In this chapter, we learned about metrics-based, one-shot learning methods. We explored two neural network architectures that have been used for one-shot learning in both the research community and software industry as well. We also learned how to evaluate trained models. Then, we executed an exercise in Siamese networks using the MNIST dataset. In conclusion, we can say that both the matching networks and Siamese network architectures have successfully proven that by changing the loss function or feature representation, we can achieve our objective with a limited amount of data.
In the next chapter, we will be exploring different optimization-based methods and learn how they differ from metrics-based methods.