# Learning Efficient Coding of Natural Images with Maximum Manifold Capacity Representations

**Authors:** Thomas Yerxa, Yilun Kuang, Eero Simoncelli, SueYeon Chung  
**arXiv:** [2303.03307](https://arxiv.org/abs/2303.03307)  
**Date:** 2023-03-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.03307v2/

## Key points

- Sensory systems finely tuned to extract maximum information while considering biological constraints
- Challenging to measure information theoretic properties in practical settings
- Computational models use metrics like reconstruction error
- New efficiency metric quantifying number of linearly separable object categories
- Introduced as a way to optimize manifold capacity directly, resulting in MMCRs
- Competitive performance on SSL benchmarks
- Insights into how manifold compression enhances class separability
- Empirical analyses comparing MMCRs with other SSL frameworks:
- Revealed unique characteristics
- Effective models for the ventral stream in neural predictivity benchmarks
- Study by Thomas Yerxa, Yilun Kuang, Eero Simoncelli, and SueYeon Chung:
- Accepted at NeurIPS 2023
- Provides valuable contributions to understanding efficient coding of natural images through MMCR representations.

## Summary

The efficient coding hypothesis suggests that sensory systems are finely tuned to extract maximum information from their inputs while considering biological constraints. However, measuring information theoretic properties in practical settings is challenging, leading computational models to use various metrics like reconstruction error to test this hypothesis. Recent advancements have introduced a new efficiency metric called manifold capacity, which quantifies the number of object categories that can be linearly separated. Despite being computationally intensive, researchers have developed a method to optimize manifold capacity directly, resulting in Maximum Manifold Capacity Representations (MMCR). Inspired by self-supervised learning (SSL), MMCRs have shown competitive performance on SSL benchmarks and offer insights into how manifold compression enhances class separability. Furthermore, empirical analyses comparing MMCRs with other SSL frameworks reveal unique characteristics and suggest that MMCRs serve as effective models for the ventral stream in neural predictivity benchmarks. This study by Thomas Yerxa, Yilun Kuang, Eero Simoncelli, and SueYeon Chung was accepted at NeurIPS 2023 and provides valuable contributions to understanding efficient coding of natural images through MMCR representations.