Berkeley team's V1-inspired diffusion model reveals brain-AI parallels

A new paper from UC Berkeley (Yun et al., arXiv 2607.15693) proposes a minimal diffusion model directly inspired by the primary visual cortex (V1). The model extends classic sparse coding by introducing a non-factorial prior over latent variables via an unconstrained pairwise interaction matrix, turning the inference process into a general recurrent dynamical system. The team trained these recurrent dynamics using a denoising score-matching objective with implicit differentiation, avoiding the need for expensive rollout simulations.

The results are striking: the learned interaction matrix structurally resembles the horizontal connections in superficial layers of V1 that link neurons with similar orientation tuning. The model achieves denoising performance nearly on par with standard black-box diffusion architectures, even reconstructing extended contours under extreme ambiguity. Crucially, the model's simplicity allows direct decomposition of its Jacobian in terms of the interaction matrix, revealing how the recurrent dynamics assign high probability to continuous families of natural structural deformations. An unexpected finding is that a large fraction of latent variables learn to disconnect from visual input entirely, forming a hierarchical representation that enforces global consistency across image features. This work not only generates testable hypotheses for functional connectivity in biological circuits during perceptual inference, but also provides mechanistic insight into how diffusion models generalize from finite training data to generate infinite novel images.