Meta learning by optimization

This chapter introduces the optimization-based approaches to meta learning, which model the inner loop of meta learning as solving an optimization problem. The central observation is that the vanilla stochastic gradient descent algorithm is unsuitable for optimizing learning models under data scarcity constraints. Optimization-based meta learning algorithms address this limitation by seeking effective update rules or initialization that allows efficient adaptation to novel tasks with few training samples. We also discuss the impact of inner loop optimization on meta learning performance.