Generative Reasoning Integrated Label Noise Robust Deep Image Representation Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning-based image representation learning (IRL) methods are crucial for image understanding.
• High-quality annotated data is essential but costly to acquire, leading to the use of noisy labels from crowdsourcing or automated methods.
• Existing methods often overfit to noisy labels, resulting in suboptimal performance.

Purpose of the Study:

• To develop a label noise robust deep representation learning approach.
• To address the challenge of inaccurate image characterization caused by noisy training data.
• To integrate generative reasoning with discriminative reasoning for improved IRL under noisy labels.

Main Methods:

• Introduction of the Generative Reasoning Integrated Label Noise Robust Deep Representation Learning (GRID) approach.
• Integration of generative reasoning into discriminative reasoning using a supervised variational autoencoder.
• Implementation of a hybrid representation learning strategy that utilizes generative reasoning for noisy samples and discriminative reasoning for clean samples.

Main Results:

• The GRID approach effectively detects training samples with noisy labels.
• It adjusts the learning procedure to mitigate the impact of label noise.
• Experimental results demonstrate the superiority of GRID over state-of-the-art methods in handling noisy labels for IRL.

Conclusions:

• The proposed GRID approach offers a robust solution for IRL in the presence of label noise.
• It is independent of the IRL method, annotation type, network architecture, loss function, or learning task.
• GRID provides a versatile and effective strategy for various image understanding problems.