Meta-Hyperband: Hyperparameter optimization with meta-learning and Coarse-to-Fine

Abstract

Hyperparameter optimization is one of the main pillars of machine learning algorithms. In this paper, we introduce Meta-Hyperband: a Hyperband based algorithm that improves the hyperparameter optimization by adding levels of exploitation. Unlike Hyperband method, which is a pure exploration bandit-based approach for hyperparameter optimization, our meta approach generates a trade-off between exploration and exploitation by combining the Hyperband method with meta-learning and Coarse-to-Fine modules. We analyze the performance of Meta-Hyperband on various datasets to tune the hyperparameters of CNN and SVM. The experiments indicate that in many cases Meta-Hyperband can discover hyperparameter configurations with higher quality than Hyperband, using similar amounts of resources. In particular, we discovered a CNN configuration for classifying CIFAR10 dataset which has a 3% higher performance than the configuration founded by Hyperband, and is also 0.3% more accurate than the best-reported configuration of the Bayesian optimization approach. Additionally, we release a publicly available pool of historically well-performed configurations on several datasets for CNN and SVM to ease the adoption of Meta-Hyperband.