AutoPilot "Learning"

[lolachampcar]

I was very surprised to hear all the talk of Tesla's AP learning as it went along. Updating weekly. Neural network learning. And all the rest of the speculation that followed some of Elon's comments.

My AP did not change appreciably from the original release. Where others saw learning and changed behavior, I saw small changes in ambient conditions that could easily explain the differing performance.

Fast forward to 7.1 and we have some real meaningful changes in AP functionality. This dove tails nicely with how these types of developments occur. There is no secret sauce special fancy learning going on. Its just a bunch of engineers and programmers hammering on fixing the thing when they confirm it is doing something stupid.

It is really hard not to follow the bread crumbs Elon leaves but I'm getting better at separating marketing BS from reality. Like horsepower, Tesla's AP did not need the damage that comes from BS hype. It is more than good enough to stand on its own merits.

 

[PaulR]

Lolachampcar,

Well stated.... I agree. About a month ago in the midst of all the speculation about the software and hardware "learning" I found the following article and speech (referred to by someone on this forum - thanks to that person) given by Usman Pirzada who gives meaning to the whole process and explains what Tesla's hardware and software REALLY does. Take a few minutes and you will appreciate even more how forward thinking Mr. Musk and those who developed the hardware and software which supports his thinking really are:

Want to know what Deep Neural Networks are and do? What will the future probably bring Tesla and us?

Take a few minutes and read this article -

Exclusive: The Tesla AutoPilot - An In-Depth Look At The Technology Behind the Engineering Marvel

Enjoy!

 

[AWDtsla]

Lolachampcar,

Well stated.... I agree. About a month ago in the midst of all the speculation about the software and hardware "learning" I found the following article and speech (referred to by someone on this forum - thanks to that person) given by Usman Pirzada who gives meaning to the whole process and explains what Tesla's hardware and software REALLY does. Take a few minutes and you will appreciate even more how forward thinking Mr. Musk and those who developed the hardware and software which supports his thinking really are:

Want to know what Deep Neural Networks are and do? What will the future probably bring Tesla and us?

Take a few minutes and read this article -

Exclusive: The Tesla AutoPilot - An In-Depth Look At The Technology Behind the Engineering Marvel

Enjoy!

That's not Tesla's hardware or software. Tesla's value add is supposed to be fleet learning.

 

[lolachampcar]

My rub is mentioning something like real time learning then leaving the world to speculate and fill in the gaps with imagined intelligence. It looks like true product migration will occur on a firmware update schedule and not weekly (weakly and not weekly ).

 

[ohmman]

The thing that bothered me the most about the fleet learning commentary is that Elon himself retweeted someone saying they'd noticed improvements. This was, if I recall correctly, days after the release. If indeed the new models are bundled with the firmware releases (as I have publicly assumed would be the case), then that retweet was a questionable move. At least, that's how I see it.

 

[dirkhh]

The thing that bothered me the most about the fleet learning commentary is that Elon himself retweeted someone saying they'd noticed improvements. This was, if I recall correctly, days after the release. If indeed the new models are bundled with the firmware releases (as I have publicly assumed would be the case), then that retweet was a questionable move. At least, that's how I see it.

It may not be very ethical (assuming you are correct), but it's smart. No one's going to sue him over it, yet it supports a perception that is very positive for Tesla...

 

[AWDtsla]

The thing that bothered me the most about the fleet learning commentary is that Elon himself retweeted someone saying they'd noticed improvements. This was, if I recall correctly, days after the release. If indeed the new models are bundled with the firmware releases (as I have publicly assumed would be the case), then that retweet was a questionable move. At least, that's how I see it.

The real answer is we know nothing about how the system works. Maybe it needs 10 cars all doing the same correction in the same place. Maybe 100. Maybe it doesn't learn anything without user intervention. Maybe you can teach it by pushing it slightly. Maybe it learns nothing at all. There's a lot of engineers and other technically oriented people that could be doing a better job being expert trainers if they would clue us in, but it doesn't look like they have any intention of doing so. Apple-esque secrecy without Apple-esque quality.

 

[techmaven]

It is certainly possible for there to be limited amounts of learning that is built into the deployed fleet that might be subtle. If there are more glaring errors with maps on your route, maybe you'd see more of a correction.

