And another recent one. That one shows that too much duplicated data (here, ~ training for many epochs) leads to overfitting:

→ View original post on X — @rasbt