Top eigenvector of AGOP of two separate models, MLPs and Laplace kernel machines, captured similar features (cosine similarity greater than .99) when trained on the same data from CelebA across ...