Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Tesla Live Stream – Autonomy Day [video]

livestream.tesla.com

Post reply on HN
Print view View original

301–310 of 550 posts

Re: Tesla Live Stream – Autonomy Day [video]

#301

Post by RivieraKid »

trixie_ wrote: post #290
RivieraKid wrote: post #253

Earlier quoted context omitted.

If this is true, why do Waymo (or any of the top 10 self-driving companies besides Tesla) use Lidar? Two facts: 1. Tesla can't use Lidar because of cost and aesthetics. 2. Elon has a strong incentive to oversell their approach to self-driving.

From the presentation they mention that Lidar is good at developing tech demos quickly, but they've found it is insufficient in reality. Which of those 10 companies you reference have moved far beyond tech demos?

Waymo operates the first autonomous taxi service in Phoenix with plans to expand geographically, no info about the others.

Re: Tesla Live Stream – Autonomy Day [video]

#302

Post by paulrosenzweig »

ibeckermayer wrote: post #291
sairahul82 wrote: post #114

Karpathy's presentation is really good. Watch it later if you get some time. The key points are - Telsa is using fleet for learning (Fleet learning). This has 3 components, a. Trigger infrasture that collects the kind of data telsa is looking for training. The rational here is we don't need massive amounts of same kind of data but needs right kind of data for training neural networks correctly. b. Data Engine which l…

> People drive vision only. No we don't. We have vision which we then interpret with our _minds_. We have the ability to _understand_ what we are seeing, reason from it, and make decisions accordingly. "Deep Learning", no matter how big your dataset, fundamentally cannot possess this capability. Which goes a long way into explaining why Tesla's "self driving" tech keeps on killing people in situations that are trivia…

> "Deep Learning", no matter how big your dataset, fundamentally cannot possess this capability.

Why not? I’d agree that current models can’t, but what’s the fundamental shortcoming of deep learning?

Re: Tesla Live Stream – Autonomy Day [video]

#303

Post by lojack »

ibeckermayer wrote: post #291
sairahul82 wrote: post #114

Karpathy's presentation is really good. Watch it later if you get some time. The key points are - Telsa is using fleet for learning (Fleet learning). This has 3 components, a. Trigger infrasture that collects the kind of data telsa is looking for training. The rational here is we don't need massive amounts of same kind of data but needs right kind of data for training neural networks correctly. b. Data Engine which l…

> People drive vision only. No we don't. We have vision which we then interpret with our _minds_. We have the ability to _understand_ what we are seeing, reason from it, and make decisions accordingly. "Deep Learning", no matter how big your dataset, fundamentally cannot possess this capability. Which goes a long way into explaining why Tesla's "self driving" tech keeps on killing people in situations that are trivia…

I question how much of our driving is logic and analytics based as opposed to instinct and reaction based.

Re: Tesla Live Stream – Autonomy Day [video]

#304

Post by gthtjtkt »

rohit2412 wrote: post #293
SheinhardtWigCo wrote: post #259

Earlier quoted context omitted.

This argument misses the forest for the trees. Even level 5 autopilot will sometimes crash into objects that are clearly visible to most humans, as will lidar-based solutions. But if the robot fleet is an order of magnitude safer than the human fleet, then it makes rational sense to prefer a robot driver to a human.

I'm getting tired of saying this. Why is it one or the other? Why only a human driver or a robot driver? Could you have a human driver in charge with an accident preventing robot, getting the best of both worlds? If the human falls asleep, the robot can do trivial driving, and the human still makes the more difficult driving decisions. The question isn't if robot car is better than an average car being driven by a hu…

> Could you have a human driver in charge with an accident preventing robot, getting the best of both worlds?

We already have this and we've had it for years. What do you think automatic emergency braking is? Adaptive cruise control? Lane keeping assist?

> The question isn't if robot car is better than an average car being driven by a human. It's if the same robot car would be safer if the human was driving it, and the robot in accident prevention mode. And I don't think that is going to be true in any near future.

No clue what you're trying to say here. Driver assist features don't improve safety?

Re: Tesla Live Stream – Autonomy Day [video]

#305

Post by Qub3d »

cr0sh wrote: post #156
isoprophlex wrote: post #36

Earlier quoted context omitted.

absolutely glorious, they're taking questions. first question: can you use anything else besides ReLU activations in your neural net engine? this is wonderful

I'm not a NN expert, but based on what I have found, the point of using RELU instead of other activation functions is what is called the problem of "vanishing gradients". Basically (IIRC), during backprop the error difference gets ever smaller the further back in layers you go, ultimately getting "lost in the noise", making learning in the earlier layers more difficult to impossible. I'm not saying RELU is the only o…

ReLu is useful, but the question as asked was more about other (also essential) functions for other steps in RNNs -- in particular, sigmoid "squash" functions, as well as MaxPooling.

Re: Tesla Live Stream – Autonomy Day [video]

#306

Post by SheinhardtWigCo »

stefan_ wrote: post #122
grey-area wrote: post #115

Earlier quoted context omitted.

