Group Equivariant Capsule Networks

Abstract

We present group equivariant capsule networks, a framework to introduce guaranteed equivariance and invariance properties to the capsule network idea. Our work can be divided into two contributions. First, we present a generic routing by agreement algorithm defined on elements of a group and prove that equivariance of output pose vectors, as well as invariance of output activations, hold under certain conditions. Second, we connect the resulting equivariant capsule networks with work from the field of group convolutional networks. Through this connection, we provide intuitions of how both methods relate and are able to combine the strengths of both approaches in one deep neural network architecture. The resulting framework allows sparse evaluation of the group convolution operator, provides control over specific equivariance and invariance properties, and can use routing by agreement instead of pooling operations. In addition, it is able to provide interpretable and equivariant representation vectors as output capsules, which disentangle evidence of object existence from its pose.

arXiv commentPresented at NIPS 2018

Cited by in corpus (18)

• Stacked Capsule Autoencoders
• Attentive Group Equivariant Convolutional Networks
• Capsule Networks Need an Improved Routing Algorithm
• You Only Hypothesize Once: Point Cloud Registration with Rotation-equivariant Descriptors
• B-Spline CNNs on Lie Groups
• Coordinate Independent Convolutional Networks -- Isometry and Gauge Equivariant Convolutions on Riemannian Manifolds
• Quaternion Equivariant Capsule Networks for 3D Point Clouds
• Self-Supervised Multi-View Learning via Auto-Encoding 3D Transformations
• Improving the Robustness of Capsule Networks to Image Affine Transformations
• Improved Generalization Bounds of Group Invariant / Equivariant Deep Networks via Quotient Feature Spaces
• Group Equivariant Stand-Alone Self-Attention For Vision
• A Neural-Symbolic Architecture for Inverse Graphics Improved by Lifelong Meta-Learning
• GraphTER: Unsupervised Learning of Graph Transformation Equivariant Representations via Auto-Encoding Node-wise Transformations
• Capsule network with shortcut routing
• No Representation without Transformation
• Finite Group Equivariant Neural Networks for Games
• Autoequivariant Network Search via Group Decomposition
• Equivariant Point Network for 3D Point Cloud Analysis