[论文解读] Alpha MAML: Adaptive Model-Agnostic Meta-Learning

Harkirat Behl, Atılım Güneş Baydin+1 more|arXiv (Cornell University)|May 17, 2019

Domain Adaptation and Few-Shot Learning参考文献 18被引用 54

一句话总结

Alpha MAML 通过使用超梯度下降的在线超参数自适应，扩展了 MAML，以减少学习率的调优并提高 Omniglot 上少样本任务的训练稳定性。

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.

研究动机与目标

• Motivate reducing the heavy hyperparameter tuning required by MAML in few-shot learning.
• Introduce Alpha MAML as an online adaptive scheme for learning rates in both inner and meta updates.
• Show that adaptive hyperparameters improve training stability and reduce sensitivity to initial values.
• Demonstrate robustness and faster convergence on Omniglot compared to standard MAML.

提出的方法

• Build on MAML by introducing online updates for both inner learning rate alpha and meta learning rate beta.
• Derive update rules for alpha and beta using hypergradient descent applied to the meta-objective.
• Show that alpha and beta updates can reuse existing gradients, avoiding extra gradient computations.
• Provide four update equations that couple theta, alpha, beta, and their hypergradients.
• Extend to multi-task batches with the corresponding general update derivations in the appendix.

实验结果

研究问题

• RQ1 Can online hypergradient descent adapt inner and meta learning rates during MAML training to reduce sensitivity to initial hyperparameters?
• RQ2 Does Alpha MAML achieve stable convergence and require less hyperparameter tuning than standard MAML on few-shot tasks?
• RQ3 How does Alpha MAML perform on Omniglot compared to MAML in terms of convergence and robustness to hyperparameters?

主要发现

• Alpha MAML automatically tunes alpha and beta online during training.
• The method reduces sensitivity to initial hyperparameter choices compared to MAML.
• Alpha MAML demonstrates improved training stability and convergence on Omniglot over a range of initial values.
• Grid-search experiments indicate Alpha MAML converges under hyperparameter settings where MAML fails to converge within fixed iterations.