Multi-view learning meets state-space model: A dynamical system perspective.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Control Theory

Background:

• Multi-view learning leverages multiple data modalities for improved task performance.
• Existing deep learning methods often lack principled frameworks for dynamic feature representation, hindering interpretability.
• There is a need for unified, interpretable models that capture the evolution of multi-view features.

Purpose of the Study:

• To introduce the Multi-view State-Space Model (MvSSM) for principled multi-view representation learning.
• To develop a framework that integrates feature integration and label prediction within a single, interpretable model.
• To enable theoretical analysis of system dynamics and representation transitions in multi-view learning.

Main Methods:

• Formulating multi-view representation learning as a continuous-time dynamical system inspired by control theory.
• Treating view-specific features as external inputs and a shared latent representation as the evolving system state.
• Developing MvSSM variants (MvSSM-Lap, MvSSM-iLap) using Laplace and inverse Laplace transformations for system dynamics.

Main Results:

• The MvSSM framework unifies feature integration and label prediction, allowing for theoretical analysis.
• MvSSM variants show structural similarities to graph convolutions, facilitating efficient feature propagation.
• Experimental results on IAPR-TC12 and ESP datasets demonstrate significant performance gains over state-of-the-art methods.

Conclusions:

• The MvSSM provides a theoretically grounded and interpretable approach to multi-view representation learning.
• The proposed dynamical system framework enhances the modeling of feature evolution and cross-view interactions.
• MvSSM achieves superior performance, highlighting its potential for advancing multi-view learning research.