Dataset2Vec: learning dataset meta-features

Abstract Meta-learning, or learning to learn, is a machine approach that utilizes prior experiences expedite the process on unseen tasks. As data-driven approach, meta-learning requires meta-features represent primary tasks datasets, and are estimated traditonally as engineered dataset statistics require expert domain knowledge tailored for every meta-task. In this paper, first, we propose meta-feature extractor called Dataset2Vec combines versatility of with expressivity learned by deep neural networks. Primary datasets represented hierarchical sets, i.e., set esp. predictor/target pairs, then DeepSet architecture employed regress them. Second, novel auxiliary task abundant data similarity aims predict if two batches stem from same different ones. an experiment large-scale hyperparameter optimization 120 UCI varying schemas task, show outperform thus demonstrate usefulness first time.