A small randomly connected recurrent net tends to bubble around chaotically and often contain the function you are looking for, and you can find it by training only an output layer. That's the basic idea
Small random recurrent net contains function, train only output layer
By
–