Autoencoder in Autoencoder Networks | JoVE Visualize

Area of Science:

Machine Learning
Data Science
Artificial Intelligence

Background:

Modeling complex correlations in high-dimensional multiview data presents significant challenges, particularly with noisy features.
Existing methods struggle to effectively integrate information from diverse data sources while handling high dimensionality.

Purpose of the Study:

To propose a novel unsupervised multiview representation learning (UMRL) algorithm, Autoencoder in Autoencoder Networks (AE2-Nets), to address challenges in modeling complex correlations.
To develop a framework capable of encoding high-dimensional, heterogeneous data into a compact and informative representation.

Main Methods:

The proposed AE2-Nets utilize a bidirectional encoding strategy: inner autoencoder networks for view-specific information (forward encoding) and outer autoencoder networks for integrating information across views (backward encoding).
A probabilistic explanation and extension from hierarchical variational autoencoder are provided for the nested architecture.
The forward-backward strategy is designed to handle high-dimensional and noisy features within views and capture complementarity across multiple views.

Main Results:

AE2-Nets demonstrated effective encoding of information from high-dimensional heterogeneous data into a compact representation.
The bidirectional encoding strategy successfully addressed noisy features and integrated complementary information across multiple views.
Extensive results on benchmark datasets validated the superiority of AE2-Nets over state-of-the-art algorithms.

Conclusions:

AE2-Nets offer a powerful and flexible solution for unsupervised multiview representation learning, particularly for complex, high-dimensional datasets.
The proposed framework effectively models intricate correlations and extracts informative representations by leveraging a novel bidirectional encoding strategy.
The method shows significant advantages over existing approaches, paving the way for improved analysis of multiview data.