I agree with much of what Gary Marcus says. Using trained deep models with out of sample data in the real world will cause problems. There is a good reason that

I agree with much of what Gary Marcus says. Using trained deep models with out of sample data in the real world will cause problems. There is a good reason that self driving car companies want millions of miles of training data: to reduce out of training sample incidents. For almost four years, my work has been almost 100% development of sometimes novel, sometimes mundane deep learning models so I am a fan but I still wait eagerly for breakthroughs in hybrid AI systems, new Bayesian techniques that can work with less data, handle counter factual conjectures, etc.