In Chapter 9, Emerging Neural Network Designs, we introduced new neural network (NN) architectures to tackle some of the limitations of existing deep learning (DL) algorithms. We discussed graph neural networks that are used to process structured data, represented as graphs. We also introduced memory augmented neural networks, which allow networks to use external memory. In this chapter, we'll look at how to improve DL algorithms by giving them the ability to learn more information using fewer training samples.
Let's illustrate this problem with an example. Imagine that a person has never seen a certain type of object, say a car (I know—highly unlikely). They will only need to see a car once to be able to recognize other cars as well. But this is not the case with DL algorithms. A DNN needs a lot of training samples (and sometimes data augmentation...
