|
Abstract:
|
In this talk I will discuss distribution shift, both as an obstacle to be overcome to achieve generalization, and as a device for obtaining generalization guarantees. In the first part, I will discuss the problem of label shift, where the proportion among the labels can shift but the class conditional distributions do not change, including connections to some practical problems and some theoretical results. Then I will discuss a new work in which we deliberately perturb the distribution of training data in order to obtain a generalization guarantee.
|
