What is the current state of the art for finding similar images? Is it auto-encoders nn's, convnets and then just use some sort of distance measure?