Deep Spectral Representation Learning From Multi-View Data.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Multi-view representation learning (MvRL) seeks consensus from diverse data sources.
• Shallow models in existing MvRL methods often yield suboptimal results, particularly in unsupervised settings.
• Deep learning offers greater representative capacity but integrating it with spectral methods is challenging.

Purpose of the Study:

• To propose a novel deep unsupervised multi-view representation learning method, Multi-view Laplacian Network (MvLNet).
• To address the limitations of shallow models in MvRL by leveraging deep learning.
• To develop the first deep version of multi-view spectral representation learning.

Main Methods:

• MvLNet integrates deep learning with multi-view spectral representation learning.
• An orthogonal constraint, reformulated via Cholesky decomposition, is introduced to prevent trivial solutions.
• An orthogonal layer is incorporated into the embedding network to learn a common space for consensus representation.

Main Results:

• Extensive experiments were conducted on seven challenging datasets.
• MvLNet demonstrated effectiveness across three multi-view tasks: clustering, recognition, and retrieval.
• The proposed method outperforms numerous recent approaches.

Conclusions:

• MvLNet provides an effective deep unsupervised approach for multi-view representation learning.
• The method successfully learns consensus representations for various downstream tasks.
• The source code is publicly available for further research and application.