November 2018
Intermediate to advanced
322 pages
7h 54m
English
The structure of the decoder is shown in the following diagram:

The decoder essentially tries to reconstruct the image from the correct DigitCaps capsule for each image. You can view this as a regularization step, with loss being the Euclidean distance between the predicted output and the original label. You could argue that you don't require reconstruction in this application as you are just carrying out classification. However, Hinton specifically shows in his original paper that adding reconstruction loss does improve the accuracy of the model.
The decoder's structure is pretty simple, and consists of only three ...
Read now
Unlock full access