I agree that the _goal_ should be 10x or even 100x better than the best human drivers, but better than average would result in net good and it's hard for me to see an argument in which that's not true because it's less net suffering/death.
>If you want self-driving cars to be successful then you should be demanding from the carmakers to make these systems flawless.
We're still talking about tons of metal moving at very high speeds around other vehicles. That's just not possible. Should my self driving car handle a catastrophic tire failure at 90mph better than me? Absolutely. Is that still a situation that's likely to result in a crash regardless of the best inputs on a compromised system? Yes. There will _always_ be situations in which fatalities can occur with cars or any other kind of vehicle no matter how well they're engineered.
Having seat belts is obviously statistically safer, but when they were first introduced, many people loudly protested that they'd rather be thrown clear in an accident.
No, that’s not the definition of mean, that’s (something like) the definition of median.
In a skewed distribution (which driving skill may very well be), the mean can be very far from the middle. If a relatively small number of people are extraordinarily bad at driving, most people are above average. If a relatively small number of people are extraordinarily good at driving, most people are below average.
Why this nitpick matters: if the leading point of your thesis depends on is quantifying how many people are above average, you should really know if the average is at the 10th, 50th or 90th percentiles (or somewhere else entirely).
Average means something else, technically.
And I strongly doubt "average [X]er" would be interpreted as a median by most people. (Except when median and mean overlap)
The quote is "by definition like 49% of the drivers are technically better than the "average driver"." It's obvious to anyone who's not pedantic that they're referring to the median.
As for an example of a mode average: "The average participant's favorite polygon was a triangle."
That's not the original use. That's a response to the (unattributed) idea that self-driving cars would need to be "better than the average driver". In the original context of people giving thresholds for self-driving cars, I disagree that median is what is meant.
> As for an example of a mode average: "The average participant's favorite polygon was a triangle."
But that mode might be 5 out of 40, and I'd just call that a lie. In a lot of distributions the mode just gives you the biggest or smallest number. And it's affected a lot by bucket size; if I measure "how long did you sleep last night" with a resolution of seconds, the mode is probably going to be 0, which is hilariously misleading. Mode, much like a broken clock, might pick a suitable number sometimes, but all the factors you need to check to see if it's suitable basically render the idea of 'mode' redundant. Just use those factors to pick your number. Mean and median are pretty reliable in being useful. Mode isn't.
Of course, to impute a categorical variable in a data set with empty values, for example.
I find it crazy that so many people think it is. First off, by definition like 49% of the drivers are technically better than the "average driver".
Second, just like with "human-like translation" from machines, errors made by machines tend to be very different than errors made by humans. Perhaps self-driving cars can never cause a "drunken accident", but they could cause accidents that would almost never happen with most drivers.
Third, and perhaps the most important to hear by fans of this "better than average driver" idea, is that self-driving cars are not going to take off if they "only" kill almost as many people as humans do. If you want self-driving cars to be successful then you should be demanding from the carmakers to make these systems flawless.