Neural population geometry and optimal coding of tasks with shared latent structure

Area of Science:

• Neuroscience
• Computational Neuroscience
• Cognitive Science

Background:

• The brain forms neural representations to navigate environments and generalize learning.
• Understanding which features of neural activity facilitate this generalization is crucial.
• Existing research highlights the role of neural representations but lacks specific geometric insights.

Purpose of the Study:

• To analytically determine the geometric properties of neural activity governing linear readout generalization.
• To identify key statistics of neural representations that enable cross-task learning.
• To link the geometry of neural population activity to its information content.

Main Methods:

• Analytical determination of geometric properties of neural activity.
• Analysis of four statistics: dimensionality, factorization, and correlation structures.
• Validation using biological and artificial neural data.

Main Results:

• Four statistics of neural activity geometry predict generalization performance.
• Optimal neural representations are lower-dimensional and more correlated early in learning.
• These geometric properties directly influence the linear decodability of information.

Conclusions:

• The geometry of neural population activity is a fundamental determinant of learning generalization.
• Specific geometric features, like dimensionality and correlations, are critical for efficient information readout.
• This work provides a framework for understanding how neural representations support flexible behavior.