Dimensionality-induced information loss of outliers in deep neural networks

TL;DR

This paper investigates how the intrinsic low-dimensional structure of deep neural networks influences the distinguishability of out-of-distribution samples, revealing that low-dimensional weights cause information loss and proposing a dimensionality-aware detection method.

Contribution

It uncovers the role of low-dimensional feature representations in OOD detection and introduces a novel, efficient method leveraging dimensionality alignment for improved detection performance.

Findings

Low-dimensional features make OOD samples more distinguishable in deeper layers.

Low-dimensional weights cause information loss, leading to misclassification due to dataset bias.

The proposed dimensionality-aware method achieves high performance with lower computational cost.

Abstract

Out-of-distribution (OOD) detection is a critical issue for the stable and reliable operation of systems using a deep neural network (DNN). Although many OOD detection methods have been proposed, it remains unclear how the differences between in-distribution (ID) and OOD samples are generated by each processing step inside DNNs. We experimentally clarify this issue by investigating the layer dependence of feature representations from multiple perspectives. We find that intrinsic low dimensionalization of DNNs is essential for understanding how OOD samples become more distinct from ID samples as features propagate to deeper layers. Based on these observations, we provide a simple picture that consistently explains various properties of OOD samples. Specifically, low-dimensional weights eliminate most information from OOD samples, resulting in misclassifications due to excessive attention to dataset bias. In addition, we demonstrate the utility of dimensionality by proposing a dimensionality-aware OOD detection method based on alignment of features and weights, which consistently achieves high performance for various datasets with lower computational cost.