Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Unified Classification and Rejection: A One-versus-All Framework

Zhen Cheng, Xu-Yao Zhang, Cheng-Lin Liu

TL;DR

This work presents a unified framework for open set recognition and OOD rejection by recasting the problem as a $K{+}1$-class task trained with one-versus-all (OVA) classifiers. Binary OVA posteriors are fused via Dempster–Shafer theory to yield a coherent set of $K$ known-class posteriors plus an explicit OOD posterior, enabling simultaneous classification and rejection without requiring OOD data during training. To preserve closed-set accuracy while maintaining strong rejection capability, the authors introduce a hybrid learning strategy that combines OVA loss with a regularized multi-class objective, and instantiate the approach on convolutional prototype networks with ViT backbones. Empirical results on OSR and OOD benchmarks (CIFAR and ImageNet-scale) show competitive or superior performance in InD accuracy, OOD detection, and misclassification detection, with notable gains when using modern architectures like ViT. The work highlights the practicality of a single, unified model for both classification and outlier rejection in open-world settings, and outlines future directions including broader backbones and joint evaluation of multiple failure modes.

Abstract

Classifying patterns of known classes and rejecting ambiguous and novel (also called as out-of-distribution (OOD)) inputs are involved in open world pattern recognition. Deep neural network models usually excel in closed-set classification while performs poorly in rejecting OOD inputs. To tackle this problem, numerous methods have been designed to perform open set recognition (OSR) or OOD rejection/detection tasks. Previous methods mostly take post-training score transformation or hybrid models to ensure low scores on OOD inputs while separating known classes. In this paper, we attempt to build a unified framework for building open set classifiers for both classification and OOD rejection. We formulate the open set recognition of $ K $-known-class as a $ (K+1) $-class classification problem with model trained on known-class samples only. By decomposing the $ K $-class problem into $ K $ one-versus-all (OVA) binary classification tasks and binding some parameters, we show that combining the scores of OVA classifiers can give $ (K+1) $-class posterior probabilities, which enables classification and OOD rejection in a unified framework. To maintain the closed-set classification accuracy of the OVA trained classifier, we propose a hybrid training strategy combining OVA loss and multi-class cross-entropy loss. We implement the OVA framework and hybrid training strategy on the recently proposed convolutional prototype network and prototype classifier on vision transformer (ViT) backbone. Experiments on popular OSR and OOD detection datasets demonstrate that the proposed framework, using a single multi-class classifier, yields competitive performance in closed-set classification, OOD detection, and misclassification detection.

Unified Classification and Rejection: A One-versus-All Framework

TL;DR

This work presents a unified framework for open set recognition and OOD rejection by recasting the problem as a -class task trained with one-versus-all (OVA) classifiers. Binary OVA posteriors are fused via Dempster–Shafer theory to yield a coherent set of known-class posteriors plus an explicit OOD posterior, enabling simultaneous classification and rejection without requiring OOD data during training. To preserve closed-set accuracy while maintaining strong rejection capability, the authors introduce a hybrid learning strategy that combines OVA loss with a regularized multi-class objective, and instantiate the approach on convolutional prototype networks with ViT backbones. Empirical results on OSR and OOD benchmarks (CIFAR and ImageNet-scale) show competitive or superior performance in InD accuracy, OOD detection, and misclassification detection, with notable gains when using modern architectures like ViT. The work highlights the practicality of a single, unified model for both classification and outlier rejection in open-world settings, and outlines future directions including broader backbones and joint evaluation of multiple failure modes.

Abstract

Classifying patterns of known classes and rejecting ambiguous and novel (also called as out-of-distribution (OOD)) inputs are involved in open world pattern recognition. Deep neural network models usually excel in closed-set classification while performs poorly in rejecting OOD inputs. To tackle this problem, numerous methods have been designed to perform open set recognition (OSR) or OOD rejection/detection tasks. Previous methods mostly take post-training score transformation or hybrid models to ensure low scores on OOD inputs while separating known classes. In this paper, we attempt to build a unified framework for building open set classifiers for both classification and OOD rejection. We formulate the open set recognition of -known-class as a -class classification problem with model trained on known-class samples only. By decomposing the -class problem into one-versus-all (OVA) binary classification tasks and binding some parameters, we show that combining the scores of OVA classifiers can give -class posterior probabilities, which enables classification and OOD rejection in a unified framework. To maintain the closed-set classification accuracy of the OVA trained classifier, we propose a hybrid training strategy combining OVA loss and multi-class cross-entropy loss. We implement the OVA framework and hybrid training strategy on the recently proposed convolutional prototype network and prototype classifier on vision transformer (ViT) backbone. Experiments on popular OSR and OOD detection datasets demonstrate that the proposed framework, using a single multi-class classifier, yields competitive performance in closed-set classification, OOD detection, and misclassification detection.
Paper Structure (14 sections, 1 theorem, 24 equations, 3 figures, 8 tables)

This paper contains 14 sections, 1 theorem, 24 equations, 3 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Proposed One-versus-All Framework
  4. Background of Classification with Rejection
  5. One-versus-All Learning Framework
  6. Hybrid Learning Strategy
  7. Instantiation with Convolutional Prototype Network
  8. Experiments
  9. Experimental Setup
  10. Evaluation on OOD Detection
  11. Evaluation on MisD
  12. Ablation Studies
  13. Results on Modern Architecture and Larger Dataset
  14. Conclusion and Future Work

Key Result

Theorem 3.1

The multiple binary classifiers in Eq. (equ:ova_posterior) can be combined into a multi-class classifier. Formally, the posterior probability of the combined classifier for class $i$, denoted as $p_{i}^{m}\left( \boldsymbol{x}\right)$, is

Figures (3)

  • Figure 1: A model unifying classification and rejection. The model's primary objective is to classify InD samples as accurately as possible (InD ✔), and also to reject possibly misclassified samples (InD ✘) and OOD samples.
  • Figure 2: Illustration of the difference between CE-based multi-class classifier and OVA-based multi-class classifier while processing OOD inputs,. The threshold parameters play an important role in OOD detection.
  • Figure 3: Classification with rejection on large-scale ImageNet-200/ResNet-50. Our method achieves consistent improvement on classification accuracy, MisD, and OOD detection, compared with CNN (CE). AURC is multiplied by $10^3$. Other values are percentages. The rejection rule of our method is based on $(K+1)$-class posterior probabilities. Other methods are based on $K$-class posterior probabilities.

Theorems & Definitions (1)

  • Theorem 3.1