| name | beyond-neural-activity-prediction |
| description | Multi-level representational probing framework for evaluating digital twins of sensory cortex beyond standard prediction accuracy. Probes latent representations (linear decodability, latent-unit tuning, population geometry) in mouse V1 digital twins. Based on arXiv:2605.23122 (May 2026). Use when evaluating brain digital twins, comparing model architectures for neural prediction, or studying latent representations in vision models. |
Beyond Neural Activity Prediction: Probing Latent Representations in Mouse V1 Digital Twins
Methodology from arXiv:2605.23122 (May 2026).
Authors: Adriano Lima, Yuchen Hou, Michael Beyeler, Marius Schneider
Subjects: q-bio.NC
Overview
This paper addresses a critical gap in evaluating digital twins of sensory cortex: although prediction accuracy is the central metric, it provides limited insight into the latent representations that support those predictions. Models with similar prediction accuracy may rely on different latent representations, which matters increasingly as digital twins are used for in silico experimental design.
Key Findings
1. Prediction Accuracy Correlates with Representation Quality
- Across architectures, better neural-response prediction correlates with:
- Stronger probe accuracy (linear decodability of visual features)
- Flatter hidden-population eigenspectra (higher-dimensional representations)
- Closer population-geometry signatures to mouse V1
2. Comparable Accuracy ≠ Comparable Representations
- Digital twins with comparable prediction scores can differ substantially in:
- Probe performance
- Latent-unit tuning properties
3. Multi-Level Probing Framework
Three levels of latent representation characterization:
Level 1: Linear Decodability
- Controlled visual probes of orientation, contrast, and motion
