Multiview Orthonormalized Partial Least Squares: Regularizations and Deep Extensions.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Multiview learning aims to integrate information from multiple data sources.
• Existing methods often struggle with high-dimensional and heterogeneous data.
• A unified framework for subspace-based multiview learning is needed.

Purpose of the Study:

• To propose a novel unified multiview learning framework based on least squares.
• To introduce regularization techniques to enhance the framework's robustness and performance.
• To develop nonlinear extensions for improved feature extraction and cross-modal retrieval.

Main Methods:

• Developed multiview orthonormalized partial least squares (MvOPLSs) for a common latent space.
• Applied regularization on model parameters, decision values, and latent projected points.
• Incorporated deep networks for learning nonlinear transformations.

Main Results:

• The proposed MvOPLSs framework effectively learns a common latent space for multiview data.
• Regularization techniques successfully integrated existing and novel multiview learning models.
• Nonlinear extensions significantly boosted performance in feature extraction and cross-modal retrieval.

Conclusions:

• Subspace-based learning in a common latent space is a powerful approach for multiview data.
• Nonlinear extensions, particularly with deep networks, offer substantial performance gains.
• One proposed nonlinear method outperformed all existing methods in experiments.