Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

BootOOD: Self-Supervised Out-of-Distribution Detection via Synthetic Sample Exposure under Neural Collapse

Yuanchao Wang, Tian Qin, Eduardo Valle, Bruno Abrahao

TL;DR

BootOOD addresses the challenge of detecting semantically similar OOD samples without relying on external outliers. It leverages Neural Collapse to enforce a radial, norm-based separation between ID and pseudo-OOD signals via a lightweight radius head that is trained with feature-space mixup-derived pseudo-OOD features, while keeping the primary classifier unchanged. The method achieves strong near-OOD performance on CIFAR-10, CIFAR-100, and ImageNet-200, outperforming post-hoc and OE-free training methods and remaining competitive with outlier-exposure approaches while preserving ID accuracy. By decoupling OOD modeling from the main classifier and exploiting NC geometry, BootOOD provides a stable, self-supervised route to robust OOD detection in challenging semantic regimes with practical impact for safety-critical deployment.

Abstract

Out-of-distribution (OOD) detection is critical for deploying image classifiers in safety-sensitive environments, yet existing detectors often struggle when OOD samples are semantically similar to the in-distribution (ID) classes. We present BootOOD, a fully self-supervised OOD detection framework that bootstraps exclusively from ID data and is explicitly designed to handle semantically challenging OOD samples. BootOOD synthesizes pseudo-OOD features through simple transformations of ID representations and leverages Neural Collapse (NC), where ID features cluster tightly around class means with consistent feature norms. Unlike prior approaches that aim to constrain OOD features into subspaces orthogonal to the collapsed ID means, BootOOD introduces a lightweight auxiliary head that performs radius-based classification on feature norms. This design decouples OOD detection from the primary classifier and imposes a relaxed requirement: OOD samples are learned to have smaller feature norms than ID features, which is easier to satisfy when ID and OOD are semantically close. Experiments on CIFAR-10, CIFAR-100, and ImageNet-200 show that BootOOD outperforms prior post-hoc methods, surpasses training-based methods without outlier exposure, and is competitive with state-of-the-art outlier-exposure approaches while maintaining or improving ID accuracy.

BootOOD: Self-Supervised Out-of-Distribution Detection via Synthetic Sample Exposure under Neural Collapse

TL;DR

BootOOD addresses the challenge of detecting semantically similar OOD samples without relying on external outliers. It leverages Neural Collapse to enforce a radial, norm-based separation between ID and pseudo-OOD signals via a lightweight radius head that is trained with feature-space mixup-derived pseudo-OOD features, while keeping the primary classifier unchanged. The method achieves strong near-OOD performance on CIFAR-10, CIFAR-100, and ImageNet-200, outperforming post-hoc and OE-free training methods and remaining competitive with outlier-exposure approaches while preserving ID accuracy. By decoupling OOD modeling from the main classifier and exploiting NC geometry, BootOOD provides a stable, self-supervised route to robust OOD detection in challenging semantic regimes with practical impact for safety-critical deployment.

Abstract

Out-of-distribution (OOD) detection is critical for deploying image classifiers in safety-sensitive environments, yet existing detectors often struggle when OOD samples are semantically similar to the in-distribution (ID) classes. We present BootOOD, a fully self-supervised OOD detection framework that bootstraps exclusively from ID data and is explicitly designed to handle semantically challenging OOD samples. BootOOD synthesizes pseudo-OOD features through simple transformations of ID representations and leverages Neural Collapse (NC), where ID features cluster tightly around class means with consistent feature norms. Unlike prior approaches that aim to constrain OOD features into subspaces orthogonal to the collapsed ID means, BootOOD introduces a lightweight auxiliary head that performs radius-based classification on feature norms. This design decouples OOD detection from the primary classifier and imposes a relaxed requirement: OOD samples are learned to have smaller feature norms than ID features, which is easier to satisfy when ID and OOD are semantically close. Experiments on CIFAR-10, CIFAR-100, and ImageNet-200 show that BootOOD outperforms prior post-hoc methods, surpasses training-based methods without outlier exposure, and is competitive with state-of-the-art outlier-exposure approaches while maintaining or improving ID accuracy.

Paper Structure

This paper contains 57 sections, 20 equations, 2 figures, 9 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Post-hoc OOD Detection
  4. Training-based OOD Detection Without Outlier Exposure
  5. Training-based Methods with Outlier Exposure
  6. Neural Collapse and Feature Norm for OOD Detection
  7. Our differences.
  8. Methodology
  9. Problem Setup
  10. Neural Collapse as a Geometric Prior for BootOOD
  11. NC essentials and geometry.
  12. Phase-1 $\rightarrow$ Phase-2 warm start.
  13. Why NC helps OOD detection.
  14. Radius head and directional separation.
  15. Pseudo-OOD Feature Generation
  16. ...and 42 more sections

Figures (2)

  • Figure 1: Overview of the BootOOD geometry. ID class features collapse around their class means with consistent norms under Neural Collapse. BootOOD synthesizes pseudo-OOD features in between these means and trains an auxiliary radius-based head to classify them as having smaller feature norms. The global mean defines the center, ID samples occupy a reference radius $r_{\text{ref}}$, and a sequence of inner radii $r_1, r_2, r_3$ specifies relaxed OOD constraint regions. By separating ID and OOD purely in feature-norm space, BootOOD avoids enforcing strict angular (orthogonality) constraints and relieves the main classifier from modeling OOD structure.
  • Figure 2: CIFAR-100 feature diagnostics: (left) radius $\|\mathbf{z}\|_2$; (right) max cosine to class weights $\max_c \cos(\mathbf{z},\mathbf{w}_c)$. ID (blue) vs. OOD (orange).