1) Aurélien Géron’s tutorial
- If you want to watch just one video on capsule networks, make it this one. I found myself taking four pages of notes while watching Aurélien Géron’s tutorial.
- The video starts with an explanation of how capsule networks (CapsNets) attempt to perform inverse computer graphics. It then uses a recurring example involving a house and a sailboat to develop the intuition and the mechanics in a step-by-step, easy-to-follow fashion.
- Geoffrey Hinton’s comment says it all: “This is an amazingly good video.”
2) Charles Martin’s overview
- In many ways, this overview provides a link between the two other videos.
- It helps explains Hinton’s critique of max-pooling and mentions some of the neuroscientific ideas and findings that he drew upon as inspiration for capsule networks, such as cortical minicolumns and the two-streams hypothesis.
- Other highlights include an introduction to a Keras implementation and a discussion of how the basic elements of CapsNets compare to those of classic feed-forward networks.
3) Geoffrey Hinton’s talk
- I would recommend to watch this video last. The knowledge you will gain from the other two videos will help you better understand and enjoy Geoffrey Hintons’s presentation.
- The talk starts with a critique of existing neural net architectures and provides four arguments against the max-pooling operation that is commonly used in convolutional neural networks. Along the way, we learn what the difficulty we experience in putting together two pieces of a tetrahedron can tell us about our visual system.
- The second half of the talk explains how the CapsNet architecture is designed to force the network to learn a hierarchy of the object parts, without manually engineering the parts.
Learn more about capsule networks at aisummary.com.
Source: Deep Learning on Medium