Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Class Relevance Learning For Out-of-distribution Detection

Butian Xiong, Liguang Zhou, Tin Lun Lam, Yangsheng Xu

TL;DR

The paper tackles OOD detection in image classification, identifying limitations of max-logit based cues that ignore interclass relationships. It introduces class relevance learning (CRL), which constructs a class relevance matrix $P_{crm}$ from per-class prototype logits $L_k$ and uses a final score $P_{cf}$ that fuses max-logit information with a class-relateness distance $P_{cr}$ (where $P_{cr}= sum_i P_t ext{log} rac{P_t}{P_{crm}(I_{pc})}$ and $P_{cf}=- ext{max}( heta(z)) \\alpha - rac{1}{P_{cr}} \\beta$). The approach offers a simple post-processing step to estimate $P_{crm}$ and demonstrates improved OOD detection over state-of-the-art baselines on Near-OOD and Far-OOD benchmarks, validated by extensive experiments. Overall, CRL shows that exploiting interclass relationships via a class relevance matrix can substantially enhance the reliability and robustness of OOD detection in practical deployments.

Abstract

Image classification plays a pivotal role across diverse applications, yet challenges persist when models are deployed in real-world scenarios. Notably, these models falter in detecting unfamiliar classes that were not incorporated during classifier training, a formidable hurdle for safe and effective real-world model deployment, commonly known as out-of-distribution (OOD) detection. While existing techniques, like max logits, aim to leverage logits for OOD identification, they often disregard the intricate interclass relationships that underlie effective detection. This paper presents an innovative class relevance learning method tailored for OOD detection. Our method establishes a comprehensive class relevance learning framework, strategically harnessing interclass relationships within the OOD pipeline. This framework significantly augments OOD detection capabilities. Extensive experimentation on diverse datasets, encompassing generic image classification datasets (Near OOD and Far OOD datasets), demonstrates the superiority of our method over state-of-the-art alternatives for OOD detection.

Class Relevance Learning For Out-of-distribution Detection

TL;DR

The paper tackles OOD detection in image classification, identifying limitations of max-logit based cues that ignore interclass relationships. It introduces class relevance learning (CRL), which constructs a class relevance matrix from per-class prototype logits and uses a final score that fuses max-logit information with a class-relateness distance (where and ). The approach offers a simple post-processing step to estimate and demonstrates improved OOD detection over state-of-the-art baselines on Near-OOD and Far-OOD benchmarks, validated by extensive experiments. Overall, CRL shows that exploiting interclass relationships via a class relevance matrix can substantially enhance the reliability and robustness of OOD detection in practical deployments.

Abstract

Image classification plays a pivotal role across diverse applications, yet challenges persist when models are deployed in real-world scenarios. Notably, these models falter in detecting unfamiliar classes that were not incorporated during classifier training, a formidable hurdle for safe and effective real-world model deployment, commonly known as out-of-distribution (OOD) detection. While existing techniques, like max logits, aim to leverage logits for OOD identification, they often disregard the intricate interclass relationships that underlie effective detection. This paper presents an innovative class relevance learning method tailored for OOD detection. Our method establishes a comprehensive class relevance learning framework, strategically harnessing interclass relationships within the OOD pipeline. This framework significantly augments OOD detection capabilities. Extensive experimentation on diverse datasets, encompassing generic image classification datasets (Near OOD and Far OOD datasets), demonstrates the superiority of our method over state-of-the-art alternatives for OOD detection.
Paper Structure (8 sections, 3 equations, 2 figures, 3 tables)

This paper contains 8 sections, 3 equations, 2 figures, 3 tables.

Table of Contents

  1. INTRODUCTION
  2. Methodology
  3. Problem Statement
  4. Class Relevance Learning
  5. Experiments and Results
  6. Experimental Setup
  7. Experimental Results
  8. Conclusion

Figures (2)

  • Figure 1: Proposed class relevance learning framework for measuring the class relevance score of a test sample to the constructed class relevance matrix of training dataset.
  • Figure 2: This figure displays the difference between max logits and class relevance learning. We train the ResNet18 model on the CIFAR10 dataset and test it on the TinyImageNet le2015tiny and Places365 dataset. The left column shows the confidence score distribution of CRL and right column shows the distribution of Maxlogits.