News for hackers. Stuff that matters.
Re:Tesla Live Stream – Autonomy Day [video] by gamblor956 on Monday April 22, 2019 @05:43PM (#19723289)
This is definitely not true. The engineer explicitly stated that Tesla only gets data when trigger conditions are met (and that this was in fact always the case), so most of the miles that Tesla drivers have been driving was actually never even sent back to the mothership for use in ML training.

IOW, despite Elon claiming for years that they're ahead because of the "millions of miles" of free data collection by Tesla drivers, it turns out once the engineers on the ground actually speak...they're actually pretty far behind on the data front, and most of the data they are getting is the same safe routes over and over again since they're not trying to put Autopilot into novel situations to acquire new data or test the ML algorithm's capabilities to handle unique/novel situations (like trucks turning in the highway, or freeway dividers).

Re:Tesla Live Stream – Autonomy Day [video] by dreaminvm on Monday April 22, 2019 @05:57PM (#19723421)
They cherry-pick rare cases and use their fleet to get more examples of these situations. This seems like the right approach given more miles following the same car in a straight line is pretty useless.

My takeaway from the presentation is that Tesla will perform better than other companies in this space (although I don't know enough about Waymo to comment) due to the following:

-You want a large dataset (Tesla and many companies have this and can simulate) -You want a varied/diverse dataset (Tesla and many companies have this and can simulate)--the point here is simulations for simple cases work (you can only simulate when you know), but for complex ones are close to the difficulty of actual FSD -You want a real dataset (Tesla is the only company who can say this and can say they have data on how X00Ks of drivers will handle these situations)

Re:Tesla Live Stream – Autonomy Day [video] by gamblor956 on Monday April 22, 2019 @07:37PM (#19724180)
My point was that it appears that Tesla doesn't have a large and varied dataset. It has a small and pre-selected data set, since the cars only transmit data when pre-determined triggers are fired. Thus, it doesn't matter how many 1000s of drivers Tesla "has" or how many "situations" they're in, since it's not actually collecting data from most of these situations.

And Autpilot's performance (including its numerous regressions) suggests very strongly either that it doesn't have a very large data set, or else that it has a large data set of everyone doing roughly the same thing almost all of the time. These are the two most logical explanations for Autopilot's tendency to veer toward freeway dividers even (especially?) after updates.

Re:Tesla Live Stream – Autonomy Day [video] by Joky on Monday April 22, 2019 @07:49PM (#19724259)
> My point was that it appears that Tesla doesn't have a large and varied dataset. It has a small and pre-selected data set,

I don't thinks it is a fair characterization: first the notion of "large" is fairly subjective. But more importantly the fact that they collect data on pre-determined triggers is just a guarantee that the dataset is not over-fitted (let say to the 280 and 101 in the Bay Area and to Elon's commute in LA) and instead has good coverage of the world.

Their capabilities of triggering on situation allows them to grow the dataset quickly in a supervised way. It is all about the granularity of these triggers, imagine that you can express "collect situation in tunnels with jerk higher than X m/s3" or "collect all lane change abort in snow condition", ... In the next 24h you get data from all over the world and this data is automatically tagged and classified by the neural net.

Re:Tesla Live Stream – Autonomy Day [video] by millettjon on Tuesday April 23, 2019 @08:58PM (#19734601)
Shadow mode would also transmit data no?
Re:Tesla Live Stream – Autonomy Day [video] by identity_zero on Monday April 22, 2019 @05:59PM (#19723436)
Isn't the "millions of miles of free data" the edge cases? Yes, they're not literally streaming back everything but they have millions of miles with people driving auto-pilot and sending back a significant amount of edge case scenarios, which is the truly valuable information. Sending back data of "successful auto-pilot" can be sending back 10 hours of driving on a straight two-lane road; which seems quite useless.

They also have cars all over the planet while Waymo seems to only be tested in a few cities. I can't find the podcast, but Elon mentioned how one of the main difficulties is the varying types of intersections. There are only so many intersections in Mountain View. But Tesla is getting data for all kinds of intersections.

Re:Tesla Live Stream – Autonomy Day [video] by andrewtbham on Monday April 22, 2019 @05:51PM (#19723350)
I have wondered about this myself... and I am working on an CV ML project myself... but data that is correctly predicted is not not as useful.. what you want are edge cases, images with low confidence, and errors. So when users take over and disengage, that is useful data. when the system beeps and forces a disengagement.. this is the data you need. That has been my experience and I'm glad to hear them validate that. Your training data needs diversity.