Correlational Neural Networks.

Area of Science:

• Machine Learning
• Data Science
• Natural Language Processing

Background:

• Common representation learning (CRL) embeds data from multiple views into a shared subspace.
• Existing CRL methods include canonical correlation analysis (CCA) and autoencoders (AEs).
• CCA excels at transfer learning but lacks scalability, while AEs are scalable but may not learn maximally correlated representations.

Purpose of the Study:

• To propose a novel AE-based approach, correlational neural network (CorrNet), for CRL.
• To explicitly maximize the correlation among different data views within the common subspace.
• To evaluate CorrNet's effectiveness in learning correlated representations and its performance on cross-language tasks.

Main Methods:

• Developed CorrNet, an AE-based model incorporating a correlation maximization objective.
• Trained and evaluated CorrNet on various datasets for representation learning.
• Compared CorrNet against traditional AE and CCA methods.

Main Results:

• CorrNet demonstrated superior ability in learning correlated common representations compared to AE and CCA.
• Representations learned by CorrNet showed improved performance on cross-language tasks.
• The proposed method offers a scalable and effective approach to CRL.

Conclusions:

• CorrNet effectively learns correlated common representations by combining AE scalability with explicit correlation maximization.
• The proposed model outperforms existing state-of-the-art methods in cross-language applications.
• CorrNet presents a promising advancement in common representation learning.