[hatmatrix]

Yes - but how do you know when your test data and training data are different in the way that the model cares? With conventional methods, you can assess similarity by distances in the feature space of the model, or you have a physical understanding of why the model works than you have a better intuition about the differences in data that are important.