Inference and Learning for Generative Capsule Models

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Capsule networks aim to represent hierarchical relationships between objects and their parts.
• Existing methods for inferring object-part relationships and transformations can be computationally intensive or limited in scope.

Purpose of the Study:

• To develop a generative model for capsule network data.
• To derive variational and RANSAC-based algorithms for inferring object transformations and part assignments.
• To create a learning algorithm for object models using variational expectation maximization.

Main Methods:

• Specification of a generative model for capsule network data.
• Derivation of a variational inference algorithm for object transformation and part assignment.
• Development of a variational expectation maximization learning algorithm for object models.
• Application of RANSAC (Random Sample Consensus) for an alternative inference method.

Main Results:

• Successful inference of object transformations and part assignments on synthetic geometric data (constellations) and face data.
• Demonstrated significant performance improvement over amortized inference methods (Kosiorek et al., 2019) on constellation data.
• Validation of the generative model and inference algorithms on complex datasets.

Conclusions:

• The proposed generative model and variational inference algorithms offer an effective approach for reasoning about object-part relationships in capsule networks.
• The developed methods provide a more robust and accurate way to infer object transformations and part assignments compared to prior work.
• This research advances the capabilities of capsule networks in scene understanding and object recognition.