Practical Bayesian Optimization of Machine Learning Algorithms

TLDR

Machine learning algorithms require careful tuning of hyperparameters, yet this process is often a black art relying on expert intuition or brute‑force search, motivating the development of automatic optimization methods. The study investigates automatic tuning of learning algorithms using Bayesian optimization with a Gaussian process model of generalization performance. Bayesian optimization uses a Gaussian process posterior to efficiently select parameters, and the authors propose cost‑aware, parallelizable algorithms that adaptively choose experiments. The authors demonstrate that careful selection of the Gaussian process prior and inference can dramatically affect Bayesian optimization, and that their cost‑aware, parallel algorithms outperform prior methods, achieving or surpassing expert‑level tuning across several modern machine learning models.

Abstract

Machine learning algorithms frequently require careful tuning of model hyperparameters, regularization terms, and optimization parameters. Unfortunately, this tuning is often a black art that requires expert experience, unwritten rules of thumb, or sometimes brute-force search. Much more appealing is the idea of developing automatic approaches which can optimize the performance of a given learning algorithm to the task at hand. In this work, we consider the automatic tuning problem within the framework of Bayesian optimization, in which a learning algorithm's generalization performance is modeled as a sample from a Gaussian process (GP). The tractable posterior distribution induced by the GP leads to efficient use of the information gathered by previous experiments, enabling optimal choices about what parameters to try next. Here we show how the effects of the Gaussian process prior and the associated inference procedure can have a large impact on the success or failure of Bayesian optimization. We show that thoughtful choices can lead to results that exceed expert-level performance in tuning machine learning algorithms. We also describe new algorithms that take into account the variable cost (duration) of learning experiments and that can leverage the presence of multiple cores for parallel experimentation. We show that these proposed algorithms improve on previous automatic procedures and can reach or surpass human expert-level optimization on a diverse set of contemporary algorithms including latent Dirichlet allocation, structured SVMs and convolutional neural networks.