In the last chapter, we discussed two optimization-based methods. We attempted to train models with a learn to learn mechanism, similar to what is seen in humans. Of course, apart from the ability to learn new things, humans also have access to a large amount of memory when performing any task. This enables us to learn a new task more easily by recalling past memories and experiences. Following the same thought process, model-based architecture is designed with the addition of external memory for the rapid generalization of one-shot learning tasks. In these approaches, models converge with only a few training steps using information stored in external memory.
The following topics will be covered in this chapter:
- Understanding Neural Turing Machines
- Memory-augmented neural networks
- Meta networks
- Coding exercises
