By searching for shared inductive biases across tasks, meta-learning promises to accelerate learning on novel tasks, but with the cost of solving a complex bilevel optimization problem. We introduce and rigorously define the trade-off between accurate modeling and optimization ease in meta-learning. At one end, classic meta-learning algorithms account for the structure of meta-learning but solve a complex optimization problem, while at the other end domain randomized search (otherwise known as joint training) ignores the structure of meta-learning and solves a single level optimization problem. Taking MAML as the representative meta-learning algorithm, we theoretically characterize the trade-off for general non-convex risk functions as well as linear regression, for which we are able to provide explicit bounds on the errors associated with modeling and optimization. We also empirically study this trade-off for meta-reinforcement learning benchmarks.

When trading off between MAML and DRS, the learning rate in MAML provides a simple and explicit mechanism to trade between these two algorithms (ignoring task dataset size). In particular, as the learning rate goes to zero, MAML becomes DRS. Thus, if proper hyperparameter tuning is performed on the learning rate, then I would expect MAML to always out-perform DRS given a reasonable number of samples. In the meta-RL benchmarks, the original MAML paper [1] reports using a learning rate of 0.1 for TRPO-MAML but the largest value evaluated in this work is 0.01. Does that change the performance?

This paper addresses a trade-off between MAML and DRS which are two opposite methods. Authors did a good job in the rebuttal which well answered most of reviewers' concerns. Two of reviewers raised their scores to 6. The paper well investigates an important question on deeper understanding of MAML and DRS for meta-learning.

By searching for shared inductive biases across tasks, meta-learning promises to accelerate learning on novel tasks, but with the cost of solving a complex bilevel optimization problem. We introduce and rigorously define the trade-off between accurate modeling and optimization ease in meta-learning. At one end, classic meta-learning algorithms account for the structure of meta-learning but solve a complex optimization problem, while at the other end domain randomized search (otherwise known as joint training) ignores the structure of meta-learning and solves a single level optimization problem. Taking MAML as the representative meta-learning algorithm, we theoretically characterize the trade-off for general non-convex risk functions as well as linear regression, for which we are able to provide explicit bounds on the errors associated with modeling and optimization. We also empirically study this trade-off for meta-reinforcement learning benchmarks.