Hyperbolic Prototype-Residual Autoencoder for Interpretable Generative Latent Organization

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Generative models often lack interpretable latent space organization.
• Continuous latent variables in current models hinder semantic control and sample organization.
• Existing methods struggle to provide explicit structure for understanding generative processes.

Purpose of the Study:

• To develop a generative model with an interpretable and structured latent space.
• To enable explicit control over semantic variations in generated data.
• To propose a novel framework for organizing latent representations using hyperbolic geometry.

Main Methods:

• Proposed a hyperbolic prototype-residual autoencoder utilizing a fixed prototype tree in the Poincaré ball.
• Encoded samples are assigned to prototypes via hyperbolic distance.
• Decoder reconstructs images from prototype-residual representations, combining semantic anchors with instance-level variations.

Main Results:

• Demonstrated interpretable coarse-to-fine behavior in image reconstruction and generation.
• Showcased effective prototype-guided generation and reconstruction on MNIST and CelebA datasets.
• Validated the framework's ability to align encoder-decoder behavior with a predefined hyperbolic hierarchy.

Conclusions:

• Fixed hyperbolic prototype trees offer an effective scaffold for organizing generative latent spaces.
• The proposed autoencoder facilitates interpretable control over image generation and reconstruction.
• This approach advances the field of interpretable generative modeling.