A Receptor Skeleton for Capsule Neural Networks
Authors: Jintai Chen, Hongyun Yu, Chengde Qian, Danny Z Chen, Jian Wu
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive experiments verify that our approach facilitates efficient clustering processes, and Caps Nets with our approach significantly outperform Caps Nets with previous routing algorithms on image classification, affine transformation generalization, overlapped object recognition, and representation semantic decoupling. |
| Researcher Affiliation | Academia | 1College of Computer Science and Technology, Zhejiang University, Hangzhou, China; 2School of Statistics and Data Science, Nankai University, China; 3Department of Computer Science and Engineering, University of Notre Dame, Notre Dame, IN 46556, USA; 4The First Affiliated Hospital, and Department of Public Health, Zhejiang University School of Medicine, Hangzhou, China; |
| Pseudocode | Yes | Algorithm 1 A procedure for synthesizing the j-th parent capsule s representation by our proposed approach. |
| Open Source Code | No | The paper states that the performances of *known approaches* are obtained by running their open source codes, but does not provide open-source code for its *own* methodology. |
| Open Datasets | Yes | We generate the Multi-MNIST dataset similarly as in (Sabour et al., 2017)...For the Small Norb dataset, we follow the experimental design as in (Ribeiro et al., 2020b)...The JSRT dataset contains 247 chest radiographs, and segmentation annotations...from (Shiraishi et al., 2000). |
| Dataset Splits | No | The paper mentions "official train-test splits" for some datasets but does not explicitly specify details for a separate validation split or its size/methodology. |
| Hardware Specification | Yes | The models are trained and tested on Ge Force RTX 3090 Ti GPUs. |
| Software Dependencies | No | The paper mentions "We use the Py Torch library to implement our approach" but does not specify a version number for PyTorch or any other software dependency. |
| Experiment Setup | Yes | We let K = 6 for the K nearest neighbor grouping...and the number of receptors per capsule M = 2. ... we use the SGD optimizer with weight decay 5e-4 and momentum 0.9... We run 300 epochs, and the initial learning rate is 0.1, which is reduced by 10 at the 60-th, 120-th, and 160th epochs. ... in computing GPU memory, batch size = 64, capsule number = 40, capsule size = 16, width = 8, and height = 8; in computing fps, batch size = 1. |