There's another thing he doesn't seem to mention: the problem in all learning systems of converging on and then getting stuck in local maxima.
Perhaps when things get stuck it is in fact rational to be irrational-- to strategically adopt odd heterodox modes of thought or even employ a bit of randomness (Monte Carlo methods) to shake yourself free of a local maximum.
Perhaps when things get stuck it is in fact rational to be irrational-- to strategically adopt odd heterodox modes of thought or even employ a bit of randomness (Monte Carlo methods) to shake yourself free of a local maximum.