Learning Extremal Representations with Deep Archetypal Analysis.

Area of Science:

• Machine Learning
• Data Analysis
• Computer Vision

Background:

• Traditional linear archetypal analysis assumes data additivity, limiting its application to complex data like images.
• Existing methods struggle with learning representations and archetypes simultaneously.

Purpose of the Study:

• To develop a generative archetypal analysis model capable of learning latent feature spaces.
• To enable end-to-end learning of optimal representations and archetypes.
• To allow for interpretable archetypes by integrating side information.

Main Methods:

• A generative formulation of the linear archetype model using neural networks.
• Integration into a deep variational information bottleneck framework.
• Introduction of a distance-dependent archetype loss function.

Main Results:

• Successfully learned interpretable archetypes for female facial expressions using emotion scores as side information.
• Demonstrated archetype exploration in the chemical space of small organic molecules.
• Showcased how side information influences archetype interpretation.

Conclusions:

• The proposed method effectively learns interpretable archetypes in latent spaces.
• The framework accommodates complex side information for enhanced archetype understanding.
• This approach offers a flexible and powerful tool for data exploration and representation.