What does it mean when autopilot is learning??

[cityberty]

With v9 I commonly see Tesla Tubers and some people here on the forum say “autopilot is ok but its still learning”

Does this mean it is learning locally with your car or across the fleet?

If it is across the fleet then I have 2 questions.

- Has anyone ever tested this? maybe drove a hard intersection with one tesla on AP that was previously failing and then correct it tested it again and then test it in a seperate Tesla on that same spot?

- Does it learn and update constantly(connected to wifi with downloads some people see overnight that brings down data) or only with new releases.

Or is it just a myth like when everyone claimed that cameras read speedlimit signs(which was pretty much disproved) and AP is just really updated through hardware with AP 2.0 and AP 2.5

Thanks

 

[trace]

Teaching neural nets requires a lot of computing, so all data (cameras, weather, speed, etc) is offloaded to Teslas servers. Basically every software update gets a neural net update.

You going through an intersection correctly does not directly teach the software to do it the exact same way because all conditions are different (could be raining next time AP goes through there). Ideally it will match your successful intersections interaction to hundreds of other successful intersection interactions and go through it the most successful way (eg: slowing down before the intersection like you do, but speed up after the intersection like user 81 does).

Lots of very smart Youtubers (smarter than me) and other journalists talk about this process more in depth.

 

[drklain]

If you read some of the posts on neural nets by some real experts on this forum (as opposed to the very opinionated, self-professed experts who really don't know what they are talking about), you'll get a feel for how machine learning works and specifically how Tesla Autopilot is doing it. @jimmy_d is one of those experts and recently posted an excellent post on what his analysis of the Version 9 AP neural net activity points to. You can read the post here. Bottom line, don't think of this machine learning like writing a computer program where you have a bunch of "if/then" statements for every possible situation. While not exactly the same, it is much closer to how a human or animal learns...and requires a TON of data. Think about a human learning to walk (or drive a car) and the literally thousands of factors constantly being evaluated and compared to past experience and knowledge before decision/actions are taken. Note - this analogy is not perfect and not technically correct/accurate, but it is the best one I can think of without making this post really long and complicated!

All of the data our cars collect with autopilot on and with autopilot off (so called "ghost mode") provides data points into the neural net.... Really fascinating stuff and completely different than traditional computer programming...