Self-Driving Cars [S1E3: AI Failure?]
Вставка
- Опубліковано 1 лют 2019
- PATREON: / welchlabs
TWITTER: @welchlabs
MORE: www.welchlabs.com
CODE: github.com/stephencwelch/self...
FURTHER READING + REFEERENCES
For a bayesian approach to the same reliability testing math: stats.stackexchange.com/quest...
Rand Report: www.rand.org/pubs/research_re...
Zeiler and Fungus 2013: arxiv.org/abs/1311.2901
WHO top 10 causes of death: www.who.int/news-room/fact-sh...
WHO injuries and violence facts: apps.who.int/iris/bitstream/h...
Bojarski et al: arxiv.org/abs/1704.07911
Dean Pomerleau: www.cs.cmu.edu/~bhamner/neural...
Dean Pomerleau: www.ri.cmu.edu/pub_files/pub3...
Dean Pomerleau: www.ri.cmu.edu/publications/l...
Nvidia End to End: arxiv.org/abs/1604.07316
Way ChauffeurNet: arxiv.org/abs/1812.03079
VIDEO REFERENCES
Amnon Shashua Talk: • Autonomous Car - what ...
NAVLAB: • Navlab 97: Automated a...
MUSIC
www.premiumbeat.com/royalty-f...
www.premiumbeat.com/royalty-f...
www.premiumbeat.com/royalty-f...
SPECIAL THANKS TO
Tony Fast
Krish Ravindranath
Karthik Naga
Alex Mones
Michael Grabchak
- AND -
Sid Sarasvati
Ross Hanson
Vin Soma
Antoine Pintout
Jaewon Jung
Raphael J Vasquez
AJ Englehardt
Nate Fuller
Hitoshi Yamauchi - Наука та технологія
![Self-Driving Cars [S1E4: RALPH]](http://i.ytimg.com/vi/DIAKL0s4Mzs/mqdefault.jpg)
![Self-Driving Cars [S1E4: RALPH]](/img/tr.png)
![Channel Update [April 2017]](http://i.ytimg.com/vi/gZPy7OxAjRM/mqdefault.jpg)
![Learning To See [Part 15: Information]](/img/n.gif)
You are providing a public service. Hope you can keep going with this. Please show the good and the bad without prejudice. Development of autonomous driving will be crucial for a lot of elderly like I am.
We like to keep our independence.
Can't wait for the next video!
2nd
same
He is gone :(
Phenomenal! Your story telling and technical accuracy and fantastic.
Your series of videos are insightful and easy to understand concepts. Thank You!
Perfectly well done. Awaiting your next video
Another a cliffhanger!! I’m becoming a patron to get the next episode!!! So glad you’re back! #bestchannelever
Alison Young love it so much #fangirls
Alison Young who let you guys on here? :)
Loving the complex insights made easy.
This and the next video has not been added to the playlist of self-driving cards.
I haven't seen ads for nearly a decade... I always disable adblock before watching your videos... and I don't skip the ads.
I would have supported you on patreon but I'm broke.
Madao Wow, hopefully the ad was bearable - thanks!
@@WelchLabsVideo I have only one wish: Please don't stop making videos.
Great Work and thank you very much. Always like the way Welch Labs explain complex idea such a simple way. Keep providing and keep growing.
Your teaching skills are very impressive (you should teach somewhere for sure!). You make a good summary of Shashua's arguments about the problems of end to end learning for autonomous vehicles. However, I have some counter arguments:
1) Nvidia graphics hardware can simulate cameras (and other sensors) very accurately. They can simulate both the lighting and the environment to produce synthetic images that cannot be distinguished from the real images. They have simulated the moon landing pictures so accurately that they debunk the idea that it was faked. Nvidia technology can therefore simulate what any autonomous vehicle would have as input from every sensor. This is analogous to using flight simulators to train pilots; this works because the simulators are very good.
