beyond-neural-activity-prediction

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
• Tests whether visual features are linearly accessible in latent space

Level 2: Latent-Unit Tuning

• Orientation selectivity index
• Contrast response functions
• Spatial-frequency tuning

Level 3: Population Geometry

• Hidden-layer activity eigenspectra
• Dimensionality of representations
• Comparison with mouse V1 population signatures

Methodology

1. Train digital twins of mouse V1 with different visual-encoder architectures sharing:

   • Same training data (naturalistic videos from freely moving mice)
   • Same neural-prediction objective

2. Freeze models after training

3. Systematically probe latent representations at three levels

4. Correlate representation quality with prediction accuracy

5. Compare models with comparable prediction but different representations

Implications

• Digital twin validation: Prediction accuracy alone is insufficient — latent representation quality matters
• Model selection: Different architectures with similar accuracy may support different in silico experiments
• Brain-AI alignment: Representation probing provides mechanistic understanding beyond correlation-based evaluation

Activation Keywords

• digital twin, neural prediction, latent representation
• V1 modeling, mouse cortex, representational probing
• population geometry, linear decodability, neural encoding
• model comparison, brain digital twin evaluation