Gary Marcus

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 would hope any serious self driving effort does not do single frame detection but rather integrates information across time (and uses Lidar). Also agree somewhat this is contrived, like inputting giberrish into Google translate - not too surprising you get weird outputs.

Of course the single largest vendor of “autopilot” systems doesn’t use Lidar, and no present regulation requires them to do so.

Can you explain to me what is bad about fitting curves (or complex statistical models, as I would call them)? Are you making a technical argument here?

I suspect you could create a traffic jam with a couple of cardboard celebrity cutouts standing in the road. Will take a little while for self-driving cars to recognise that they’re a fake obstacle.