Gary Marcus

Driverless cars just got a whole lot harder. This technical paper by @amirrosenfeld raises some profound questions about the robustness of #DeepLearning as a perceptual mechanism. https://arxiv.org/abs/1808.03305 

I am a little torn on papers like this i.e. when they argue with their importance to medicine. Understanding error modes is absolutely crucial, but these adversarial examples are so contrived, they are like saying: "Ha, your lab test does not work if you mix blood with lemonade."

Not certain about implications for medicine; implications for driverless cars seem scary. What the results really show is that we don’t exactly understand how deep learning does perception in context. And driving is all about perception in context.

I think the core of the argument is "driverless cars will never see non-natural images, barring the actions of bad actors." In one sense the model is right. There are no real objects in those pictures. It is true humans can cope better, but not unexpected.