ANN representation.
- Evaluating reconstruction quality with cross-trial correlation (rather than just within-trial) is crucial for distinguishing genuine stimulus-driven signal preservation from noise reconstruction in alignment methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.