I realized one cool thing that I didn’t consider at AI day.
When Tesla are using auto-labelling, they use multiple drives at the same area to label lane markings etc to generate the labels, ie the outputs. Then they can train the neural network using raw video from snowy drives where the lane markings are very diffucult to see. This is much better than using human labeled images where the human labeler will struggle to see the lane lines. The neural network will become superhuman at figuring out where the lane lines should be.
When Tesla are using auto-labelling, they use multiple drives at the same area to label lane markings etc to generate the labels, ie the outputs. Then they can train the neural network using raw video from snowy drives where the lane markings are very diffucult to see. This is much better than using human labeled images where the human labeler will struggle to see the lane lines. The neural network will become superhuman at figuring out where the lane lines should be.