On the conditions of MAML convergence
On the conditions of MAML convergence
复制标题
DOI:
--
复制
发表时间:
2019
期刊:
影响因子:
--
通讯作者:
S. Takagi;Yoshihiro Nagano;Yuki Yoshida;M. Okada
中科院分区:
文献类型:
--
作者:
S. Takagi;Yoshihiro Nagano;Yuki Yoshida;M. Okada
Model-agnostic meta-learning (MAML) is known as a powerful meta-learning method. In this paper, we derive the conditions that inner learning rate (cid:11) and meta-learning rate (cid:12) must satisfy for a simplified MAML to locally converge to local minima from any point. We find that the upper bound of (cid:12) depends on (cid:11) , in contrast to the case of using the normal gradient descent method. Moreover, we show that the threshold of (cid:12) increases as (cid:11) approaches its own upper bound. This result is verified by experiments on various few-shot tasks and architectures; specifically, we perform sinusoid regression and classification of Omniglot and MiniImagenet datasets with a multilayer perceptron and a CNN. Based on this outcome, we present a guideline for determining the learning rates: first, search for the largest possible (cid:11) ; next, tune (cid:12) based on the chosen value of (cid:11) .