Trusted-Data-Guided Label Enhancement on Noisy Labels.

Area of Science:

• Machine Learning
• Computer Vision
• Natural Language Processing

Background:

• Label distribution estimation is crucial for depicting data ambiguity.
• Data annotation often introduces label noise, complicating accurate label distribution recovery.
• Existing methods struggle with label enhancement on corrupted datasets.

Purpose of the Study:

• To propose a novel label enhancement (LE) method, TALEN, for recovering and refining label distribution from noisy data.
• To address the challenge of implementing LE on corrupted labels by leveraging trusted data.
• To improve the performance of classification models trained on noisy datasets.

Main Methods:

• TALEN recovers and progressively refines label distribution guided by a small batch of trusted data.
• An LE process is applied to untrusted data to identify samples with clean labels.
• A combined loss function is utilized for training the predictive classification model.

Main Results:

• Experiments validate the feasibility of identifying clean labels using the recovered label distribution on synthetic label noise.
• TALEN demonstrates a clear advantage over existing noise-robust learning methods on both synthetic and real-world label noise.
• The method shows effectiveness on image and text datasets, highlighting its versatility.

Conclusions:

• TALEN effectively recovers and refines label distribution even in the presence of significant label noise.
• The proposed method offers a robust solution for noise-robust learning, outperforming existing approaches.
• TALEN's ability to leverage trusted data makes it a valuable tool for real-world noisy data scenarios.