De Sousa Ribeiro, Fabio, Leontidis, George and Kollias, Stefanos (2020) Introducing Routing Uncertainty in Capsule Networks. In: 2020 Conference on Neural Information Processing Systems, December 6-12, 2020, Vancouver, Canada.
Full content URL: https://proceedings.neurips.cc/paper/2020/file/47f...
This is the latest version of this item.
| Documents |
| ||||||
|
PDF
NeurIPS-2020.pdf - Whole Document 3MB |
| Item Type: | Conference or Workshop contribution (Paper) |
|---|---|
| Item Status: | Live Archive |
Abstract
Rather than performing inefficient local iterative routing between adjacent capsule layers, we propose an alternative global view based on representing the inherent uncertainty in part-object assignment. In our formulation, the local routing iterations are replaced with variational inference of part-object connections in a probabilistic capsule network, leading to a significant speedup without sacrificing performance. In this way, global context is also considered when routing capsules by introducing global latent variables that have direct influence on the objective function, and are updated discriminatively in accordance with the minimum description length (MDL) principle. We focus on enhancing capsule network properties, and perform a thorough evaluation on pose-aware tasks, observing improvements in performance over previous approaches whilst being more computationally efficient.
| Keywords: | capsule neural networs, routing uncertainty, variational inference, global latent variables |
|---|---|
| Subjects: | G Mathematical and Computer Sciences > G500 Information Systems G Mathematical and Computer Sciences > G400 Computer Science G Mathematical and Computer Sciences > G760 Machine Learning G Mathematical and Computer Sciences > G730 Neural Computing |
| Divisions: | College of Science > School of Computer Science |
| ID Code: | 43047 |
| Deposited On: | 18 Nov 2020 11:48 |
Available Versions of this Item
- Introducing Routing Uncertainty in Capsule Networks. (deposited 18 Nov 2020 11:48) [Currently Displayed]
Repository Staff Only: item control page