A Parametrical Model for Instance-Dependent Label Noise.

Area of Science:

• Machine Learning
• Computer Science

Background:

• Estimating transition matrices is crucial for statistically consistent classifiers in label-noise learning.
• Existing methods like the clean-label transition matrix (CLTM) struggle with instance-dependent label noise.

Purpose of the Study:

• To propose a novel approach for estimating instance-dependent label-noise transition matrices.
• To develop a classifier that predicts Bayes optimal labels by modeling the transition from Bayes optimal labels to noisy labels (BLTM).

Main Methods:

• Directly modeling the Bayes-Label Transition Matrix (BLTM) from Bayes optimal labels to noisy labels.
• Utilizing the inherent certainty of Bayes optimal labels (one-hot vectors) to identify reliable training examples.
• Employing a deep neural network to parameterize and estimate the instance-dependent transition matrix.

Main Results:

• The proposed BLTM approach effectively handles instance-dependent label noise, outperforming traditional CLTM methods.
• The method allows for the collection of theoretically guaranteed Bayes optimal labels from noisy data.
• Deep neural networks enable accurate estimation of the instance-dependent transition matrix, leading to improved generalization.

Conclusions:

• Modeling the BLTM offers significant advantages over CLTM for instance-dependent label noise.
• The proposed deep learning-based method enhances classification performance and generalization in noisy label learning.
• This work provides a more robust framework for handling complex label noise patterns in machine learning.