Controlling Neural Collapse Enhances Out-of-Distribution Detection and Transfer Learning

2025-02-15

AI Summary

This paper addresses the inherent trade-off between out-of-distribution (OOD) detection and OOD generalization in deep neural networks—specifically, the conflict wherein stronger neural collapse (NC) improves OOD detection but degrades generalization, and vice versa. We propose a hierarchical, controllable NC mechanism: strengthening NC in lower layers to enhance OOD detection while suppressing NC in higher layers to improve transfer generalization. We establish, for the first time, theoretical connections between NC intensity and both OOD detection and generalization performance. Furthermore, we design a plug-and-play, architecture-agnostic framework for hierarchical NC control, integrating entropy regularization (to suppress NC) and Simplex ETF projection (to promote NC). Extensive experiments on benchmarks including CIFAR/SVHN and ImageNet-O demonstrate state-of-the-art performance: our method simultaneously improves AUROC by +2.3% and transfer accuracy by +1.8%, effectively reconciling the detection–generalization tension.

Technology Category

• Machine Learning: Deep Generative Models & Autoencoders
• Computer Vision: Learning & Optimization for CV
• Search and Optimization: Learning to Search

Application Category

• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods
• Security and Privacy: Data transparency and provenance

Abstract

Out-of-distribution (OOD) detection and OOD generalization are widely studied in Deep Neural Networks (DNNs), yet their relationship remains poorly understood. We empirically show that the degree of Neural Collapse (NC) in a network layer is inversely related with these objectives: stronger NC improves OOD detection but degrades generalization, while weaker NC enhances generalization at the cost of detection. This trade-off suggests that a single feature space cannot simultaneously achieve both tasks. To address this, we develop a theoretical framework linking NC to OOD detection and generalization. We show that entropy regularization mitigates NC to improve generalization, while a fixed Simplex Equiangular Tight Frame (ETF) projector enforces NC for better detection. Based on these insights, we propose a method to control NC at different DNN layers. In experiments, our method excels at both tasks across OOD datasets and DNN architectures.

Problem

• Neural Collapse affects OOD detection and generalization
• Trade-off between detection and generalization in DNNs
• Control NC to improve both detection and generalization

Innovation

• Entropy regularization improves generalization
• ETF projector enhances OOD detection
• Control NC at different DNN layers