Authors: Yule Wang, Joseph Yu, Chengrui Li, Weihan Li, and Anqi Wu, Georgia Tech, USA.

Official code for MIG-Vis (ICLR 2026 Poster): a framework to visualize and interpret neural latent groups via mutual-information guided diffusion.

• [2025/12] The abstract version of MIG-Vis has been accepted as a poster presentation at COSYNE 2026 in Lisbon, Portugal!

MIG‑Vis is a method to discover and visualize semantically meaningful latent subspaces encoded in neural activity from higher visual cortex recordings. It combines a group‑wise disentangled VAE for inferring structured neural latent subspaces, and a mutual information (MI)‑guided diffusion image synthesis procedure to visualize what each latent group encodes.

The full code and installation & experiment instructions will be available soon.

MIG‑Vis/
├── neural_vae/                  # VAE module for inferring neural latent groups
├── mig_diffusion/               # MI‑guided diffusion synthesis procedures
├── diffusion_model/             # Image diffusion function modules
├── utils_scripts/               # Utility scripts e.g., data loading, training helpers