Meta-Learning Assistants Using a Novel Characterization of Data Landscapes

摘要

Our project focused on the mathematical foundations needed to build meta-learning assistants. The overall goal is to know how we can acquire and exploit knowledge about learning (i.e., meta-knowledge) to understand and improve the performance of learning algorithms. To that end, our work focused on the following research problem: how can we decide if one single complex model, or rather a combination of simple models, is the best strategy to use when we face a supervised learning task. Our results show that a combination of simple models is often the best choice, as a minimum increase in model complexity is equivalent to tenths of simple models.