-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Neural Networks
- Thread starter lunitiks
- Start date
Yeah, not denying that Tesla could be much more transparent about when the functions will be ready and the buyers risks.
Yeah this isn't true. You can argue about scale, but Waymo and others are also training "deep learning" neural nets.
Also part of Tesla’s Smart Summon code was recently posted here and it was full of good old programmed control algorithms.
I don’t think things mean what @CarlK thinks they mean in this context.
CarlK
Active Member
They might use some degree of machine learning but not full fledged deep machine learning like Tesla does. Again without huge amount of data deep learning will not give you the accurate enough data. Also when Waymo started it's program deep learning was not even ready at that time. Waymo might have taken a different approach otherwise.
Here is a recent video from Lex Fridman. The entire lesson is informative and worth watching. Skip to 41:50 for the particular subject we're talking here. Vision Sensors + Deep Learning is what Tesla is using. Lidar + Maps is what everyone else is using.
Last edited:
Looks like that’s code in the mobile app for controlling Summon. Not saying that the car side of Summon is some AI genius that shoots yellow laser beams out of an infinity stone on its forehead but still....
Possible, although I find some of the data there indicates algorithmic inputs from the car. The bigger point is — and this has been covered elsewhere before — the control algorithms inside a Tesla Autopilot definitely are still ”Software 1.0.” The parts that are taught and can learn are not blindly or even directly driving the car.
Everyone mixes and matches different ideas and technologies.
EVNow
Well-Known Member
Yes, during autonomy day Tesla said somethings are better done in hueristics than NN.
BTW, is there an article that talks about various FSD efforts by different companies and their approach etc ?
Not enough, so far one has to keep up with many sources (which I guess is the prudent thing to do anyway
). I would expect the likes of Ars Technica to offer a best crack at a balanced technical look though.It will be interesting to keep watching this space, but realistically it is all speculation for the time being. The bets are on the table so to speak (though new bets are still coming in too), but it will take time to determine a winner. Until then it is all guesswork, educated as it may or may not be.
In other words: People believe what they want to believe, until reality concurs or shows otherwise.
Last edited:
You are saying simultaneously that they are not using "full fledged" deep learning and then implying that they are using deep learning, but without enough data to make it accurate. Which is it, are they using deep learning but without enough data, or are they not using deep learning?
Deep learning is a specific set of machine learning techniques. It is independent of the amount of data you have. You can do non-deep learning with a vast data set, or deep learning with a small data set. It is basically incontrovertible that many many other companies are using deep learning for AVs, by the way, so please be careful what you claim.
CarlK
Active Member
Yes the link I had two post above yours. That's one from the probably most authoritative and objective person in this field. Most articles you see are written by people who do not have any understanding of the technology or from those having their own financial agenda. Just need to take a large chunk of salt when you read them.
Just noticed that link was voted down by two perennial Tesla haters on this forum whom I have already ignored. That gives you more reason than anything else to watch this very informative video.
EVNow
Well-Known Member
I'll watch that. I've been listening to Lex's AI podcast.
Its strange there is no one good article comparing and contrasting all the companies, their strategy and status. I've been searching for one.
diplomat33
Average guy who loves autonomous vehicles
Thanks for sharing that video from Lex Friedman. It was very informative. I really like his comparison between the vision+deep learning approach versus the LIDAR+maps approach. It really helps put what Tesla and other companies are doing in their proper perspective. I do think that Tesla will get to FSD with their approach, it just requires more time and effort with deep learning.
CarlK
Active Member
Part of the reason is people writing articles are mostly auto people but not tech people. All they could understand are superficial things like who are using Lidar and who are not or who had demos and who did not have. The comparison in Lex's video (@41:50) is the first and the only time I saw a clear comparison of those different approaches. I don't think any of those auto writers could even get a gasp of that.
People have been very confused of state of the art of the technology. You can't blame them because it's always been he say she say but we don't have a (logical) measuring stick. The situation starts to change with recent opinions from people like Lex Fridman, he did not want to offend anyone but obviously favors Vision+Deep Learning over Lidar+Map, and Anthony Levandowski who had been Google/Waymo's "Lidar Tsar" but now says Elon is right Lidar is a crutch. Deep learning does require a huge amount of data. That's the reason why Tesla wants to put hundreds of thousands cars on the road first and why others will never be able to implement it anytime soon.
The sad thing is mainstream media is never interested to dig deep and get stuff like those. The only hope is when the true self driving system comes out where everyone could see. Kind of like what electrical vehicles had been though from it will never work to everyone is jumping on the bandwagon after you showed them how to do. The optimistic in me thinks the wait is not going to be very long.
Last edited:
EVNow
Well-Known Member
Ars Technica is actually extremely biased against Tesla. Their writers / editors (like Jonathan) openly, brazenly support Tesla shorts and block anyone who might support Tesla. Can you imagine any "journalist" actually using a block list maintained by extremely biased shorts (who say Tesla is a total fraud) and blocking anyone and everyone who might refute their articles ? 90% of those who would have never interacted with them ?
Nafnlaus on Twitter
EVNow
Well-Known Member
What I'd like to hear is someone who is an industry expert (preferably an academic) who interviews the major players to figure out where they are. Lex is actually well situated to do that - but he didn't press Musk enough to figure out where they are. You need a more aggressive interviewer.
It is difficult to tell whether Musk is talking past, present or future. He just mixes all those. Yes, lots of data is good - but where is that being used today and how ? Shadow mode is great, but how is it really being used today ? i.e. Musk keeps talking about the potential of big data - but we don't know where they are today.
To me it looks like in City NOA, Tesla is still figuring out basic things - they don't need 400k cars to be sending the data for those basic things (grass vs road !).
Well, I said ”the likes of Ars Technica” as in the type of website that focuses on technical content. Many EV sites for example lack the technical expertise to really discuss the merits of autonomous systems, so the likes of Ars Technica are a better bet generally.
That said it is also true that many tech-heads are highly sceptical of Elon Musk’s autonomous driving proclamations. I can’t shake the feeling that there may be some merit to that scepticism unfortunately.
Similar threads
- Article
