Trusted Multi-View Learning Under Noisy Supervision

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Multi-view learning often prioritizes accuracy over uncertainty, limiting use in safety-critical areas.
• Existing trusted multi-view methods require high-quality labels, hindering application with noisy data.
• Developing reliable multi-view learning models with noisy labels is a significant challenge.

Purpose of the Study:

• To propose a reliable multi-view learning model that can effectively handle noisy labels.
• To address the challenge of modeling label noise stemming from poor data features and class confusion.
• To improve decision accuracy and uncertainty estimation in multi-view learning systems operating under label noise.

Main Methods:

• Proposed Trusted Multi-view Noise Refining (TMNR) using evidential deep neural networks and noise correlation matrices.
• Developed Trusted Multi-view Noise Re-Refining (TMNR2) to disentangle co-training complexities via distinct module objectives.
• TMNR2 utilizes evidence-label consistency for mislabeled sample identification and neighbor-based pseudo-label generation.

Main Results:

• TMNR2 significantly outperforms state-of-the-art baselines on 7 multi-view datasets.
• Achieved average accuracy improvements of 7% on datasets with 50% label noise.
• Demonstrated stabilized training by reducing mapping interference between evidential networks and noise correlation matrices.

Conclusions:

• TMNR2 offers a robust solution for multi-view learning in the presence of noisy labels.
• The proposed methods enhance both decision accuracy and uncertainty estimation.
• The findings have implications for applying multi-view learning in real-world, data-scarce scenarios.