Mechanisms of dimensionality reduction and decorrelation in deep neural networks

TL;DR

This paper develops a mean-field theoretical framework to understand how deep neural networks reduce dimensionality and decorrelate features across layers, enhancing interpretability and insight into sensory processing.

Contribution

It introduces a unified theory explaining how deep networks perform dimensionality reduction and decorrelation, applicable to both deterministic and generative models.

Findings

Deep networks implement dimensionality reduction across layers.

Neurons maintain weak correlations for feature extraction.

Unified framework applies to various deep network types.

Abstract

Deep neural networks are widely used in various domains. However, the nature of computations at each layer of the deep networks is far from being well understood. Increasing the interpretability of deep neural networks is thus important. Here, we construct a mean-field framework to understand how compact representations are developed across layers, not only in deterministic deep networks with random weights but also in generative deep networks where an unsupervised learning is carried out. Our theory shows that the deep computation implements a dimensionality reduction while maintaining a finite level of weak correlations between neurons for possible feature extraction. Mechanisms of dimensionality reduction and decorrelation are unified in the same framework. This work may pave the way for understanding how a sensory hierarchy works.