Learning With Auxiliary Less-Noisy Labels.

Area of Science:

• Machine Learning
• Computer Science

Background:

• Acquiring sufficient accurate labels for training classifiers is challenging due to limited reliable resources.
• Real-world applications often use less-accurate labels from non-expert sources, leading to performance degradation from high noise rates.

Purpose of the Study:

• To propose a novel learning method that leverages both noisy and auxiliary less-noisy labels.
• To address the performance deterioration caused by high label noise in classification tasks.

Main Methods:

• A learning method based on a flipping probability noise model and logistic regression classifier.
• Simultaneous estimation of noise rate parameters, inference of ground-truth labels, and classifier learning via maximum likelihood.
• Development of three learning algorithms based on different prior knowledge states of less-noisy labels.

Main Results:

• The proposed method demonstrates tolerance to label noise.
• The method outperforms classifiers that do not explicitly utilize auxiliary less-noisy labels.
• Experimental results validate the effectiveness of the proposed approach.

Conclusions:

• The proposed method offers a robust solution for classification with noisy labels by incorporating auxiliary less-noisy data.
• This approach improves classification performance in scenarios with limited access to perfectly labeled data.