Capsule Networks -- A Probabilistic Perspective

TL;DR

This paper presents a probabilistic generative model for capsule networks that explicitly encodes object pose relationships, aiming to improve robustness to viewpoint changes and providing insights into inference failures.

Contribution

It introduces a probabilistic perspective on capsule networks, separating generative modeling from inference, and explores test-time optimization to address inference challenges.

Findings

The probabilistic model clarifies capsule assumptions and their limitations.

Test-time optimization improves inference quality.

The unified objective demonstrates practical applicability.

Abstract

'Capsule' models try to explicitly represent the poses of objects, enforcing a linear relationship between an object's pose and that of its constituent parts. This modelling assumption should lead to robustness to viewpoint changes since the sub-object/super-object relationships are invariant to the poses of the object. We describe a probabilistic generative model which encodes such capsule assumptions, clearly separating the generative parts of the model from the inference mechanisms. With a variational bound we explore the properties of the generative model independently of the approximate inference scheme, and gain insights into failures of the capsule assumptions and inference amortisation. We experimentally demonstrate the applicability of our unified objective, and demonstrate the use of test time optimisation to solve problems inherent to amortised inference in our model.