So far, we have discussed metric-based learning, which uses a special similarity measure (which is hard to overfit) to adapt the representational power of NNs with the ability to learn from datasets with few training samples. Alternatively, model-based approaches rely on improved network architectures (for example, memory augmented networks) to solve the same issue. In this section, we'll discuss optimization-based approaches, which adjust the training framework to adapt to the few-shot learning requirements. More specifically, we'll focus on a particular algorithm called model-agnostic meta learning (MAML; Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks, https://arxiv.org/abs/1703.03400). As the name suggests, MAML can be applied over any learning problem and model that is trained with gradient descent.
To quote the original...
