(And it seems like further one could identify the feature in pixelspace by taking the gradient of v⋅∇vf(x,Θ+v) with respect to the pixels? Might be useful for interpretability? Not sure.)
(And it seems like further one could identify the feature in pixelspace by taking the gradient of v⋅∇vf(x,Θ+v) with respect to the pixels? Might be useful for interpretability? Not sure.)