Shortest Program Interpolation Learning

Abstract

We prove that the Minimum Program Length learning rule exhibits tempered overfitting. We obtain tempered agnostic finite sample learning guarantees and characterize the asymptotic behavior in the presence of random label noise.