Robust Trusted Conflictive Multiview Collaborative Contrastive Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Multiview learning methods often prioritize accuracy over decision uncertainty.
• Real-world multiview data frequently exhibits misalignment, leading to conflicting instances and limiting applications in safety-critical domains.
• Existing methods for improving multiview reliability struggle with performance degradation when handling conflicting instances.

Purpose of the Study:

• To propose a novel method, Robust Trusted Conflictive Multiview Collaborative Contrastive Learning (RCMCL), to enhance robustness and generalization in conflictive multiview scenarios.
• To address the limitations of current multiview learning techniques in handling decision uncertainty and data misalignment.
• To improve the reliability of multiview learning for safety-critical applications.

Main Methods:

• Utilizes an evidential deep neural network to generate view-specific opinions.
• Employs dissonance-based evidence contrastive learning for opinion consistency across views.
• Incorporates collaborative learning of consistent and complementary evidence, introducing vacuity degree and category-level contrastive learning.

Main Results:

• The proposed RCMCL method demonstrates enhanced robustness and generalization capabilities in conflictive multiview settings.
• Experimental results on eight benchmark datasets show RCMCL outperforms state-of-the-art methods.
• The method effectively integrates consistent and complementary evidence for improved joint decision-making.

Conclusions:

• RCMCL offers a superior approach to multiview learning by effectively managing decision uncertainty and conflicting instances.
• The method provides a more reliable and robust solution for applications requiring high accuracy and trustworthiness.
• The successful validation on benchmark datasets confirms the practical efficacy of RCMCL.