New Optimal Transport Method Improves Cold-Start Active Learning

Ning Zhu, Xiaochuan Ma, Juntao Xu, Jingze Liang, Mengfei Zhao, An Chen, Liang-Jian Deng· August 5, 2026 View original

Key takeaways

  • A new optimal transport framework unifies and improves Cold-Start Active Learning.
  • The epsilon-Adaptive Selection (epsilon-AS) algorithm adapts to data and tasks automatically.
  • epsilon-AS achieves state-of-the-art performance, improving accuracy and reducing selection time.
  • This method is particularly beneficial for data-scarce or new ML projects.

Who benefits

AI/ML DevelopmentData Annotation ServicesAutonomous VehiclesHealthcare (medical imaging)E-commerce

Summary

Researchers propose a unified optimal transport framework for Cold-Start Active Learning (CSAL) that adapts automatically to data and tasks, overcoming limitations of existing methods. Their new algorithm, epsilon-Adaptive Selection (epsilon-AS), consistently achieves state-of-the-art performance across various datasets by dynamically adjusting regularization strength.

Cold-Start Active Learning (CSAL) aims to efficiently select valuable data subsets for labeling without prior knowledge, a critical challenge in machine learning. Current CSAL methods often rely on specific inductive biases, leading to inconsistent performance across different tasks. This new research introduces a unified framework based on optimal transport, which can adaptively adjust to the specific data and task at hand.The framework reveals a shared allocation structure among existing methods and provides a theoretical analysis of the trade-off controlled by entropic regularization. Based on this, the researchers developed epsilon-Adaptive Selection (epsilon-AS), a novel Sinkhorn-based CSAL algorithm. Extensive experiments demonstrate that epsilon-AS consistently outperforms state-of-the-art methods, showing significant accuracy improvements and reduced selection times on large datasets like ImageNet-1k.

Why it matters

Data scientists and ML engineers can leverage this new active learning approach to significantly reduce the cost and time associated with data labeling, especially in scenarios with limited initial data or when starting new projects.

How to implement this in your domain

  1. 1Review the upcoming code release for epsilon-AS to understand its implementation details.
  2. 2Integrate epsilon-AS into your data labeling pipelines for new machine learning projects.
  3. 3Compare its performance against existing active learning strategies on your specific datasets.
  4. 4Evaluate the trade-off between labeling budget and model accuracy using this adaptive selection method.
  5. 5Apply the framework to cold-start scenarios where initial labeled data is scarce.

Original post by Ning Zhu, Xiaochuan Ma, Juntao Xu, Jingze Liang, Mengfei Zhao, An Chen, Liang-Jian Deng

"arXiv:2608.03249v1 Announce Type: new Abstract: Cold-Start Active Learning (CSAL) aims to select a valuable subset from an unlabeled pool without any prior knowledge or human assistance. Existing methods take diverse routes based on typicality, coverage, or diversity. Each rests…"

View on X

Primary sources

Originally posted by Ning Zhu, Xiaochuan Ma, Juntao Xu, Jingze Liang, Mengfei Zhao, An Chen, Liang-Jian Deng on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses