Reviews: Learning To Learn Around A Common Mean

This paper studies how to define an algorithm that, given an increasing number of tasks sampled from a defined environment, will train on them and learn a model that will be well suited for any new task sampled from the same environment. The scenario just described corresponds to the'learning to lean' problem where a learning agent improves its learning performance with the number of tasks. Specifically in this work the focus is on the'ridge regression' family of algorithms and the environment consists in tasks that can be solved by ridge regression with models around a common mean. In other words, we need a learning algorithm that besides solving regression problems, progressively learns how to approximate the environment model mean. The transfer risk is a measure of how much the knowledge acquired over certain available tasks allow to improve future learning.