These weakly abstract models can generalize, but only locally so. They still break down upon simple problem modifications or when encountering novel problems. This has been true of all deep learning generated abstractions, across every domain and across all architectures.

→ View original post on X — @fchollet,