Deep Learning Essentials

Capsule networks

Capsule networks are one of the recent developments in deep learning. They are meant to address the existing limitations of the Convolution Neural Network (CNN). CNNs have shown a remarkable ability to learn image features that are invariant to orientation and spatial changes. However, a core representation of an object in CNN is simply an unordered pooling of such invariant features. It has no understanding of the relative spatial relationship among these features. For example, a CNN trained on recognizing human faces will still detect a synthetic image of a distorted human face (facial features such as nose and eyes in wrong places) as long as the facial features are present in the image. It does so because all the trained ...