6th ICML Workshop on Automated Machine Learning (2019)
Alpha MAML: Adaptive Model-Agnostic Meta-Learning
Harkirat Singh Behl harkirat@robots.ox.ac.uk
Atılım G¨une¸s Baydin gunes@robots.ox.ac.uk
Philip H.S. Torr phst@robots.ox.ac.uk
University of Oxford
Abstract
Model-agnostic meta-learning (MAML) is a meta-learning technique to train a model on a
multitude of learning tasks in a way that primes the model for few-shot learning of new
tasks. The MAML algorithm performs well on few-shot learning problems in classification,
regression, and fine-tuning of policy gradients in reinforcement learning, but comes with the
need for costly hyperparameter tuning for training stability. We address this shortcoming
by introducing an extension to MAML, called Alpha MAML, to incorporate an online
hyperparameter adaptation scheme that eliminates the need to tune meta-learning and
learning rates. Our results with the Omniglot database demonstrate a substantial reduction
in the need to tune MAML training hyperparameters and improvement to training stability
with less sensitivity to hyperparameter choice.
1. Introduction
Meta-learning—or “learning to learn”—concerns machine learning models that can improve
their learning quality by altering aspects of the learning process such as the model architecture,
optimization rules, initialization, or learning hyperparameters (Thrun and Pratt, 2012;
Schmidhuber, 1987; Hochreiter et al., 2001). An important application of meta-learning is in
few-shot learning problems (Vinyals et al., 2016; Behl et al., 2018), where one is concerned
with developing methods able to learn new concepts from one or only a few instances (Lake
et al., 2015). In this paper we focus on the state-of-the-art model-agnostic meta-learning
(MAML) (Finn et al., 2017) method, which is a conceptually simple and general algorithm
that has been shown to outperform existing approaches in tasks including few-shot image
classification and few-shot adaptation in reinforcement learning (Antoniou et al., 2019).
MAML aims to solve the few-shot learning problem by being just few gradient descent
steps away from any new concepts, doing so by making the assumption that learning a new
concept will just involve few parameter updates (Algorithm 1). In other words, MAML is
based on learning an initial representation that can be efficiently fine-tuned for new tasks in
a few steps.
The generality of MAML comes with the difficulty of choosing hyperparameters to
achieve stable training in practice (Antoniou et al., 2019). MAML has two important
hyper-parameters, namely the learning rate
α
and the meta-learning rate
β
, thus increasing
any hyperparameter grid search computation by an order, and making it significantly more
time and resource consuming than comparable methods. Another complication to this
problem is the fact that it is currently not established whether the technique can benefit
from a conventional decaying schedule for the inner learning rate
α
. Furthermore, a good
value of α in MAML is even more important than for any conventional stochastic gradient
descent (SGD) optimization, because only a handful of samples are available in the few-shot
learning case. This has significant consequences, making it difficult to scale this algorithm
c
2019 H.S. Behl, A.G. Baydin, and P.H.S. Torr.
arXiv:1905.07435v1 [cs.LG] 17 May 2019
剩余9页未读，继续阅读
评论0