Common population codes produce extremely nonlinear neural manifolds.

Area of Science:

• Computational neuroscience
• Neural coding
• Data analysis

Background:

• Neural populations encode variables through distributed activity patterns.
• Neural activity often resides in lower-dimensional subspaces within the full state space.
• Understanding the geometry of these subspaces is crucial for interpreting neural computations.

Purpose of the Study:

• To investigate the linearity of low-dimensional structures in neural population activity.
• To determine if common population codes result in linear or nonlinear manifolds.
• To assess the impact of nonlinearity on dimensionality reduction techniques like PCA.

Main Methods:

• Analysis of point clouds and manifolds derived from common population codes.
• Theoretical analysis of the dimension of best-fitting linear subspaces.
• Comparison of theoretical predictions with the performance of linear methods.

Main Results:

• Neural population activity data frequently exhibit exceedingly nonlinear structures.
• The dimension of the best-fitting linear subspace grows exponentially with the true data dimension.
• Linear methods, including PCA, dramatically fail to identify the true underlying structure.

Conclusions:

• The geometry of neural population codes is often highly nonlinear.
• Linear dimensionality reduction techniques are inadequate for accurately characterizing many neural population codes.
• New methods are needed to analyze the complex, nonlinear structure of neural activity.