Thermodynamics of natural images

TL;DR

This paper explores the thermodynamic properties of natural images, revealing criticality and special configurations with error-correcting features, linking image statistics to neural responses in visual cortex.

Contribution

It introduces a thermodynamic framework for analyzing natural images, demonstrating criticality and identifying configurations with error correction properties.

Findings

Evidence of criticality through diverging specific heat.

Identification of special image configurations with error-correcting features.

Neural models resembling visual cortex cells detect these configurations.

Abstract

The scale invariance of natural images suggests an analogy to the statistical mechanics of physical systems at a critical point. Here we examine the distribution of pixels in small image patches and show how to construct the corresponding thermodynamics. We find evidence for criticality in a diverging specific heat, which corresponds to large fluctuations in how "surprising" we find individual images, and in the quantitative form of the entropy vs. energy. The energy landscape derived from our thermodynamic framework identifies special image configurations that have intrinsic error correcting properties, and neurons which could detect these features have a strong resemblance to the cells found in primary visual cortex.