Then on significant software releases, a lot more functionality is released and you see a bigger difference.

The fleet wide learning is probably about high resolution mapping and driver corrections. The high resolution mapping part is likely the candidate for essentially crowd sourcing and re-release. If there are few drivers in your area and/or the maps are pretty good already and/or the road markings are pretty good, then you might not see any improvement from just that aspect.

 

[lolachampcar]

My goal is a better and more accurate understanding of exactly how much I can trust what Elon (as he really is the voice for Tesla) says. I'm quickly coming to the conclusion that he is very bright and says just enough to wind people up in the direction he wants them to go even if there little to no truth in that direction. I'm afraid I'll have to disagree with you dirkhh. I do not consider that type of communication to be smart as it erodes credibility.

 

[AWDtsla]

My goal is a better and more accurate understanding of exactly how much I can trust what Elon (as he really is the voice for Tesla) says. I'm quickly coming to the conclusion that he is very bright and says just enough to wind people up in the direction he wants them to go even if there little to no truth in that direction. I'm afraid I'll have to disagree with you dirkhh. I do not consider that type of communication to be smart as it erodes credibility.

He clearly is forming a reality distortion field. And maybe they have some software in the works that does exactly what he says, but maybe it will be a year before our cars are actually running it, maybe it will only work on future cars.

 

[RubberToe]

It seems perfectly reasonable to me that Tesla is not telling anyone how the AP functionality works. The amount of $$$ at stake are mind boggling for whomever figures out fully autonomous driving. You really wouldn't even want to give your competitors a hint at how you are going about it. Especially since Google is on record as believing that AP is NOT the way to go, and Tesla is obviously going that way. One or the other is missing something.

My $0.02

RT

 

[Lesifass]

It seems perfectly reasonable to me that Tesla is not telling anyone how the AP functionality works.

There's a difference between "not telling anyone anything" and "telling everyone about features that apparently don't exist / don't work as advertised".

 

[ohmman]

It seems perfectly reasonable to me that Tesla is not telling anyone how the AP functionality works. The amount of $$$ at stake are mind boggling for whomever figures out fully autonomous driving. You really wouldn't even want to give your competitors a hint at how you are going about it. Especially since Google is on record as believing that AP is NOT the way to go, and Tesla is obviously going that way. One or the other is missing something.

Not sure I agree with this fully (most deep learning tools are open these days - see TensorFlow for instance), but your last sentence is interesting. To add to it, Baidu Chief Scientist Andrew Ng says that they believe neither Google nor Tesla have the right idea. Baidu is working on autonomous buses which will drive the same route over and over, and then slowly expand the route over time as confidence goes up. I find that squarely in the middle of both.

My belief is that none of them are wrong; they all have their own applications. And eventually, they're likely to converge.

 

[AWDtsla]

It seems perfectly reasonable to me that Tesla is not telling anyone how the AP functionality works. The amount of $$$ at stake are mind boggling for whomever figures out fully autonomous driving. You really wouldn't even want to give your competitors a hint at how you are going about it. Especially since Google is on record as believing that AP is NOT the way to go, and Tesla is obviously going that way. One or the other is missing something.

My $0.02

RT

They could reveal a plethora of information that would put owners in a better place without revealing anything secret or anything that would help someone else develop their own system.

For example:
"Am I teaching the fleet if I keep my hands on the wheel?"

 

[LetsGoFast]

The thing that bothered me the most about the fleet learning commentary is that Elon himself retweeted someone saying they'd noticed improvements. This was, if I recall correctly, days after the release. If indeed the new models are bundled with the firmware releases (as I have publicly assumed would be the case), then that retweet was a questionable move. At least, that's how I see it.

There's a difference between "not telling anyone anything" and "telling everyone about features that apparently don't exist / don't work as advertised".

I would agree. If in fact there no incremental changes, it would border on a basis for security fraud class action. If it was a knowingly false statement, it would be a serious violation. I'm unwilling to conclude that Elon would ever completely lie like that. Exaggerations or optimistic dates are a completely different thing than claiming a capability that simply doesn't exist.