Group Equivariant Capsule Networks

Authors: Jan Eric Lenssen, Matthias Fey, Pascal Libuschewski

NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Last, we apply this combined architecture as proof of concept application of our framework to MNIST datasets and verify the properties experimentally.
Researcher Affiliation Academia Jan Eric Lenssen Matthias Fey Pascal Libuschewski TU Dortmund University Computer Graphics Group 44227 Dortmund, Germany {janeric.lenssen, matthias.fey, pascal.libuschewski}@udo.edu
Pseudocode Yes Algorithm 1 Group capsule layer
Open Source Code Yes For further details, we refer to our implementation, which is available on Github1. 1Implementation at: https://github.com/mrjel/group_equivariant_capsules_pytorch
Open Datasets Yes MNIST datasets [Le Cun et al., 1998]... standard MNIST dataset with 50k training examples and the dedicated MNIST-rot dataset with the 10k/50k train/test split [Larochelle et al., 2007]. In addition, we replicated the experiment of Sabour et al. [2017] on the aff NIST dataset2, a modification of MNIST where small, random affine transformations are applied to the images.
Dataset Splits No The paper mentions "MNIST-rot dataset with the 10k/50k train/test split" but does not explicitly detail a validation split or how it was used.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions software components like "spline-based convolution operator" and the GitHub link implies PyTorch, but no specific version numbers for software dependencies are provided in the text.
Experiment Setup Yes Our canonical architecture consists of five capsule layers where each layer aggregates capsules from 2 2 spatial blocks with stride 2. ... The numbers of output capsules are 16, 32, 32, 64, and 10 per spatial position for each of the five capsule layers, respectively. In total, the architecture contains 235k trainable parameters (145k for the capsules and 90k for the CNN). ... We use the spread loss as proposed by Hinton et al. [2018] for the capsule part and standard cross entropy loss for the convolutional part and add them up. We trained our models for 45 epochs.