Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning Multi-Manifold Embedding for Out-Of-Distribution Detection

Jeng-Lin Li, Ming-Ching Chang, Wei-Chao Chen

TL;DR

A novel Multi-Manifold Embedding Learning (MMEL) framework is introduced, optimizing hypersphere and hyperbolic spaces jointly for enhanced OOD detection and analyzing the effects of learning multiple manifolds and visualize OOD score distributions across datasets.

Abstract

Detecting out-of-distribution (OOD) samples is crucial for trustworthy AI in real-world applications. Leveraging recent advances in representation learning and latent embeddings, Various scoring algorithms estimate distributions beyond the training data. However, a single embedding space falls short in characterizing in-distribution data and defending against diverse OOD conditions. This paper introduces a novel Multi-Manifold Embedding Learning (MMEL) framework, optimizing hypersphere and hyperbolic spaces jointly for enhanced OOD detection. MMEL generates representative embeddings and employs a prototype-aware scoring function to differentiate OOD samples. It operates with very few OOD samples and requires no model retraining. Experiments on six open datasets demonstrate MMEL's significant reduction in FPR while maintaining a high AUC compared to state-of-the-art distance-based OOD detection methods. We analyze the effects of learning multiple manifolds and visualize OOD score distributions across datasets. Notably, enrolling ten OOD samples without retraining achieves comparable FPR and AUC to modern outlier exposure methods using 80 million outlier samples for model training.

Learning Multi-Manifold Embedding for Out-Of-Distribution Detection

TL;DR

A novel Multi-Manifold Embedding Learning (MMEL) framework is introduced, optimizing hypersphere and hyperbolic spaces jointly for enhanced OOD detection and analyzing the effects of learning multiple manifolds and visualize OOD score distributions across datasets.

Abstract

Detecting out-of-distribution (OOD) samples is crucial for trustworthy AI in real-world applications. Leveraging recent advances in representation learning and latent embeddings, Various scoring algorithms estimate distributions beyond the training data. However, a single embedding space falls short in characterizing in-distribution data and defending against diverse OOD conditions. This paper introduces a novel Multi-Manifold Embedding Learning (MMEL) framework, optimizing hypersphere and hyperbolic spaces jointly for enhanced OOD detection. MMEL generates representative embeddings and employs a prototype-aware scoring function to differentiate OOD samples. It operates with very few OOD samples and requires no model retraining. Experiments on six open datasets demonstrate MMEL's significant reduction in FPR while maintaining a high AUC compared to state-of-the-art distance-based OOD detection methods. We analyze the effects of learning multiple manifolds and visualize OOD score distributions across datasets. Notably, enrolling ten OOD samples without retraining achieves comparable FPR and AUC to modern outlier exposure methods using 80 million outlier samples for model training.
Paper Structure (21 sections, 15 equations, 3 figures, 5 tables)

This paper contains 21 sections, 15 equations, 3 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work and Preliminary
  3. Hypersphere Embedding
  4. Hyperbolic Embedding
  5. Multiple Manifold Learning
  6. Method
  7. Learning Multiple Manifold Embedding
  8. Learning hypersphere manifold
  9. Learning hyperbolic manifold
  10. Loss optimization
  11. OOD Score Calculation
  12. Test-time OOD Sample Enrollment
  13. Experiments
  14. Out-of-distribution Detection Accuracy
  15. In-distribution Classification Accuracy
  16. ...and 6 more sections

Figures (3)

  • Figure 1: Overview of the proposed Multi-Manifold Embedding Learning (MMEL) framework for OOD detection. The upper part indicates the network structure trained with the hypersphere and hyperbolic manifolds which are illustrated in the right box for details. The lower part indicates the OOD score computation in the inference phase.
  • Figure 2: Results of (a) FPR$_{95}$ and (b) AUC percentage (%) using $N_e$ OOD samples as a negative anchor, where each bar denotes the score using $N$ numbers of samples. CIFAR-100 is used as the ID samples and test on the six OOD datasets.
  • Figure 3: Histogram visualization of OOD scores for ID samples in blue and OOD samples in green.