Gameify object labelling to help train the FSD

phelo

Do Tesla have employees sifting through images and labeling objects to help train the FSD, if so why don't they gameify it and release it as an app. They could give points to top contributors to encourage participation that could be redeemed for charging or other such incentives. To ensure the accuracy of the labeling they could compare multiple user inputs etc.. I believe google does something similar to train AI using the dreaded captcha, are you a human test that I can never pass :)

M-A-B-MCMLXXX

“ Do Tesla have employees sifting through images and labeling objects to help train the FSD”

Yes. I don’t think it’s a wise idea to outsource that to the public. They would need to expend just as much effort to check the submissions.

phelo

Yes i agree outsourcing to the public would have some challenges but I do not think verifying of submissions would be one of them as they could just select the results that had the highest concurrence among users. Obviously I have no idea how the labeling system works so it might just be completely impractical. Although people already suspect that google uses the recaptcha as part of its training for waymo

https://www.ceros.com/originals/recaptcha-waymo-future-of-self-driving-cars/

M-A-B-MCMLXXX

Nefarious people could easily game (haha) that, though, and flood the system with a bunch of false tags. Say some Romanian hacker wrote code to submit millions of photos of stop signs to Tesla tagged as yield signs. No bueno. Tesla is the responsible and liable party.

stingray.don_98527447

Labeling is not the bottleneck. For Tesla, the bottleneck is the computing power of the supercomputer that conducts the machine learning. This is the reason that Tesla is building a new supercomputer called Dojo.

phelo

Hmm, I am not so sure that it is not a bottle neck i read that Tesla is going to double the amount of people they have doing annotations to 1,000, I believe dojo will speed up the process but they will still need humans to label single frames that dojo will then apply to complete videos instead of humans having to label every frame which will speed up the process but still more humans would mean faster training.



https://www.google.com/amp/s/syncedreview.com/2020/08/17/elon-musk-seeks-ai-talents-for-video-data-processing-supercomputer-dojo/amp/

https://www.tesla.com/careers/search/job/data-annotationspecialist-71323

phelo

> @M-A-B-MCMLXXX said:
> Nefarious people could easily game (haha) that, though, and flood the system with a bunch of false tags. Say some Romanian hacker wrote code to submit millions of photos of stop signs to Tesla tagged as yield signs. No bueno. Tesla is the responsible and liable party.

Yes thats prob a big risk alright, but if it was limited to tesla owners that might reduce that risk

stingray.don_98527447

> @phelo said:
> Hmm, I am not so sure that it is not a bottle neck i read that Tesla is going to double the amount of people they have doing annotations to 1,000, I believe dojo will speed up the process but they will still need humans to label single frames that dojo will then apply to complete videos instead of humans having to label every frame which will speed up the process but still more humans would mean faster training.
>
>
>
> https://www.google.com/amp/s/syncedreview.com/2020/08/17/elon-musk-seeks-ai-talents-for-video-data-processing-supercomputer-dojo/amp/
>
> https://www.tesla.com/careers/search/job/data-annotationspecialist-71323

Sure, but Tesla is able to hire the necessary personnel to perform this function and keep direct control. There is also a significant amount of training involved for the annotators. They aren't even looking to hire part time employees for this effort. I like the idea, I just don't think this is the direction that Tesla is going with this effort. Also, it is Dojo that is driving the need for more annotators not the other way around.