The paper demonstrates that in certain artificial scenarios neural nets can be rather easily tricked. This has in fact long been known. But intellectual opponents of deep learning (like Gary Marcus) are seizing on it to suggest that the problems can't be fixed. That's premature.https://twitter.com/GaryMarcus/status/1034075062674935808 …
Where on earth are you getting that current cars are safer than any human population? Intervention rates are not consistent with that, as far as I can tell.
-
-
Not sure why you think I said that. I said we should define "safe" for DL cars as "causing fewer deaths than human drivers". Obviously we're not ready to test that today. But country stats provide a benchmark. My guess is DL cars will only be accepted when safer than US drivers.
-
i think you said it because ... you said it above. “[deep learning] likely already better”
- 2 more replies
New conversation -
-
-
Deaths/year/100K cars (Wikipedia stats): Somalia: 4480 Brazil: 58 US: 13 Norway: 3 I'm guessing with no proof that DL is already better than Somalia. But DL cars will never be accepted at Brazil let alone Somalia levels. Maybe they will if they get between US and Norway levels.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Another way to frame this debate is to ask "Does driving a car in the full spectrum of conditions that human drivers face require AGI?" We now know that playing chess or go at human expert levels does NOT require AGI. For cars maybe we'll find out fairly soon.
-
Would you consider this question too? Is reasoning needed to predict pedestrian & driver behavior? Or pattern recognition is enough?
- 1 more reply
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.