Visual information flow in Wilson-Cowan networks

Area of Science:

• Computational Neuroscience
• Computer Vision
• Information Theory

Background:

• The retina-V1 pathway's communication efficiency is crucial for visual processing.
• Existing models often focus on statistical independence, but biological systems may use different principles.
• Wilson-Cowan networks offer a biologically plausible model for neural interactions.

Purpose of the Study:

• To analyze the communication efficiency of a psychophysically tuned cascade of Wilson-Cowan and divisive normalization layers simulating the retina-V1 pathway.
• To investigate the reduction of total correlation in neural responses along this simulated pathway.
• To explore the potential of neural field models for image compression.

Main Methods:

• First-time analysis of Wilson-Cowan networks using multivariate total correlation.
• Derivation of cortical model parameters from the relationship between Wilson-Cowan and divisive normalization models.
• Theoretical expression for total correlation reduction and empirical study using natural scenes and advanced statistical tools for estimating multivariate total correlation.

Main Results:

• The cascade of layers substantially reduces redundancy between neural responses, despite not being optimized for statistical independence.
• Wilson-Cowan networks exhibit similar efficiency to equivalent divisive normalization models.
• Nonlinear local contrast computation and oriented filters contribute most significantly to total correlation reduction.

Conclusions:

• Psychophysically tuned models are more efficient in regions with higher luminance-contrast.
• The findings provide an alternative confirmation of the efficient coding hypothesis for Wilson-Cowan systems.
• Neural field models are suggested as a viable alternative to divisive normalization for image compression.