Interpolation Learning With Minimum Description Length

Understand

We prove that the Minimum Description 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.