Yushi Nakaya

From internal representations to model improvement through prediction errors

It is suggested that linking a model's internal features to its errors and their effects on performance may help select training images that improve performance, thereby allowing the model's current state to guide which images are labeled next.