OpenMvL: An Attribute-Inspired Dual-Head Framework for Open-Set Multiview Learning

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Existing multiview learning methods struggle in open-set scenarios with unknown classes.
• Intra-view heterogeneity and inter-view conflicts hinder performance in real-world applications.

Purpose of the Study:

• To propose a novel attribute-inspired dual-head framework for open-set multiview learning.
• To address intra-view heterogeneity and inter-view conflicts in multiview data.

Main Methods:

• Developed a multiview attribute-inspired open-set network with a deep unfolding layer for intra-view conflict resolution.
• Employed an evidential fusion approach to manage inter-view conflicts by quantifying global uncertainty.
• Designed a dual-head perception-adaptive open-set training loss with Softmax and Evidence heads.

Main Results:

• The proposed framework effectively resolves intra-view conflicts, reducing heterogeneity and improving topological representations.
• Evidential fusion successfully tackles inter-view conflicts by selectively integrating complementary features.
• The dual-head training loss enhances representation learning and calibrates uncertainty for better discrimination.

Conclusions:

• The novel framework demonstrates outstanding performance on challenging open-set multiview tasks.
• The method effectively adapts to both known and unknown samples, showcasing its robustness.
• This work advances open-set multiview learning by addressing critical real-world challenges.