How can the author draw conclusions without citing any code or experiments. For example, if we try to use the features DeepDream leaned we can see that our imag

How can the author draw conclusions without citing any code or experiments. For example, if we try to use the features DeepDream leaned we can see that our image will have terrible compression artifacts (faces of cats). DNNs do dimensional reduction, it is not clear if this correctly preserves image features.