I've seen a paper that argued that, but it was utterly idiotic since it used "L2 interpolation in encoding space" as its definition for "interpolation". Effectively, it was saying, "models that aren't linear aren't linear".

→ View original post on X — @fchollet,