Revisiting the high-dimensional geometry of population responses in the visual cortex

Area of Science:

• Neuroscience
• Computational Neuroscience
• Systems Neuroscience

Background:

• Large-scale neural recordings reveal complex, high-dimensional neural codes.
• Characterizing neural geometry from noisy data, especially with more neurons than trials, presents statistical challenges.

Purpose of the Study:

• Develop tools for accurate estimation of high-dimensional signal geometry.
• Investigate the geometry of neural representations in the mouse primary visual cortex (V1).
• Re-evaluate previous findings of power-law geometry in V1 representations.

Main Methods:

• Applied novel statistical tools for high-dimensional geometry estimation.
• Analyzed neural response data from the mouse primary visual cortex.
• Utilized principal component analysis and eigenmode analysis of population activity.

Main Results:

• V1 neural geometry is better described by a broken power law, with distinct exponents for early and late eigenmodes.
• Later modes decay more rapidly, indicating a lower-dimensional representation concentrated in early modes.
• Population modes encode visual features with higher fidelity and are more predictable by models than single neurons.

Conclusions:

• The neural code in V1 exhibits emergent structure, characterized by a broken power law.
• Population-level analysis offers greater tractability and insight into sensory representation compared to single-neuron studies.
• Deep network and classical models show significantly improved predictive performance for population eigenmodes over single neurons.