2) Even if the simulated data is not perfectly accurate because of transfer learning neural networks can still be trained very well. There are papers showing good training results on completely simulated data, but we do not have to go this far. Conservatively, very good results can be obtained on around 90% simulated data, and 10% real data. By analogy pilots trained on flight simulators do need some time on real airplanes, but they can spend the vast majority of their time on the simulators.
3) Data simulators can easily create data for the "edge" conditions that are rare in practice but are very important for safety. Shashua claims that certain edge conditions, such as a big tractor on the side of the road would cause problems for any end to end system. The reasoning is that such an edge condition is so rare that it would require a vast amount of physical data to be collected for it to occur naturally. But when using simulated data because we can directly enumerate such "edge" conditions. In the flight simulator analogy they focus mostly on the rare disaster conditions (i.e. all engines fail). They do not need to spend much training time on the vast majority of "easy" situations.
4) We can do the training and testing of neural networks completely in parallel. Very many miles of simulated data will be required for good results but we can easily replicate the computing environment. So requiring a large number of simulated miles will not be a problem. A government could even set up a testing environment for the use of different manufactures of autonomous vehicles.
5) Once there is enough training data and the neural networks are big enough they have outperformed customized feature based systems in computer vision for many applications. I don't see why this will not be the case eventually for autonomous vehicles. An obvious start is to pick a relatively small environment, such as a downtown core, or say a race track for testing. Then build a system which drives autonomously but completely end to end in all weather conditions. Finally compare its performance to the best but more traditional hybrid system. As more data is collected, and hardware becomes faster, just keep growing the size of the environment.
6) As more training/testing data is created governments can force this data to be shared among manufacturers. Then the end to end systems will all steadily improve in performance. Improving a traditional decomposition based computer vision solution in this way is not so simple. This requires the sharing of algorithms which are considered to be intellectual property, and not just the sharing of data.
7) I am sure that Shashua is aware of the arguments I have made above but he does not address them in his talks. This is logical because they involve technology from Nvidia, not Intel. He is after all, a VP of Intel.
8) To conclude completely end to end learning is not yet practical for autonomous vehicles. Therefore some hybrid approach is the only option. However, the argument that completely end to end learning will never be practical is not convincing. End to end systems for autonomnous driving are practical now for small environments. As time goes on the end to end approach will become practical for larger and larger environments. My opinion is that an unconstrained end to end trained autonomous car that outperforms hybrid systems is inevitable. But how long it will take is not clear. Remember that current neural networks have many orders of magnitude fewer neurons than the human brain.
That's exactly what I was thinking while watching the video. You can't patch the weights but you can "patch" the inputs.
Lots of great points here. I struggle with a few of these myself while writing. I definitely cannot prove your argument wrong, and I wasn't 100% convinced by Shashua's argument. However, after my research, as I say in the video, I don't believe that we're going to see production end-to-end any time soon, and maybe ever. I just don't believe that current deep learning techniques are going to scale that well with data. But, I could definitely be wrong! Thanks for watching + commenting, i'm happy to start a conversation about this stuff, I think it's important to talk about at SDCs become a reality.
@@WelchLabsVideo In 2010 only one in one hundred computer vision researchers thought that deep learning would be as successful as it is now. There is no obvious reason that the cycle of more computer power, bigger networks and more training data will not continue to create performance improvements. Only time will tell, but since companies like Tesla, which have driven over 1 billion miles, have so much training data these questions will not take long to be answered.
love this series
Great serie of video! keep going!
Thanks for this video!
Why only fatalities are taken into account in the calculation? Injuries or collisions in general are good enough metrics and give far faster way to validate self driving cars abilities.
I love your videos alot
I am 12th standard and i literally dont understand anything about this clearly but the thing that makes me love ur videos is that i am trying to understand something new
Why base improvement only on fatalities- when accident injuries are so much higher and still serious?
Informative and amazing as always
Stephen, now I became your patreon ;) but pls keep doing this videos
Deserves 10 times more view, pity still not 30000 views
Great Video!
Make a video on convergence and divergence of a sequence and series .please
Thanks a lot.
Thanks as always.
Thank you for your video series.
