Listening to the interview again, I’m not so sure what Elon was saying. It’s ambiguous. It’s hard to tell when he’s talking about perception, and when he’s talking about action (i.e. path planning and driving policy) — if at all.
This quote (14:25) sure sounds like it could refer to imitation learning:
And we’re really starting to get quite good at not even requiring human labelling. Basically the person, say, drives the intersection and is thereby training Autopilot what to do.
But I wonder if it could just mean, for example, that if Tesla drivers stop for a traffic light, it is automatically labelled as red. The cars upload a picture of the traffic light, it is automatically labelled as red since the cars were stopped, and those labelled images are used to train the perception neural network. This would be a form of weakly supervised learning applied to computer vision (e.g. object recognition).
An example of weakly supervised learning for image recognition is Facebook training a neural network based on images from Instagram, weakly labelled with hashtags. Instagram hashtags only loosely correspond to what’s in the image. That’s why this is weak labelling.
So, I’m not sure if Elon was talking about imitation learning for path planning and driving policy, or if he was talking about weakly supervised learning for computer vision.