Second order dimensionality reduction using minimum and maximum mutual information models.

Area of Science:

• Computational Neuroscience
• Systems Neuroscience
• Information Theory

Background:

• Characterizing neural feature selectivity is crucial for understanding brain function.
• Conventional methods like spike-triggered covariance (STC) and maximally informative dimensions (MID) are limited to Gaussian stimuli and struggle with high-dimensional data due to the curse of dimensionality.

Purpose of the Study:

• To develop novel dimensionality reduction techniques for analyzing multidimensional neural feature selectivity.
• To overcome the limitations of existing methods when dealing with non-Gaussian stimulus distributions and high-dimensional neural data.

Main Methods:

• Proposed two new information-theoretic dimensionality reduction methods based on minimum and maximum information principles.
• These methods extend spike-triggered covariance (STC) to handle non-Gaussian stimulus distributions.
• Evaluated methods using simulated neurons responding to natural images and electrophysiological recordings from macaque retinal and thalamic cells responding to naturalistic stimuli.

Main Results:

• The proposed minimum and maximum information methods significantly outperformed STC when using non-Gaussian stimuli.
• Maximally informative dimensions (MID) performed best in low-dimensional feature spaces.
• The new methods successfully identified relevant linear subspaces of arbitrary dimensionality with complex stimuli.

Conclusions:

• The novel information-theoretic approaches provide a powerful framework for analyzing neural feature selectivity with complex, naturalistic stimuli.
• These methods offer significant advantages over conventional techniques, particularly in non-Gaussian regimes and high-dimensional scenarios.
• The findings advance our ability to decode neural representations and understand sensory processing in the brain.