[AN #109]: Teaching neural nets to generalize the way humans would — LessWrong