The presentation makes it clear your claim is entirely false, you should watch it.

No, it's spot on. It's entirely what I said: the car can only deliver a few raw frames, and only in response to particular triggers. Notice the cherry-picked examples in the presentation. There is a whole class of problems the field cars can never help with, since they lack the dead-reckoning sensor setup and precise odometry a development car would have.

> There is a whole class of problems the field cars can never help with, since they lack the dead-reckoning sensor setup and precise odometry a development car would have.

Can you give an example? I'm curious what kind of triggers strictly require lab-calibrated hardware.

Re: Tesla Live Stream – Autonomy Day [video]

#307

Post by X6S1x6Okd1st »

xedeon wrote: post #99
iandanforth wrote: post #87

Tesla: "We have a global network of cameras that can be queried to find anything on or near roads and send back photos of it." Every law enforcement entity on the planet: drool

It was pointed out that All data is anonymized.

anonymizing data is non-trival.

Re: Tesla Live Stream – Autonomy Day [video]

#308

Post by dragontamer »

cromwellian wrote: post #258
grandmczeb wrote: post #165

Earlier quoted context omitted.

Convolutions have a much higher arithmetic intensity than dot products.

But in terms of CNNs, it's basically dot product is it not?Can you give an example of a CNN where it isn't?

He has a point. There's a bit amount of data-reuse in CNNs.

Hmm... it will depend on the CNN. There's probably a good neural network design that would take advantage of this architecture. IE: A well recycled convolutional layer that probably fits within the 32MB (load those weights once, use them across the whole picture).

So the whole NN doesn't necessarily have to fit inside of 32MB to be useful. But at least, large portions have to fit. (say, a 128x128 tile with 20 hidden-layers is only 300kB). Recycling that portion across the 1080 x 1920 input would be relatively cheap.

I herp-derped early on, there seem to be CNNs that would make good use of the architecture. Still, the memory bandwidth of that chip is very low, I'd expect GDDR6 or HBM2 to definitely be superior to the 68GBps LPDDR4 chip they put in there.

Re: Tesla Live Stream – Autonomy Day [video]

#309

Post by mlindner »

RivieraKid wrote: post #255
grecy wrote: post #184

Holy cow Elon said a lot of things that will drastically change the world in the next 2-3-5 years. I think it's very, very clear they've been doing a lot of stuff behind that scenes that nobody has given them credit for over these last 5 years. Everyone thought they were just throwing everything at the Model 3 ramp and doing nothing else, but it turns out the opposite is true. At the end Elon said autonomy is basical…

It's so frustrating to see that people eat Elon's BS. Even after: - the Unsworth lie - the "zero concern and I mean zero about 10k Model 3s per week in Dec 2018" lie - the "funding secured" and "only uncertainty is shareholder approval" lies - the "short burn of the century" lies I mean, they were even wrong about Nvidia's chip today, Nvidia had to issue a statement: https://www.marketwatch.com/story/nvidia-says-tesl…

There was no "Unsworth lie". There was Elon saying some very unkind things after Unsworth personally attacked him. Neither thing is appropriate, but let's not pretend this wasn't a tit-for-tat.

Re: Tesla Live Stream – Autonomy Day [video]

#310

Post by GeorgeTirebiter »

Fricken wrote: post #240
metafunctor wrote: post #207

Earlier quoted context omitted.

I think the point was that lidar is too low resolution. You must have very high resolution (in both time and space) to achieve L4 or L5; to accurately determine if a pedestrian is not paying attention, or if a bicyclist is looking to cross the road. Not that Tesla is doing any of that right now, they most likely are not. But a lidar is not going to, either. Not ever, which is the point. I thought the fleet learning a…

Musk would have you think it's a Lidar vs. cameras debate. It's not. It's a Lidar + cameras vs. just cameras debate. Lidar has significant shortcomings. So do cameras. But together they make for a decent perception system. With regards to fleet learning, Tesla has been going on about it for years, but more data is not better. Better data is better. Clifornia's disengagement reports have seen criticism because they pl…

That's correct, it's lidar+cameras vs just cameras. But it should be RF + radar + sonar + lidar + cameras vs anything else. If the 737Max8 tells us anything, it's that one sensor just isn't enough.

EM radiation from UV down to radio has different propagation characteristics, and materials form different attenuators. Radio easily passes through most non-metals with less attenuation than 400-800 THz radiation (light). When I was at Tesla, I tried like heck to get people to internalize this, but the 'cameras + one forward-looking radar is enough' crowd has clearly won.

Heck, RF can shoot through mud, rain, sleet, and snow. Cameras? not so much. Although I agree that the number of 'RF pixels' one gets is nowhere near what a camera provides, it does provide some pixels when the cameras are covered in road muck.

And what about when it's dark? Are the headlights going to provide enough light for the cameras to function well in all road conditions?

Post reply on HN