Mobileye, a leading vendor of autonomous vehicle technology, is basing its safety case on an elementary statistical fallacy: multiplying together two probabilities as if they're independent when they're not.
Mobileye is planning to build two different self-driving stacks—one based entirely on cameras and the other based entirely on radar and lidar. Then after testing the two separately, they'll combine them into one system.
The theory is that if one system has a 1/10,000 chance of crashing in any given hour and the other system also has a 1/10,000 chance of crashing, a combined system has a 1 in 100 million (10,000 times 10,000) chance of crashing per hour.
Hence Mobileye argues that it will be able to prove its system has human-level safety (<1 crash per 100M miles) without actually doing 100 million miles of testing. The assumption of independence lets them test each system to a lower level of safety and then multiply.
As I wrote three years ago when they first started saying this, this is a ridiculous assumption! Among other problems, scenarios that confuse cameras are likely to also confuse lidar, and vice versa. arstechnica.com/cars/2018/05/i…
I bring this up because Mobileye CEO Amnon Shashua made this argument again during a CES presentation yesterday. I assume he knows it's a marketing gimmick rather than a serious statistical argument, but it's weird that he keeps saying it.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
