개발하며 Nolsigan

Dynamic Routing Between Capsules

November 12, 2017 | 0 Minute Read

Introduces new concept called capsule whose output represents the instantiation parameters of a specific type of entity such as an object or object part. Output’s orientation represents its instantiation parameters and length represents its probability.

Key Points

  • Lower level capsule influences higher level capsule if two capsule’s outputs agrees
  • Agreement is caculated by scalar product of outputs
  • Uses shrinking algorithm to make output vector size from 0 to 1
  • Total loss is sum of each class’s loss
  • Because there’s no max pooling, capsule network can capture not only the most significant object but also others
  • Even two digits highly overlapped
  • But also generates drawback that images with high variance background are hard to classify
  • Network depends more on forward propagation through routing-by-agreement
  • Reduce burden for backpropagation to do the whole learning job

Thoughts

  • Nice, interesting, novel method
  • Hard to predict if this will substitute CNN completely
  • But giving more responsibility for making prediction to dynamic forward propagation seems like a good direction of progress
  • But takes longer time to train