3 ms·
If they're not training autonomous systems for these extremely simple use cases, I'd imagine there wouldn't be much confidence in what else will be missed by th
by Y7ZCQtNo39 9y ago
If they're not training autonomous systems for these extremely simple use cases, I'd imagine there wouldn't be much confidence in what else will be missed by the "autonomous" system.
Out of the infinite number of possibilities one can encounter while driving, can the systems be trained to apply common sense? You can't program an infinite number of if-then conditionals for every possibility. The system has to be able to "think" and apply reasonable interpretation, like any human would when taking in any visual or audible input and converting that into a vehicle maneuver.
The data today suggests that autonomous vehicles get into less accidents per vehicle mile than human drivers, which is promising. But who knows if autonomous vehicles are being used in complex situations representative of human driving conditions: poor weather, non-highway driving, etc. If we're putting them on the 101 in sunny weather, then yes, I'd expect them to perform reasonably well and definitely better than the average human.
- xapata 9y ago> extremely simple use cases When's the last time you saw someone put black tape over a stop sign? These adversarial examples are not simple use cases. They're situations that humans rarely if ever encounter. > apply common sense ... "think" Ever seen the "invisible rope prank"? Humans are easily fooled as well. > if-then conditionals That's not the way statistical machine learning works. It's layers upon layers of probabilistic functions. > who knows if ... representative of human driving conditions I've read reports that a few of the autonomous vehicle companies are indeed testing/training on regular roads, which would allow a representative sample of human driving conditions. Frankly, that's just basic science -- don't experiment with a biased sample.