Fast Bayesian hyperparameter optimization on large datasets (Q1688974)

A Bayesian optimization procedure is proposed in this paper. It treats the size of a randomly subsampled dataset as an additional input to the loss function. At each evaluation, the optimization procedure allows the optimizer to actively choose such to yield most information about the loss-minimizing configuration on the full dataset per unit time spent. An automatically trade off high information gain about the global optimum against computational cost is established. The results of experiments with support vector machines and neural networks to evaluate the speed of the procedure in identifying good hyperparameter settings are presented.