Subspace Capsule Network

Explore a 22-minute conference talk from AAAI 2020 presented by researchers from the University of Central Florida on Subspace Capsule Networks (SCN). Delve into the key differences between SCN and CapsNet, understand the intuition behind subspace capsules, and learn about their principles and implementation using orthogonal projection. Discover how subspace capsules function in intermediate layers and the challenges they address. Examine the proposed method, which ensures no information loss, norm preservation, and angle preservation. Investigate subspace capsule convolution using P-activation functions and review experimental results on various datasets, including semi-supervised image classification, image generation with SCN-GAN, and high-resolution image generation. Analyze qualitative and quantitative comparisons, explore interpolation in the latent space, and understand the impact of capsule size on ImageNet supervised classification performance.

Introduction
Capsule Convention
Differences of SCN and CapsNet
Subspace Capsule Intuition
Subspace Capsule Network Principles
Subspace Capsules Based on Orthogonal Projection
Subspace Capsules in Intermediate Layers
Challenge
Proposed Method
No Information Loss
Norm preserving
Angle Preserving
Subspace Capsule Convolution using P
Activation Functions
Experiments
Datasets
Semi-supervised Image Classification
Semi-supervised classification with SCN-GAN
Image Generation with SCN-GAN
High Resolution Image Generation
Qualitative Results
Quantitative Comparison
Interpolation in the Latent Space
ImageNet Supervised Classification
Last Resnet Block Architecture
Effect of Capsule Size
Summery