Multiview Representation Learning via Information-Theoretic Optimization.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Multiview data offers rich features but poses challenges in feature extraction and integration within multiview learning (MVL).
• Traditional deep learning methods often embed class features implicitly, hindering explicit mapping to principal subspaces and potentially causing class confusion.
• Existing approaches lack structured feature organization, impacting the accurate capture of intrinsic data structures.

Purpose of the Study:

• To develop an innovative multiview representation learning (MVRL) approach.
• To address the limitations of traditional MVL methods in feature extraction and integration.
• To enhance the performance and generalization capabilities of MVL systems.

Main Methods:

• Introduced a novel MVRL method maximizing intraview coding rate reduction and inter-view mutual information.
• Optimized intraview representations by maximizing coding rate differences for compressed, distinct class features.
• Achieved cross-view alignment and fusion via space transformation and cross-sample fusion, maximizing information transmission for consistent representations.

Main Results:

• The proposed MVRL method demonstrated excellent performance in experiments.
• Achieved more compact and distinct feature representations within each data view.
• Enhanced consistency and correlation among data representations across multiple views.

Conclusions:

• The novel MVRL method effectively extracts intraview features and integrates inter-view information.
• Maximizing mutual information between consensus and view-specific representations leads to concise intrinsic features and improved MVL performance.
• The approach enhances the accuracy and generalization ability of machine learning models using multiview data.