Extended T: Learning With Mixed Closed-Set and Open-Set Noisy Labels.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Label noise significantly impacts classifier performance.
• Traditional transition matrices struggle with open-set and instance-dependent noise.
• Realistic datasets often exhibit mixed closed-set and open-set label noise.

Purpose of the Study:

• To develop a novel transition matrix capable of modeling mixed closed-set and open-set label noise.
• To address limitations of existing methods in handling instance-dependent label noise.
• To improve the robustness and accuracy of label noise learning algorithms.

Main Methods:

• Extended the traditional transition matrix to handle mixed label noise.
• Introduced a cluster-dependent transition matrix for instance-dependent noise.
• Designed an unbiased estimator (extended T-estimator) using only noisy data.

Main Results:

• The proposed cluster-dependent extended transition matrix effectively models realistic label noise.
• The extended T-estimator accurately estimates the new transition matrix.
• Experimental results demonstrate superior performance over state-of-the-art methods.

Conclusions:

• The developed method provides a more robust solution for learning with mixed and instance-dependent label noise.
• This work advances the field of label noise learning by addressing complex real-world noise patterns.
• The proposed approach enhances classifier consistency and performance in noisy environments.