Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AsyCo: An Asymmetric Dual-task Co-training Model for Partial-label Learning

Beibei Li, Yiyuan Zheng, Beihong Jin, Tao Xiang, Haobo Wang, Lei Feng

TL;DR

AsyCo tackles partial-label learning by introducing an asymmetric dual-task co-training framework comprising a disambiguation network that learns label confidence through PLL self-training and an auxiliary network that trains on noisier pairwise similarities derived from that confidence. Key mechanisms include data-augmented CC/RC losses for disambiguation, pseudo-label-to-pairwise similarity transformation for auxiliary training, and information distillation plus confidence refinement to correct errors. Empirical results across five datasets under uniform and instance-dependent label generation show AsyCo consistently outperforms state-of-the-art PLL methods, with robustness to label noise and dataset domain. The approach demonstrates that explicit task-based asymmetry and cross-network distillation yield complementary strengths, enabling more reliable label disambiguation in weak supervision. Limitations include increased training space, motivating future work on alternative co-training architectures and efficiency improvements.

Abstract

Partial-Label Learning (PLL) is a typical problem of weakly supervised learning, where each training instance is annotated with a set of candidate labels. Self-training PLL models achieve state-of-the-art performance but suffer from error accumulation problem caused by mistakenly disambiguated instances. Although co-training can alleviate this issue by training two networks simultaneously and allowing them to interact with each other, most existing co-training methods train two structurally identical networks with the same task, i.e., are symmetric, rendering it insufficient for them to correct each other due to their similar limitations. Therefore, in this paper, we propose an asymmetric dual-task co-training PLL model called AsyCo, which forces its two networks, i.e., a disambiguation network and an auxiliary network, to learn from different views explicitly by optimizing distinct tasks. Specifically, the disambiguation network is trained with self-training PLL task to learn label confidence, while the auxiliary network is trained in a supervised learning paradigm to learn from the noisy pairwise similarity labels that are constructed according to the learned label confidence. Finally, the error accumulation problem is mitigated via information distillation and confidence refinement. Extensive experiments on both uniform and instance-dependent partially labeled datasets demonstrate the effectiveness of AsyCo. The code is available at https://github.com/libeibeics/AsyCo.

AsyCo: An Asymmetric Dual-task Co-training Model for Partial-label Learning

TL;DR

AsyCo tackles partial-label learning by introducing an asymmetric dual-task co-training framework comprising a disambiguation network that learns label confidence through PLL self-training and an auxiliary network that trains on noisier pairwise similarities derived from that confidence. Key mechanisms include data-augmented CC/RC losses for disambiguation, pseudo-label-to-pairwise similarity transformation for auxiliary training, and information distillation plus confidence refinement to correct errors. Empirical results across five datasets under uniform and instance-dependent label generation show AsyCo consistently outperforms state-of-the-art PLL methods, with robustness to label noise and dataset domain. The approach demonstrates that explicit task-based asymmetry and cross-network distillation yield complementary strengths, enabling more reliable label disambiguation in weak supervision. Limitations include increased training space, motivating future work on alternative co-training architectures and efficiency improvements.

Abstract

Partial-Label Learning (PLL) is a typical problem of weakly supervised learning, where each training instance is annotated with a set of candidate labels. Self-training PLL models achieve state-of-the-art performance but suffer from error accumulation problem caused by mistakenly disambiguated instances. Although co-training can alleviate this issue by training two networks simultaneously and allowing them to interact with each other, most existing co-training methods train two structurally identical networks with the same task, i.e., are symmetric, rendering it insufficient for them to correct each other due to their similar limitations. Therefore, in this paper, we propose an asymmetric dual-task co-training PLL model called AsyCo, which forces its two networks, i.e., a disambiguation network and an auxiliary network, to learn from different views explicitly by optimizing distinct tasks. Specifically, the disambiguation network is trained with self-training PLL task to learn label confidence, while the auxiliary network is trained in a supervised learning paradigm to learn from the noisy pairwise similarity labels that are constructed according to the learned label confidence. Finally, the error accumulation problem is mitigated via information distillation and confidence refinement. Extensive experiments on both uniform and instance-dependent partially labeled datasets demonstrate the effectiveness of AsyCo. The code is available at https://github.com/libeibeics/AsyCo.
Paper Structure (30 sections, 16 equations, 5 figures, 12 tables)

This paper contains 30 sections, 16 equations, 5 figures, 12 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Problem Settings
  4. Classifier-consistent PLL Loss
  5. Risk-consistent PLL Loss
  6. The Proposed Model
  7. Disambiguation Network
  8. Auxiliary Network
  9. Error Correction
  10. Training and Inference
  11. Complexity Analysis
  12. Experiments
  13. Experimental Setup
  14. Datasets
  15. Compared Methods
  16. ...and 15 more sections

Figures (5)

  • Figure 1: Two examples of partially labeled instances. Due to the visual similarity between Ailurus fulgens and raccoons, the two images are both annotated with 'Ailurus fulgens' and 'raccoons'.
  • Figure 2: Architecture of AsyCo (in the training phase). AsyCo comprises two networks with identical structures, namely the disambiguation network and the auxiliary network. The former is responsible for resolving label ambiguities and learning label confidence, while the latter is trained by pairwise similarity labels constructed according to the learned confidence. Besides, the auxiliary network facilitates error correction for the disambiguation network through information distillation and confidence refinement, thereby mitigating the error accumulation problem.
  • Figure 3: An example of generating noisy pairwise similarity labels according to confidence vectors.
  • Figure 4: Comparison in terms of noise rate between pairwise similarity labels and pseudo class labels during training AsyCo. (a) Noise rates on SVHN; (b) Noise rates on CIFAR-10.
  • Figure 5: Impact of the temperature parameter $\tau$ on prediction accuracy on CIFAR-10.