Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
126 tokens/sec
GPT-4o
47 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
4 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Meta Pseudo Labels (2003.10580v4)

Published 23 Mar 2020 in cs.LG and stat.ML

Abstract: We present Meta Pseudo Labels, a semi-supervised learning method that achieves a new state-of-the-art top-1 accuracy of 90.2% on ImageNet, which is 1.6% better than the existing state-of-the-art. Like Pseudo Labels, Meta Pseudo Labels has a teacher network to generate pseudo labels on unlabeled data to teach a student network. However, unlike Pseudo Labels where the teacher is fixed, the teacher in Meta Pseudo Labels is constantly adapted by the feedback of the student's performance on the labeled dataset. As a result, the teacher generates better pseudo labels to teach the student. Our code will be available at https://github.com/google-research/google-research/tree/master/meta_pseudo_labels.

Citations (611)

Summary

  • The paper presents a dynamic teacher-student mechanism that iteratively refines pseudo labels to enhance semi-supervised learning.
  • Experimental results demonstrate significant improvements, including a top-1 accuracy of 90.2% on ImageNet.
  • The approach mitigates confirmation bias and scales effectively to large datasets, paving the way for future adaptive SSL research.

Meta Pseudo Labels: A Semi-Supervised Learning Approach

The paper "Meta Pseudo Labels" introduces a novel approach to semi-supervised learning, where a teacher network is dynamically adapted based on the feedback from a student network. This mechanism addresses the conventional limitations associated with confirmation bias in pseudo-labeling, where a pre-trained teacher generates fixed pseudo labels, potentially leading to suboptimal learning outcomes.

Methodology Overview

Meta Pseudo Labels (MPL) builds upon the foundation of traditional Pseudo Labels, but with a significant enhancement: the teacher is not static. Instead, the teacher is continually refined based on the student’s performance on a labeled dataset. This iterative feedback loop ensures that the pseudo labels become progressively more accurate and beneficial for the student’s learning process.

The MPL method proceeds through a series of updates:

  1. Student Update: The student updates its parameters based on the pseudo labels provided by the teacher on unlabeled data.
  2. Teacher Update: The teacher refines its parameters using the performance feedback of the student on labeled data. This adaptive step is crucial as it aligns the pseudo label generation with the student’s learning state.

Experimental Results

The paper reports robust experimental results demonstrating the efficacy of MPL across several standard benchmarks, notably CIFAR-10-4K, SVHN-1K, ImageNet-10%, and full ImageNet.

  • ImageNet Achievement: MPL achieves a top-1 accuracy of 90.2% on ImageNet, an improvement of 1.6% over existing state-of-the-art results. This gain signifies the potential of MPL in leveraging large-scale unlabeled data effectively.
  • Comparison with Baselines: Across CIFAR-10-4K and ImageNet-10%, MPL consistently outperforms other semi-supervised methods, including Unsupervised Data Augmentation (UDA) and Label Smoothing.

Theoretical and Practical Implications

The introduction of a dynamic teacher in semi-supervised learning holds significant implications:

  • Reduction of Confirmation Bias: By adaptively updating the teacher based on student feedback, MPL mitigates the confirmation bias inherent in fixed teacher networks. This leads to more reliable pseudo labeling.
  • Scalability: The method demonstrates scalability to large datasets and complex models, such as EfficientNet-L2, making it applicable to diverse real-world tasks.

Future Developments

The adaptive mechanism introduced in MPL opens avenues for further exploration in several directions:

  • Exploring Other Architectures: Extending MPL to other neural architectures or even non-vision tasks could uncover additional benefits or adaptations needed for different contexts.
  • Integration with Other SSL Techniques: Combining MPL with complementary SSL techniques, such as contrastive learning, could enhance its performance.
  • Optimization of Teacher-Student Interplay: Investigating alternative reward structures for teacher updates might yield further improvements.

Conclusion

The paper on Meta Pseudo Labels presents a compelling advancement in semi-supervised learning by strategically incorporating student feedback to refine the teacher’s pseudo labels. The empirical results underline MPL’s potential to significantly boost performance on large-scale datasets, suggesting a promising trajectory for future AI and machine learning research.

Github Logo Streamline Icon: https://streamlinehq.com
Youtube Logo Streamline Icon: https://streamlinehq.com