6 ms·
Predictable outcome of a sensing system fully based on deep learning. Rare unusual situations don't have enough training data and lead to unpredictable output.
by crubier 5y ago
Predictable outcome of a sensing system fully based on deep learning. Rare unusual situations don't have enough training data and lead to unpredictable output.
I still think that Tesla's approach is the right one, I just think they need to gather more data before letting this product be used in the wild unsupervised.
- judge2020 5y agoCurrent TACC/auto-steer doesn’t use deep learning except on the newest Model 3/Y vehicles with “TeslaVision”. All cars with radar use the radar and radar only to determine if they should stop for the following car.
- dawnerd 5y agoDefinitely use cameras as well to determine stopping. Otherwise there wouldn’t have been the issue with bridges or shadows causing phantom braking.
- laichzeit0 5y agoHow does TeslaVision work with stationary objects at night? Like say a big ass truck with its lights off? Do you just pray the vision system recognizes “something” is there? I know they want to pursue a pure-vision system with no radar input, but it seems like there will be some crazy low light / low visibility edge cases you’d have to deal with.
- marvin 5y agoHow does a human detect a big ass truck with its lights off at night? This is solvable with computer vision. Tesla's dataset is almost nothing but edge cases, and they keep adding more all the time. My money says they'll get there.
- laichzeit0 5y agoThe thing is we don’t. Many people die from rear ending broke down trucks. I’m fine with that, I’m not so sure if regulators will be fine with a TSLA killing someone and going “welp, a human wouldn’t have seen it either, let’s just add this to our edge case dataset”.
- 360walk 5y agoI think it is necessary for the crashes to occur, to gather the data required to re-train the auto-pilot. We as a society need to decide whether we want to pay this cost of technological advancement.
- crubier 5y agoNo. Gathering data of human drivers braking in those circumstances would result in a perfectly fine dataset. This idea of needing human sacrifice is bonkers.