Deep Semisupervised Multiview Learning With Increasing Views | JoVE Visualize

Area of Science:

Machine Learning
Computer Vision
Data Science

Background:

Semisupervised cross-view learning faces challenges with sample-level correspondence assumptions and handling increasing numbers of data views.
Existing methods often require pairwise sample relationships, which are difficult to establish in semisupervised settings with limited labeled data.
Current multiview learning models struggle with scalability as new data views are introduced, necessitating complete retraining.

Purpose of the Study:

To propose a novel method, Independent Semisupervised View-specific Networks (ISVNs), to overcome limitations in semisupervised cross-view learning.
To develop a flexible approach that accommodates increasing data views without full model retraining.
To address the non-corresponding problem by relaxing the assumption of pairwise sample relationships across views.

Main Methods:

Developed multiple Independent Semisupervised View-specific Networks (ISVNs) for view-decoupled representation learning.
Employed a specifically designed autoencoder and a pseudolabel learning paradigm to leverage both labeled and unlabeled data.
Implemented a view decoupling strategy enabling separate training of ISVNs for enhanced scalability.

Main Results:

The proposed ISVNs effectively utilize labeled and unlabeled data, relaxing the need for strict sample-level correspondence.
The view decoupling strategy allows for efficient handling of increasing data views without retraining the entire model.
Experimental results demonstrate the effectiveness and efficiency of ISVNs compared to 13 state-of-the-art approaches on four multiview datasets for retrieval and classification tasks.

Conclusions:

ISVNs offer a novel solution for semisupervised cross-view learning, adeptly managing non-corresponding data and increasing view numbers.
This method represents a significant advancement in handling dynamic and complex multiview datasets.
The research paves the way for more scalable and robust semisupervised learning frameworks in the presence of multiple data modalities.