Digital Twins of Mouse V1: Probing Latent Representations Beyond Prediction Accuracy

A new study on arXiv (May 2026) moves beyond simply measuring prediction accuracy in digital twins of mouse primary visual cortex (V1). The authors — Lima, Hou, Beyeler, and Schneider — trained a family of digital twins on neural activity from freely moving mice watching naturalistic videos, varying only the visual-encoder architecture. For each frozen model, they characterized latent representations at three levels: (i) linear decodability of controlled visual probes (orientation, contrast, motion); (ii) latent-unit tuning to canonical features like orientation selectivity and spatial frequency; and (iii) population geometry of hidden-layer activity. This multi-level probing aims to uncover whether models that predict equally well actually rely on different internal representations.

The results reveal that across architectures, better neural-response prediction correlates with stronger probe accuracy and flatter hidden-population eigenspectra — a signature of higher-dimensional representations closer to actual mouse V1 population geometry. Critically, even digital twins with comparable prediction scores can differ substantially in probe performance and latent-unit tuning. This means a model might predict neural firing accurately but still misrepresent how the brain encodes specific visual features. The work establishes representational probing as a necessary complement to standard evaluation, ensuring digital twins serve as credible in silico experimental systems for stimulus design and hypothesis generation in visual neuroscience.