Ambiguity-Induced Contrastive Learning for Instance-Dependent Partial Label Learning

Shiyu Xia, Jiaqi Lv, Ning Xu, Xin GengIJCAI 2022 · OpenReview

Abstract

Partial label learning (PLL) learns from a typical weak supervision, where each training instance is labeled with a set of ambiguous candidate labels (CLs) instead of its exact ground-truth label. Most existing PLL works directly eliminate, rather than exploiting the label ambiguity, since they explicitly or implicitly assume that incorrect CLs are noise independent of the instance. While a more practical setting in the wild should be instance-dependent, namely, the CLs depend on both the true label and the instance itself, such that each CL may describe the instance from some sensory channel, thereby providing some noisy but really valid information about the instance. In this paper, we leverage such additional information acquired from the ambiguity and propose AmBiguity-induced contrastive LEarning (ABLE) under the framework of contrastive learning. Specifically, for each CL of an anchor, we select a group of samples currently predicted as that class as ambiguity-induced positives, based on which we synchronously learn a representor (RP) that minimizes the weighted sum of contrastive losses of all groups and a classifier (CS) that minimizes a classification loss. Although they are circularly dependent: RP requires the ambiguity-induced positives on-the-fly induced by CS, and CS needs the first half of RP as the representation extractor, ABLE still enables RP and CS to be trained simultaneously within a coherent framework. Experiments on benchmark datasets demonstrate its substantial improvements over state-of-the-art methods for learning from the instance-dependent partially labeled data.

Introduction

In real-world data annotation, human labelers often struggle to make exact judgments, resulting in weak supervision. Partial Label Learning (PLL) addresses this by training on data where each sample is tagged with a set of candidate labels (CLs), only one of which is the true ground truth.

Most existing PLL methods assume that the incorrect labels in the candidate set are chosen randomly (i.e., they are instance-independent). However, in the wild, label ambiguity depends heavily on the instance itself. For example, a human annotator is much more likely to confuse an image of a "Siberian tiger" with a "Bengal tiger" or "leopard" than with a "bookcase." Prior works generally try to eliminate this ambiguity as pure noise. This paper argues that instance-dependent ambiguity actually provides valuable semantic clues about the data that should be leveraged to improve model training.

The concept of instance-dependent partial label learning, where candidate labels are semantically related to the input instance.

The ABLE Framework

The authors propose AmBiguity-induced contrastive LEarning (ABLE). The core idea is to use contrastive learning to pull an image closer to other images that share its semantic ambiguities, effectively treating ambiguity as a feature rather than a bug.

The ABLE framework architecture, showing the synchronous training of the representation network and the classifier.

Key Methodological Components

  1. Ambiguity-Induced Positives Selection: For a given training image (the anchor), ABLE looks at its candidate labels. It then searches the current training batch for other images that the model currently predicts as belonging to those candidate classes, treating them as positive pairs.
  2. Weighted Contrastive Learning: ABLE pulls the anchor and these "ambiguity-induced positives" closer in the representation space while pushing other instances away. Because some candidate labels are more likely to be the ground truth than others, the contrastive loss is weighted based on the classifier's current confidence in each candidate label.
  3. Synchronous Update Strategy: There is a circular dependency in this architecture: the representation network (RP) needs the classifier (CS) to find positive pairs and calculate weights, while the CS needs the RP to extract features. ABLE solves this by training both synchronously in an end-to-end framework, allowing them to iteratively improve each other.

Experimental Evaluation

The authors evaluated ABLE on four benchmark vision datasets: MNIST, Fashion-MNIST, Kuzushiji-MNIST, and CIFAR-10. Since these datasets traditionally provide clean labels, the authors synthesized instance-dependent partial labels by training a neural network on clean labels and using its confidence predictions to set the flipping probability for incorrect labels, simulating realistic human confusion.

Key Results

  • State-of-the-Art Accuracy: ABLE outperforms five recent baseline methods (VALEN, PRODEN, RC, CC, D2CNN) across all tested datasets. On CIFAR-10, ABLE reaches 92.30% accuracy, beating the next best method (VALEN at 90.16%) by over 2%.
  • Ablation on Ambiguity: Removing the ambiguity-induced positives drops CIFAR-10 accuracy by ~2.5% (down to 89.82%), proving that explicitly exploiting the candidate labels is highly effective.
  • Architecture Scaling: The method scales well with model capacity. Upgrading the encoder from ResNet-18 to ResNet-50 improves CIFAR-10 accuracy to 92.94%. Furthermore, using a nonlinear projection head for the contrastive loss yields a massive +7% boost over using no projection head.

Conclusion

ABLE demonstrates that the specific uncertainties in weak labels contain rich structural information about the feature space. By treating samples that share candidate labels as positive pairs, contrastive learning naturally groups semantically similar instances even before the exact ground-truth label is fully disambiguated. Future work may involve using memory queues to expand the pool of available positives beyond the current batch and exploring dynamic loss-balancing strategies.

Summary updated

Abstract

Partial label learning (PLL) learns from a typical weak supervision, where each training instance is labeled with a set of ambiguous candidate labels (CLs) instead of its exact ground-truth label. Most existing PLL works directly eliminate, rather than exploiting the label ambiguity, since they explicitly or implicitly assume that incorrect CLs are noise independent of the instance. While a more practical setting in the wild should be instance-dependent, namely, the CLs depend on both the true label and the instance itself, such that each CL may describe the instance from some sensory channel, thereby providing some noisy but really valid information about the instance. In this paper, we leverage such additional information acquired from the ambiguity and propose AmBiguity-induced contrastive LEarning (ABLE) under the framework of contrastive learning. Specifically, for each CL of an anchor, we select a group of samples currently predicted as that class as ambiguity-induced positives, based on which we synchronously learn a representor (RP) that minimizes the weighted sum of contrastive losses of all groups and a classifier (CS) that minimizes a classification loss. Although they are circularly dependent: RP requires the ambiguity-induced positives on-the-fly induced by CS, and CS needs the first half of RP as the representation extractor, ABLE still enables RP and CS to be trained simultaneously within a coherent framework. Experiments on benchmark datasets demonstrate its substantial improvements over state-of-the-art methods for learning from the instance-dependent partially labeled data.
BibTeX
@inproceedings{xia2022ambiguity,
  author = {Shiyu Xia and Jiaqi Lv and Ning Xu and Xin Geng},
  title = {{Ambiguity-Induced Contrastive Learning for Instance-Dependent Partial Label Learning}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2022},
  url = {https://openreview.net/forum?id=4kjeOoKzg5r},
}
Figure