... the complexities of algorithms trained on big data mean they're very difficult to "patch."
-
-
-
The behavior that led to that accident is, in essence, part of all the other decision making the vehicle does.
-
It may be possible to train out that behavior, but it's hard to ensure those changes are truly localized to that operating condition.
-
Self-driving car algorithms are nothing that aerospace didn't already study 30 years ago. They're more advanced, more compute available...
-
... but the endemic problem of non-verifiability has remained unsolved.
-
My former company flew a reconfigurable neural net flight controller on the VISTA F-16 in the 90s/00s. That tech still isn't fielded.
-
It worked fantastically! But V&V requirements were too huge. And flight is a simpler case than driving.
-
Convolutional NN stuff is great for object detection and classification, but ultimately it's a controls problem to solve. And that's hard.
- 12 more replies
New conversation -
-
-
reminds me of the tank on a sunny/cloudy day story: https://www.jefftk.com/p/detecting-tanks …
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
looking for an article about camera-self-driving by dude called Xi I think, to ask what you thought, can't find anywhere now
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
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.