Input-output statistical independence in divisive normalization models of V1 neurons

Area of Science:

• Computational Neuroscience
• Visual Information Processing
• Machine Learning Models of Sensory Systems

Background:

• The primary visual cortex (V1) is hypothesized to be adapted to natural signals.
• Statistically-derived nonlinear divisive normalization models, proposed by Simoncelli et al., align with this adaptation hypothesis.
• These models aim to capture the statistical properties of sensory input.

Purpose of the Study:

• To provide a more rigorous mathematical formulation and analysis of statistically-derived divisive normalization models.
• To evaluate these models using mutual information as a metric for statistical independence.
• To investigate the independence properties of model outputs with respect to inputs.

Main Methods:

• Developed a rigorous mathematical framework for analyzing divisive normalization models.
• Utilized mutual information to quantify statistical independence between model outputs and inputs.
• Performed theoretical analysis on two models of natural image statistics.
• Validated findings with empirical results on natural image datasets.

Main Results:

• The specific divisive normalization parameters chosen by Simoncelli et al. do not guarantee statistical independence among output responses.
• However, these parameters ensure that each output response is statistically independent of almost all linear inputs.
• These findings hold true for theoretical analyses of two distinct natural image statistics models.

Conclusions:

• The study offers a mathematically rigorous analysis of established V1 models.
• The findings reveal a nuanced form of statistical independence within the models, where outputs are independent of most inputs, but not necessarily each other.
• Results support the principle of sensory adaptation to natural signals, with implications for understanding neural coding.