SeGMA: Semi-Supervised Gaussian Mixture Autoencoder.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Semi-supervised learning models are crucial for leveraging limited labeled data.
• Generative models, particularly those based on autoencoders, are effective for data representation and generation.
• Existing models often struggle with fine-grained control over generated data attributes.

Purpose of the Study:

• To develop a novel semi-supervised generative model, SeGMA (Semi-supervised Generative Model with Attribute control).
• To enhance generative capabilities by enabling interpolation, style transfer, and attribute manipulation.
• To achieve efficient optimization using Cramer-Wold distance for improved performance.

Main Methods:

• Implemented SeGMA within a Wasserstein autoencoder framework.
• Utilized a mixture of Gaussians in latent space for data clustering.
• Employed a Gaussian classifier with limited labeled data to associate clusters with classes.
• Optimized the model using Cramer-Wold distance as a maximum mean discrepancy penalty.

Main Results:

• SeGMA achieves competitive generative performance on benchmark datasets.
• Demonstrated realistic sample interpolation in the latent space.
• Showcased continuous style transfer between classes by disentangling class and style information.
• Enabled modification of class characteristic intensity by adjusting latent representations.

Conclusions:

• SeGMA offers advanced generative capabilities beyond existing semi-supervised methods.
• The model effectively combines data and class learning for enhanced control.
• Future work can explore applications in data augmentation and creative content generation.