Spatio-chromatic information available from different neural layers via Gaussianization

Area of Science:

• Neuroscience
• Computational Vision
• Information Theory

Background:

• The visual system processes complex retinal images through multiple layers.
• Previous studies suggested specialized subsystems for optimal information transmission.
• The efficiency of these layers operating together remains unquantified.

Purpose of the Study:

• To quantify visual information extracted across different layers of the visual pathway.
• To assess the contribution of spatial, chromatic, and nonlinear transforms.
• To analyze information transmission using multivariate information-theoretic units.

Main Methods:

• Developed a statistical tool using a Gaussianization technique for information estimation.
• Employed colorimetrically calibrated natural images as input.
• Measured total correlation across the joint spatio-chromatic response array.

Main Results:

• Deeper representations in the visual pathway capture more information.
• Information transmission aligns with the probability distribution of natural scenes.
• Spatial transforms contribute more to information capture than chromatic transforms.
• Nonlinearities substantially enhance transmitted information, but less than linear transforms.

Conclusions:

• The visual pathway progressively encodes more information in deeper layers.
• Information processing reflects natural scene statistics.
• Linear spatial transforms are key drivers of visual information extraction.