When MAML Learns Quickly, Does It Generalize Well?
O.T. Turan (TU Delft - Electrical Engineering, Mathematics and Computer Science)
D.M.J. Tax (TU Delft - Electrical Engineering, Mathematics and Computer Science)
M. Loog (TU Delft - Electrical Engineering, Mathematics and Computer Science)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Model-Agnostic Meta-Learning is a meta-learning method that achieved state-of-the-art performance on few-shot image classification benchmarks at the time of its introduction. MAML’s strength is its ability to quickly adapt to a new task in time-critical settings, while still being general enough for any gradient-based model. Although there is no need for quick adaptation in most of the few-shot learning benchmarks, often MAML is utilized as a benchmark in this context. We investigate the benefit of limiting the adaptation steps of MAML in settings where quick adaptation is not required by the problem. In this initial study,
the expected performance of MAML is compared to some conventional base learners for synthetic linear and nonlinear regression problems. Our experimental results show that limited gradient descent steps only improve generalization performance when faced with small task variance.