IC-444ViV1T produces non-differentiable population response representations when simulating mouse V1 experiments

Dario Liscai, Emanuele Luconi, Alessandro Marin Vargas, Alessandro Sanzeni

SourceBeyond single neurons: population response geometry in digital twins of mouse visual cortex

The paper tests ViV1T, a transformer-based digital twin from the Dynamic Sensorium Competition, by simulating the Stringer et al. (2019) V1 population geometry experiment. Despite achieving improved single-neuron prediction accuracy (corr-to-average 0.478 on Sensorium), ViV1T's population response eigenspectrum decays too slowly to be differentiable, matching the failure pattern observed in CNN-based digital twins. The non-differentiability is confirmed across multiple stimulus types (natural images, whitened, spatially localized, gratings), indicating the limitation is not architecture-specific.

Evidence
observational
Caveat
The stimuli used in the Stringer et al. experiment are substantially different from those used in ViV1T's training, so the result may partly reflect domain shift rather than a fundamental representational limitation.
Model
ViV1T
Concepts
Failure mode
Datasets
MICrONS [source]
Extraction
automatic-extraction