I am experimenting the capsule network, using the architecture provided here: Capsule Network in Mathematica Deep Neural Network Repository While I can figure out what the network do following the capsule paper from Hiton et al. 2017; I am really confused about the reconstructing part, especially the pick module: Capsule Network Architecture I can extract this part out. But I am wondering how this is illustrated in the paper, and what each operation is for. Thanks for the attention.

  • 1
    $\begingroup$ The construction notebook NetModel["CapsNet Trained on MNIST Data", "ConstructionNotebook"] has lots of information. It looks like the Pick layer is specially built with more low-level functionality than the rest of the network. $\endgroup$ – Carl Lange Oct 9 '18 at 6:26
  • 1
    $\begingroup$ The functions in use in that custom Pick layer are described in the MXNet NDArray API Documentation. I haven't read the paper but it seems like the functionality of the layer should be relatively clear after reading the documentation for 'broadcast_axis' and 'pick'. $\endgroup$ – Carl Lange Oct 9 '18 at 16:50
  • $\begingroup$ Can you add a link to the paper on Arxiv or similar to your question? $\endgroup$ – Carl Lange Oct 10 '18 at 9:15

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.