Reproduce “ProtoCaps: A Fast and Non-Iterative Capsule Network Routing Method” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Miles Everett, Mingjun Zhong, Georgios Leontidis

ProtoCaps introduces a fast, non-iterative routing method for Capsule Networks that employs trainable prototype clustering and a shared Capsule subspace to reduce computational cost and memory usage. The approach enables Capsule Networks to scale effectively to challenging datasets like Imagewoof, outperforming prior non-iterative methods in efficiency and accuracy.

Key insights

  • Non-iterative routing via trainable prototype clustering within a shared subspace effectively maintains Capsule Network performance while drastically improving computational efficiency and memory footprint.
  • The inclusion of residual connections is crucial for stable and effective training in deeper Capsule Networks, especially on more complex datasets like CIFAR10 and Imagewoof.
  • Increasing the depth of convolutional Capsule layers beyond 2-3 layers does not yield significant performance improvements but also does not cause degradation, suggesting a stable architecture that reaches saturation efficiently.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.