Java Deep Learning Projects

Capsule networks (CapsNet)

As discussed in Chapter 1, Getting Started with Deep Learning, CNNs perform well at classifying good quality images. However, if the images have rotation, tilt, or any other different orientation, CNNs give very poor performance. Even the pooling operation in CNNs cannot help much with positional invariance.

To overcome the this limitation of CNN, Geoffrey Hinton et al. come up with a ground breaking idea called CapsuleNetworks (CapsNet) that are particularly good at handling different types of visual stimulus and encoding things such as pose (position, size, and orientation), deformation, velocity, albedo, hue, texture.

In a regular DNN, we keep on adding layers (more layers means a deeper network). In CapsNet, ...