Recently, Bayesian optimization has been successfully applied for optimizing hyperparameters of deep neural networks, significantly outperforming the expert-set hyperparameter values. The methods approximate and minimize the validation error as a function of hyperparameter values through probabilistic models like Gaussian processes. However, probabilistic models that require a prior distribution of the errors may be not adequate for approximating very complex error functions of deep neural networks. In this work, we propose to employ radial basis function as the surrogate of the error functions for optimizing both continuous and integer hyperparameters. The proposed non-probabilistic algorithm, called Hyperparameter Optimization using RBF and DYCORS (HORD), searches the surrogate for the most promising hyperparameter values while providing a good balance between exploration and exploitation. Extensive evaluations demonstrate HORD significantly outperforms the well-established Bayesian optimization methods such as Spearmint and TPE, both in terms of finding a near optimal solution with fewer expensive function evaluations, and in terms of a final validation error. Further, HORD performs equally well in low- and high-dimensional hyperparameter spaces, and by avoiding expensive covariance computation can also scale to a high number of observations.
from cs.AI updates on arXiv.org http://ift.tt/2af7iKv
via IFTTT
No comments:
Post a Comment