Capsule Networks Quiz

Answers and explanations appear as you take the quiz.

1. In a capsule network, what does the LENGTH of a capsule's output vector represent?
   • The probability that the entity is present
   • The color of the detected feature
   • The learning rate for that layer
   • The number of neurons in the capsule
2. What problem with standard CNNs were capsule networks specifically designed to address?
   • Slow training on GPUs
   • Loss of spatial/part-whole relationships, e.g. accepting jumbled face parts
   • Inability to process color images
   • Too few parameters to learn
3. What is the name of the algorithm that decides which lower-level capsules connect to which higher-level capsules?
   • Backpropagation through time
   • Gradient clipping
   • Dynamic routing-by-agreement
   • Batch normalization
4. Who introduced the capsule network with dynamic routing in 2017?
   • Yann LeCun and colleagues
   • Ian Goodfellow and colleagues
   • Vaswani and colleagues
   • Sabour, Frosst, and Hinton
5. What is the purpose of the 'squash' function in a capsule network?
   • To normalize vector length to between 0 and 1 while keeping its direction
   • To delete unused capsules
   • To convert images to grayscale
   • To increase the network's depth automatically
6. Which result did the original 2017 CapsNet paper report as evidence for routing-by-agreement?
   • Lower memory use than every CNN on ImageNet
   • Better recognition of highly overlapping digits than a convolutional net
   • No need for a training objective
   • Perfect performance on all affine transformations
7. In the original capsule representation, what information is carried by the direction of a capsule’s output vector?
   • The probability that the entity is absent
   • The number of routing iterations
   • Instantiation or pose parameters of the represented entity
   • The loss assigned to the previous layer
8. In routing-by-agreement, what makes a lower capsule’s prediction agree with a higher-level capsule?
   • Their capsule names begin with the same letter
   • Their vector lengths are exactly zero
   • They occur in the same image crop
   • Their predicted and output vectors have a large scalar product