Neural net / Fleet-learning

[teethdood]

Yesterday evening I was using AP with my family on a highway I used daily to go to work. Maps navigation was also on. AP had always worked fine on this 2-lane HW. We were on a left lane as we approached an exit. The HW bends to the left, the exit bends to the right. All of a sudden the car jerked to the right into the right lane as if it was trying to exit fast. My hands were off the steering wheel and I did not engage auto lane change.

My question is: if Tesla is building a neural network, shouldn't a frequently traveled road get better and better for all Teslas? Like based on GPS knowledge or some such, especially with navigation engaged, the car can download expected safe steering behavior from past cars, and not just rely solely on lane markers. (Sun was low, casting a reddish hue, which made it harder to detect lane markers I guess)

 

[182RG]

My question is: if Tesla is building a neural network, shouldn't a frequently traveled road get better and better for all Teslas? Like based on GPS knowledge or some such, especially with navigation engaged, the car can download expected safe steering behavior from past cars, and not just rely solely on lane markers.

There is strong speculation that the whole "fleet learning" concept is heavy on marketing and Twitter hype, and far less on technology. In fact, it may be non-existent. It's highly unlikely that your car is benefiting from "past cars" in a granular, specific way.

 

[larmor]

There is strong speculation that the whole "fleet learning" concept is heavy on marketing and Twitter hype, and far less on technology. In fact, it may be non-existent. It's highly unlikely that your car is benefiting from "past cars" in a granular, specific way.

At least that is what Hal Series 8000 says, and hence was soon replaced with Series 9000, with stricter adherence to the mission...

 

[PedanticOne]

From my understanding, the net is not training right in your car, but instead data is sent back to the mothership and then combined with the data from all other cars and the net is trained. Then the newly trained net is given to all cars in an update. That is what is meant by fleet learning. It's aggregate.

 

[Yinn]

From my understanding, the net is not training right in your car, but instead data is sent back to the mothership and then combined with the data from all other cars and the net is trained. Then the newly trained net is given to all cars in an update. That is what is meant by fleet learning. It's aggregate.

That's correct. The data is collected from the entire fleet, and yes other cars will result in faster learning as a result of said data collection.

With that said, I haven't seen any info that the learning is dynamic; with correlation occurring on its own. The data may very well still have to be analyzed and programmed in.

I.e. 20 people jerked the wheel or aborted at the same spot, that might trigger a code review. Vs immediately recognizing that spot is a problem and auto updating the behavior.

 

[teethdood]

I don't presume to know how Tesla does its programming, but I think it should be more "google maps"-like rather than static code. If it's static code, it has to account for practically infinite scenarios involved with driving, all hard-coded in. Whereas, it's google-maps-like where it downloads expected steering/speed behavior for a stretch of road that say 50 teslas had passed by safely before, it would make more sense.

So far, I don't think it's doing the google maps behavior, so no active neural net/fleet learning.