NorMatch: Matching Normalizing Flows with Discriminative Classifiers for Semi-Supervised Learning
Abstract: Semi-Supervised Learning (SSL) aims to learn a model using a tiny labeled set and massive amounts of unlabeled data. To better exploit the unlabeled data the latest SSL methods use pseudo-labels predicted from \emph{a single discriminative classifier}. However, the generated pseudo-labels are inevitably linked to inherent confirmation bias and noise which greatly affects the model performance. In this work, we introduce a new framework for SSL named NorMatch. Firstly, we introduce a new uncertainty estimation scheme based on normalizing flows, as an auxiliary classifier, to enforce highly certain pseudo-labels yielding a boost of the discriminative classifiers. Secondly, we introduce a threshold-free sample weighting strategy to exploit better both high and low confidence pseudo-labels. Furthermore, we utilize normalizing flows to model, in an unsupervised fashion, the distribution of unlabeled data. This modelling assumption can further improve the performance of generative classifiers via unlabeled data, and thus, implicitly contributing to training a better discriminative classifier. We demonstrate, through numerical and visual results, that NorMatch achieves state-of-the-art performance on several datasets.
Submission Length: Regular submission (no more than 12 pages of main content)
Changes Since Last Submission: The camera-ready version has the following major changes:
1. Following the latest advice from Reviewer mhzF, we have updated Sec. 3.2.2 to better clarify the utilization of NF.
2. Following the suggestions from the Action Editor, we have included more discussions on our NorMatch in "Sec. 5 Limitations and Discussions".
3. We have included the "Acknowledgements" section.
This latest revision (the third revision) mainly addresses the concerns of Reviewer mhzF about the Normalizing Flow. The major changes are highlighted in cyan colour.
1. We further updated Sec. 3.1 (Motivation) to demonstrate the differences between the Normalizing Flow Classifier (NFC) and the Softmax-based discriminative classifier.
2. We added a new "Remark" paragraph to Sec. 3.2.2 to illustrate that our NFC is a Normalizing Flow model. We also clarified why the latent features, rather than the original images, were input to the Normalizing Flow.
Based on the previous revision which addresses the reviews from Reviewer Br9c, we further update the manuscript (the second revision) to address the concerns of the other two reviewers. The changes are highlighted in blue and brown colours.
1. We updated Sec. 3.1 to address the concerns on the normalizing flow classifier (the Q1 from Reviewer mhzF).
2. We added new experimental results of "NFC + NFC" in Tab. 2 and the corresponding analysis in Sec. 4.2. Tab. 2 also included the Multiply ACcumulate operations (MACs) to show the computational cost (for Q1 and Q4 from Reviewer GUVx).
3. We included more baselines on Mini-ImageNet in Tab. 7 and also provided some additional analysis in Sec. 4.3 (to address the Q3 of Reviewer GUVx).
4. We further discussed the limitations of our method, following the advice from Reviewer GUVx (Q2).
The first revision was mainly to address the concerns of Reviewer Br9c. Major changes are highlighted in red colour:
- Firstly, we have added the cross-reference to these concepts, e.g., NCUE and NUM (these two are now presented in Sections 3.2.1 and 3.2.2 to facilitate cross-reference in the Experiment section).
- Secondly, to improve readability, we have significantly reduced the use of abbreviations in the Method and Experiments sections, ensuring each term's first instance is fully defined.
- Thirdly, we have enhanced Section 4.2 to better display experimental results, including detailed explanations of the implementation and the use of background colors to highlight superior performances.
Code: https://github.com/Zhongying-Deng/NorMatch
Assigned Action Editor: ~Eleni_Triantafillou1
License: Creative Commons Attribution 4.0 International (CC BY 4.0)
Submission Number: 1711
Loading