Open Problem: Monotonicity of Learning

[edit]

Tom Viering, Alexander Mey, Marco Loog ;
Proceedings of the Thirty-Second Conference on Learning Theory, PMLR 99:3198-3201, 2019.

Abstract

We pose the question to what extent a learning algorithm behaves monotonically in the following sense: does it perform better, in expectation, when adding one instance to the training set? We focus on empirical risk minimization and illustrate this property with several examples, two where it does hold and two where it does not. We also relate it to the notion of PAC-learnability.

Related Material