Two-stage Training for Learning from Label Proportions

Jiabin Liu, Bo Wang, Xin Shen, Zhiquan Qi, Yingjie TianIJCAI 2021 · DOI

TL;DR

This paper introduces a novel two-stage training framework called PLOT for Learning from Label Proportions (LLP). PLOT addresses the limitations of existing methods by imposing strict proportion consistency through constrained optimization and employing techniques like symmetric cross-entropy and mixup to mitigate label noise, leading to significant performance improvements.

What Problem Does This Paper Address?

Learning from Label Proportions (LLP) is a challenging weakly supervised learning problem where only the proportions of class labels within groups (bags) of data are known, not the individual instance labels. Existing deep learning approaches often use end-to-end pipelines that minimize the Kullback-Leibler (KL) divergence between predicted and true label proportions. However, this unconstrained optimization can lead to solutions that don't strictly adhere to the given proportions and can result in high-entropy, noisy instance-level predictions, ultimately degrading classification performance. This paper aims to bridge this gap by proposing a more robust and accurate method for LLP.

Methods

The core contribution is a two-stage training framework:

  • Stage 1: Utilizes existing KL-divergence based LLP methods to obtain initial pseudo-labels.
  • Stage 2 (PLOT): This is the novel contribution. It treats the pseudo-labels from Stage 1 as noisy and refines them using a constrained optimization approach.
  • Optimal Transport (OT): The paper leverages Optimal Transport, specifically Kantorovich relaxation, to generate pseudo-labels that strictly comply with the bag-level label proportions. This is a key departure from methods that rely on inexact KL divergence.
  • Constrained Optimization: Instead of unconstrained KL divergence, PLOT imposes strict constraints to ensure the generated pseudo-labels match the given proportions.
  • Symmetric Cross-Entropy (SCE): To combat the inherent noise in pseudo-labels, SCE is used as the loss function in the second stage. SCE combines standard cross-entropy with reverse cross-entropy, making it more robust to label noise.
  • Mixup Augmentation: Further enhances robustness and generalization by creating new training samples through linear interpolation of existing samples and their labels.
  • Alternating Optimization: The framework alternately updates the network parameters and the pseudo-labels, allowing for continuous refinement.

The framework is designed to be model-agnostic, meaning it can be applied as a post-hoc phase to improve the performance of various existing deep LLP models.

Datasets

  • Synthetic Datasets: A "two-moon" dataset was used for initial qualitative evaluation.
  • Benchmark Datasets: CIFAR-10 and CIFAR-100 were used for extensive quantitative experiments. The paper notes that there are no off-the-shelf LLP datasets, so they construct LLP datasets from these supervised benchmarks by creating bags with specified label proportions.

Key Insights and Ideas

  • Proportion Consistency is Crucial: The paper strongly argues that simply minimizing KL divergence is insufficient for accurate LLP. Strict adherence to bag-level proportions through constrained optimization is key.
  • Pseudo-labels are Noisy: Recognizing that even pseudo-labels derived from proportions can be noisy, the paper introduces techniques (SCE, mixup) to explicitly handle this noise.
  • Optimal Transport for Precise Labeling: OT provides a principled way to generate instance-level labels that precisely match the aggregate bag proportions, offering a more accurate "guess" than KL divergence.
  • Model-Agnostic Improvement: The PLOT framework's ability to boost existing LLP methods as a post-hoc step highlights its versatility and practical utility.
  • SCE and Mixup Synergy: Combining SCE and mixup proves highly effective in reducing label noise and improving generalization, outperforming standard cross-entropy.

Key Results

  • Significant Accuracy Gains: PLOT consistently improves the accuracy of baseline LLP methods (DLLP, LLP-VAT, LLP-GAN) across different bag sizes on CIFAR-10 and CIFAR-100.
  • State-of-the-Art Performance: The combined methods (e.g., LLP-GAN + PLOT) achieve state-of-the-art results, particularly on CIFAR-100, with accuracies reaching 43.44% for bag size 128.
  • Improved Performance with Larger Bags: The improvement from PLOT is more pronounced with larger bag sizes (64 and 128), where supervision is weaker.
  • Qualitative Improvement: Visualizations on the "two-moon" dataset show a clear reduction in misclassifications when PLOT is applied.
  • Hard vs. Soft Labeling: Both hard and soft labeling strategies for updating pseudo-labels yield comparable results, with soft labeling showing a slight edge in convergence speed.

Limitations and Future Work

  • Computational Cost: While OT can be computationally efficient, the two-stage process and alternating optimization might increase overall training time compared to single-stage methods.
  • Hyperparameter Tuning: The paper mentions that mixup augmentation and SCE might require careful hyperparameter tuning, although they claim it's "lightly additional."
  • Future Work: The authors suggest exploring more advanced OT formulations and further investigating the interplay between different noise reduction techniques and OT for LLP. They also implicitly suggest that this framework could be extended to other weakly supervised learning scenarios where aggregate information is available.

Summary updated

BibTeX
@inproceedings{liu2021two,
  author = {Jiabin Liu and Bo Wang and Xin Shen and Zhiquan Qi and Yingjie Tian},
  title = {{Two-stage Training for Learning from Label Proportions}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.24963/ijcai.2021/377},
}
Figure