This was good. Thank you. A couple of things came to mind as an investor... From this keyhole look at the Tesla software trenches.
1) Labeling and what was shown is needed for sign and traffic light reading, but there needs to be a high data rate path that uses a neural network processing to filter using a physics engine. There needs to be a third domain in the picture that characterizes based on mass and motion independent of labels. Think Feynman and his talk with his dad. It is less about what it (the bird) is called and more about what it does.
2) When he talked about turn signals/blinkers, a light went on. From the Model X British comedy show frame rate (Benny Hill?)... They are using slow frame rates to attenuate the data stream and make it manageable. That will not work on blinkers. The data attentuation has to happen on motion... think like a dog, or maybe other animal. Some animals don't see non-moving objects. They get a high frame rate on motion/transition - everything else is invisible. Again calls for a separate fast physics section.
The AI should be doing the labeling based on a high data rate stream filtered by motion. Less about what it is and more about "did it move, or change state?"
Feynman's dad was underrated.
[Edit: Elon posted a link to this research on visual acuity in nature: https://www.cell.com/action/showImagesData?pii=S0169-5347(18)30052-1
The graph sums it up.
There are huge differences in how animals see the world -- we're among the crisp-eyed
Tesla is going in the right direction. Maybe they will filter on pixels rather than frame rate for the physics domain...?]
Last edited by a moderator:
