Bridging Generative and Discriminative Noisy-Label Learning via Direction-Agnostic EM Formulation.

Area of Science:

• Machine Learning
• Computer Vision
• Natural Language Processing

Background:

• Noisy-label learning is crucial for real-world data.
• Discriminative methods are common but lack principled generative insights.
• Existing generative models have limitations like extra variables and fixed data directions.

Purpose of the Study:

• To propose a novel single-stage, EM-style generative framework for noisy-label learning.
• To develop a direction-agnostic approach that avoids explicit image synthesis.
• To introduce an instance-specific prior for improved data-dependent regularization.

Main Methods:

• A single Expectation Maximization (EM) objective is derived, specializing to causal orientations.
• Intractable p(X|Y) is replaced with a dataset-normalized discriminative proxy.
• Partial-Label Supervision (PLS) is introduced as an instance-specific prior.

Main Results:

• Achieved state-of-the-art accuracy on vision and NLP noisy label benchmarks.
• Demonstrated lower transition-matrix estimation error.
• Required substantially less training computation compared to baselines.

Conclusions:

• The proposed generative framework offers a principled and efficient alternative for noisy-label learning.
• The direction-agnostic and PLS approach enhances adaptability and regularization.
• This method sets a new standard for performance and computational cost in noisy-label learning.