Centrala begrepp
Capsule Networks can achieve efficiency and performance with a non-iterative routing mechanism.
Sammanfattning
Capsule Networks offer robust performance with fewer parameters than CNNs. The slow iterative routing mechanisms hinder scalability. ProtoCaps introduces a non-iterative routing method inspired by trainable prototype clustering to enhance computational efficiency. By utilizing a shared Capsule subspace, memory requirements during training are significantly reduced. ProtoCaps outperforms current non-iterative Capsule Networks on the Imagewoof dataset, showcasing its potential for complex computational scenarios.
Statistik
1m FLOPs (in Millions)
Test Accuracy: 99.5% for ProtoCaps (ours) on MNIST dataset
Test Accuracy: 92.5% for ProtoCaps (ours) on FashionMNIST dataset
Test Accuracy: 59.0% for ProtoCaps (ours) on Imagewoof dataset
Citat
"We propose a novel, non-iterative, trainable routing algorithm for Capsule Networks."
"Our approach demonstrates superior results compared to the current best non-iterative Capsule Network."
"ProtoCaps substantially mitigates the memory consumption issue and provides an effective, efficient and scalable routing mechanism."