Latent Class-Conditional Noise Model.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Learning with noisy labels is crucial in the Big Data era, reducing annotation costs.
• Existing Class-Conditional Noise (CCN) models require impractical anchor sets for noise transition estimation.
• Neural layer adaptations for noise estimation suffer from ill-posed stochastic learning and local minima.

Purpose of the Study:

• To introduce a novel Latent Class-Conditional Noise model (LCCN) for robust noise transition parameterization.
• To develop a stable Bayesian framework that avoids the limitations of previous noise estimation methods.
• To generalize LCCN for various applications including open-set, semi-supervised, and cross-model learning.

Main Methods:

• Parameterizing noise transition within a Bayesian framework using LCCN.
• Projecting noise transition into Dirichlet space for constrained learning on a dataset simplex.
• Employing a dynamic label regression method with a Gibbs sampler for latent true label inference.

Main Results:

• LCCN ensures stable updates of noise transition, preventing arbitrary mini-batch tuning.
• The proposed method demonstrates superior performance over state-of-the-art techniques in experiments.
• Generalizations of LCCN show effectiveness in diverse noisy label scenarios.

Conclusions:

• LCCN provides a stable and theoretically grounded approach to learning with noisy labels.
• The Bayesian framework and Dirichlet space projection offer significant advantages over prior methods.
• LCCN and its variants represent a substantial advancement in handling noisy data for machine learning tasks.