The problem is that a lot of AI folks bury their heads under the sand in the name of “PAC learning” and claim that you’re unfair in asking for out-of-sample gen

The problem is that a lot of AI folks bury their heads under the sand in the name of “PAC learning” and claim that you’re unfair in asking for out-of-sample generalization. They are loathe to accept that the popular algorithms are overfitting. I’ve had this debate with numerous NN/RL researchers, and somehow it always comes to this impasse.