Read the original document by opening this link in a new tab.
Table of Contents
Abstract
1 Introduction
2 How the vector inputs and outputs of a capsule are computed
3 Margin loss for digit existence
4 CapsNet architecture
4.1 Reconstruction as a regularization method
5 Capsules on MNIST
5.1 What the individual dimensions of a capsule represent
5.2 Robustness to Affine Transformations
6 Segmenting highly overlapping digits
Summary
A capsule is a group of neurons whose activity vector represents the instantiation parameters of a specific type of entity such as an object or an object part. Dynamic routing mechanism is used to ensure proper routing of capsule outputs to appropriate parents in higher levels. The paper introduces CapsNet architecture with convolutional capsules and demonstrates state-of-the-art performance on MNIST dataset. The CapsNet model incorporates routing-by-agreement to improve recognition accuracy, surpassing traditional convolutional networks in handling overlapping digits. The paper also discusses margin loss for digit existence, reconstruction as a regularization method, and the interpretation of individual dimensions of digit capsules. CapsNet shows robustness to affine transformations, outperforming traditional models on affNIST dataset. Additionally, the paper addresses segmenting highly overlapping digits using dynamic routing.