Matrix capsules with EM routing

[changkun]

• PDF: https://openreview.net/pdf?id=HJWLfGWRb
• Code: https://github.com/www0wwwjs1/Matrix-Capsules-EM-Tensorflow/ (open-source-impl)
• Reason:
  • ICLR 2018 accepted paper, Hinton's paper
  • bring EM algorithm and conducted a better routing mechanism than original Dynamic Routing
  • impair SGD and loss function plays the most important role in neural network
  • generic definition of Capsule in Deep Learning
• Discussion date: 12th March

The capsule network is still in the very beginning, and the original accepted paper and its open source code from the author are inconsistent (see Sarasra/models, "no routing is used between Conv1 and PrimaryCapsules" from paper, but the code uses routing), we should pay attention whether it is valuable to discuss it or not. The most interesting idea from CapsNet its self-explanable and self-evident, it brings the idea of built-in interpretability for building deep models.

[changkun]

Material for understanding: capsule blue print, https://docs.google.com/presentation/d/1ccEsDtqjyQVvEvF-MDfwwnXqht5mPRE3c800vs5yJNY/edit?usp=drivesdk

Feel free to make a comment :)