P035 The geometry of primary visual cortex representations is dynamically adapted to task performance
Leyla Roksan Caglar*1,Julien Corbo*2, O.Batuhan Erkat2,3, Pierre-Olivier Polack2
1Windreich Department of AI & Human Health, Icahn School of Medicine at Mount Sinai, New York, NY, USA
2Center for Molecular and Behavioral Neuroscience, Rutgers University—Newark, Newark, NJ, USA
3Graduate Program in Neuroscience, Rutgers University—Newark, Newark, NJ, USA
*Contributed equally; Email: l.r.caglar@gmail.com; julien.corbo@gmail.com
Introduction
Perceptual learning optimizes perception by reshaping sensory representations to enhance discrimination and generalization. Although these mechanisms’ implementation remains elusive, recent advances suggest that the neural geometry of the representations is key, by preparing population activity to be read out at the next processing stage. Our previous work has shown that learning a visual discrimination task reshapes the population feature representations in the primary visual cortex (V1) via suppressive mechanisms, effectively discretizing the representational space, and favoring categorization and generalization [1]. However, it is unclear how these changes impact the discriminability of the representations when being read out and transformed into a decision variable.
Methods
Recent findings under the Manifold Capacity Theory [2]suggest that learning enhances classification capacity by altering the geometric properties of population activity, increasing the linear separability of stimulus representations. To test this, we examined the relationship between V1 feature representation, neural manifold geometry, and behavioral discrimination, hypothesizing that the previously observed discretization would enhance classification capacity and alter manifold geometry as early as V1. Using calcium imaging, we compared V1 activity between trained and naïve mice performing an orientation discrimination Go/NoGo task at varying difficulty levels.
Results
Investigating response dimensionality, we found it increased as the Go/NoGo stimuli became more similar in both trained and naïve mice. As predicted, dimensionality was lower in trained animals, suggesting the task's biological implementation relies on reducing representational dimensionality. However, dimensionality alone did not fully explain performance variability. Instead, we found that the linear separability of representations in their embedding space was a stronger predictor of individual behavioral performance. This separability of manifolds was further evidenced by measuring the neural manifold’s capacity and their geometric properties (manifold dimension and manifold radius), which all show a decrease with successful behavioral performance in the trained mice, but show no change in the naive mice.
Discussion
Taken together, our results show a clear relationship between behavioral task performance, representational dimensionality, and manifold separability in the early visual cortex of mice. Across all computational measures, we demonstrated an inverse relationship between dimensionality and successful perceptual discrimination assisted by representational separability.These results confirm that learning alters the geometric properties of early sensory representations as early as in V1, optimizing them for linear readout and improving perceptual decision-making.
Acknowledgements
The authors are grateful to the members of the Polack lab for the helpful conversations. This work was funded by The Whitehall Foundation (grant 2015-08-69). The Charles and Johanna Busch Biomedical Grant Program The National Institutes of Health National Eye Institute: Grant #R01 EY030860 Brain initiative: Grant #R01 NS120289) Fyssen Foundation postdoctoral fellowship.
References
[1] Corbo, J., Erkat, O. B., McClure, J., Khdour, H., & Polack, P.-O. (2025). Discretized representations in V1 predict suboptimal orientation discrimination.Nature Communications,16(1), 41. https://doi.org/10.1038/s41467-024-55409-1
[2]Chung, S., Lee, D. D., & Sompolinsky, H. (2016). Linear readout of object manifolds.Physical Review E,93(6), 060301. https://doi.org/10.1103/PhysRevE.93.060301
