Supplementary Materials for Niseko: a Large-Scale Meta-Learning Dataset

There have been several studies for an overview of meta-learning techniques [28, 30]. With the prior tasks and the evaluations (e.g., accuracy or time) of some learning algorithms, there have been several studies aiming to find some promising configurations of learning algorithms given a new task. Through surrogate models which are built from previous tasks, we can measure task similarity and thus apply Bayesian Optimization [4] to find the next promising model for the new dataset. Wistuba et al. [32] train Gaussian Processes as surrogate models for prior tasks and the new task, and measure their similarity based on the means. Feurer et al. [11] combine the predictive distributions of Gaussian processes into a Gaussian process as the surrogate model.