Motorcyclists have a word for this phenomenon. It's called SMIDSY - "sorry mate, I didn't see you." Clear road, good vision, bike traveling in a straight line within the speed limit, car still somehow pulls out and hits the bike. Even worse, the times when the bike is stationary at a red light and the car just drives through it.

The real world is not filled with adversarial examples. Further, NN often does not think it's doing great when providing off the wall examples. Further this is a car, with GPS, where simply stopping is a completely viable option 99% of the time.

The truth is people really really stuck a driving, but it's easy enough that the vast majority of mistakes don't result in any problems.

> Humans know when they can't see properly

There's plenty of conditions that are more difficult to humans: cars that are painted with the same color of asphalt seem to disappear in highways. Also, fog.

Usually. Drunk and sleepy drivers blur this distinction, but you still make a good point.

Aren't general adversarial networks designed to solve this?

That doesn't sound as an insurmountable problem.

Humans don't see elephants suddenly because a prior is telling them that an elephant is improbable in that context.

Something like that in a computer system shouldn't be above the current capabilities.

These look like a digital Rorschach test. With the minor difference that a human "knows" the image is not really an object and is being asked to interpret in some way; where the machine has no "knowledge" nor "understanding" but has an imperative to match this input to something in its repertoire of knowledge.

As a human, if you asked me to match these synthetic images to a most-likely real-world match, then my responses will also look "confused."

Those aren't too bad (out of context of the Tesla article), I can see how the ai might misinterpret those images. The bagel one, for example, is pretty on point.

I guess the CNN was trained only with real images. They must add some [random noise], [crap], [glitch] and [collage] images to the training set. This may reduce the efficiency of the NN in the real world, but it will make more difficult to find funny examples of misclassification.

Our vision and persistent world model are good, but all humans are overconfident in their ability to drive in poor conditions (or even normal conditions, for many people.)

A computer isn't going to look down at a text message on their phone, fall asleep at the wheel, or get in a car while intoxicated. All which occur literally every day. People take driving for granted, because (at least in the US) we have to. Our sprawling suburbs couldn't function without it, so the barrier to entry is simply being a teenager.

Would I completely trust Tesla's auto-driving? No, not at all. But I at least trust them as much as I do an average driver, which is still just barely. The difference with Tesla's system is that one failure leads to improvement of the entire fleet, whereas one human failure maybe improves that one human driver. That means self-driving cars will continuously improve, while humans will just continue to be terrible drivers.

Robustness is a key property of the methods and models used in self driving cars.

For a human, the noise isn't bad recognition, it's corrupted sensor data via distraction. It's one of the most common causes of accidents. For the AI agent an elephant will never appear unless it's been trained to be recognized. More likely it's recognized as an obstacle, which is ok. But keep in mind that years of Google's autonomous vehicle success make this pretty improbable, certainly less than in the average driver.

Much of the fear here is irrational and natural when humans lose control.

a) well, no. we only have about 100 years dealing with these problems in context of driving cars and we're really bad at it; b) that seems like shifting goalposts, what we care about is whether driver's picture of the world is correct and actions are sensible with respect to that picture, it just so happens that driverless cars' picture is visual (or close to visual), but if you're judging them on entirety of input and side-effects it's only fair to judge our brains by the same standard.

Some adversarial methods work on black-box networks - has any work been done on applying those to human vision? We know there are a wide range of simple inputs that produce unexpected results (eg. optical illusions​).