NeurIPS 2025

I was at NeurIPS 2025 in San Diego with two things to show.

The first was a poster on ArtifactGen at the GenAI4Health workshop. The paper benchmarks WGAN-GP against diffusion for label-aware EEG artifact synthesis — the practical question being whether you can generate realistic artifacts well enough to make downstream sleep staging models more robust to them. The workshop crowd asked good, hard questions about evaluation, which is the part of generative modeling for physiological signals that nobody has fully solved.

The second was a live demo of Real-time MindEye, the MedARC project reconstructing visual perception from fMRI as it happens. Demoing a real-time system is a different kind of nervous than presenting a poster — the model has to actually work, in front of people, on the day. It did. Watching someone see a reconstruction of what a subject was looking at, seconds after they looked at it, is the clearest illustration I have of why this problem is worth working on.

Both threads are on arXiv: ArtifactGen and Real-time Reconstruction of Human Visual Perception from fMRI.