There are essentially two main options to remedy this: 1. Find ways to perform active inference, so that the model adapts its learned program in contact with a new data distribution at test time. Would likely lead to some meaningful progress, but it isn't the ultimate solution,

→ View original post on X — @fchollet,