Invariant feature based label correction for DNN when Learning with Noisy Labels.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Deep Neural Networks (DNNs) struggle with datasets containing incorrectly labeled data.
• Existing Learning with Noisy Labels (LNL) methods often rely on spurious semantic features, limiting their effectiveness.
• Unstable feature-label correlations across domains can compromise LNL method performance.

Purpose of the Study:

• To propose Invariant Feature based Label Correction (IFLC), a novel method to address spurious features in LNL.
• To enhance the accuracy of label noise correction by utilizing invariant features with stable correlations.
• To mitigate the limitations of current LNL approaches by reducing reliance on unstable semantic features.

Main Methods:

• Invariant Feature based Label Correction (IFLC) framework.
• Label Disturbing (LD) process to encourage stable DNN performance across environments and reduce spurious features.
• Representation Decorrelation (RD) process to enhance independence within feature representations for accurate label correction.
• Robust linear regression applied to feature representations for label correction.

Main Results:

• IFLC effectively reduces spurious features by encouraging stable DNN performance.
• The RD process enables accurate utilization of learned invariant features for label correction.
• Experimental results on CIFAR-10, CIFAR-100, Animal-10N, and Clothing1M datasets demonstrate competitive or superior performance compared to state-of-the-art LNL methods.

Conclusions:

• IFLC offers a novel approach to Learning with Noisy Labels by tackling the issue of spurious features.
• The proposed method achieves robust label correction by leveraging invariant features and decorrelated representations.
• IFLC represents a significant advancement in improving DNN performance on datasets with label noise.