How awesome would it be of we could go into the future and collect some weights for an awesome driving neutral network.
How would you define an "Analytical approach" and an "Empirical approach" ?
Even though there are/will be better ways, I want to ask, to collect training data what if we travel simultaneously (on different type of roads) to decrease the time required (the 12 something years and 500 million miles/kms).
와 재밌다 고마워요!!
Oh you are such a tease!
But what if we use selfdriving cars on highways and in cities? It would be very easy to define where the vehicle is allowed to drive and if every car is a self driving car, we wouldnt even need a optical recognition system, because we can have something like gps which defines the position of each vehicle in the street really well. There would be no room for errors right?
kids on the road surely won't have built in GPS
If you store all the LiDAR data that cars capture you can eventually make a hyper-realistic model of the US, and reach those 275'000'000 miles virtually.
I can't get over your voice. Not that it's weird, I quite like it, but I feel like it really does not belong to the person talking to the camera. Can't shake the feeling that this is dubbed :D
Best channel ever
I'm doing a project on literally this. I'm guessing Hidden Markov Models will do the trick?
I doubt that. The problem is training data composition, and model validation, not modeling in itself.
One approach to make validation tractable is decomposing the overall system (suggested at the end of the video.)
However, "more modeling" (as in CNN/HMM/RNN/LSTM) doesn't really change that.
An intervention of a human counts, in general, as a failure of the autonomous driver. The other way around, a warning or possible (suggested) intervention by an autonomous driver could count as a success
tl;dw
you can't debug one big neural network, so it's better to make more, smaller ones that you can replace when faulty
When an autonomous vehicle is involved in an accident then who do you blame? Algorithm or Algorithm developer?
yes
The standard of fatality is not necessary for testing safety. A better standard would be a traffic accident. There are approximately 20 traffic accidents each year per traffic fatality in the US. This is a much easier number to reach. On top of this Tesla has announced that their fleet has accumulated over a billion miles on Autopilot, already more than the numbers you cite as needed. So real world testing and analysis is very achievable.
I agree that fatalities alone do not tell the whole story. My reference for this argument (www.rand.org/pubs/research_reports/RR1478.html) also does the math for injuries and accidents. When you factor in that SDCs do not drive perfectly, you still need a heck of a lot of data to validate these things. I think the fatality argument is a good starting point (that's why I picked it), but there's definitely more to it. Thanks for watching!
@@WelchLabsVideo The link to the reference says "Page Not Available".
Hoe about this: www.rand.org/pubs/research_reports/RR1478.html
Just get better!
We'll need to do most of the test driving "hours" in VR and in computer time, not real time, just like how we trained AlphaZero to play chess. This way, the AI will actually have more hours of experience and data than all real driving combined.
VR is not only a safe way to train humans and AI, but it's' also a super fast way to train AI.
The VR footage won't be nearly the same as the real life footage from cameras on a car. Because of that I don't think training via VR will be practical. On top of that we can better use the footage from people who are gonna drive anyways instead of making people drive in VR for the sole purpose of training the AI
@@frisosmit8920 What? "making people drive in VR for the sole purpose of training the AI"? I don't know what you're talking about. My guess is that you misunderstood what I was saying. I was addressing his point about the need for millions of hours of testing. This could only be done in VR, so that the millions of hours is actually virtual hours done in the real time of say a single day on a supercomputer.
@@KittyBoom360 ah that makes more sense.
I still think it would be hard to simulate real life accurately enough in VR to make it usable for testing. Even if you think it is accurate you can't be sure, so you still need to do quite a lot of real life testing to be sure your car won't crash.
aaaaaaa next video quick
3:31
Thanks so much bro. www.deeplearningbook.org/ is the book that is shown in this video.
Love the quality of the video, very interesting and well explained.
If you looking for more visibility, I would advise you to change your microphone or do something about your tone of voice. I might be the only one but I find your voice quite monotonous.
Tesla already has 1 billion miles, with increase in car production from the model 3 they should reach 5 billion in no time.
10
Wpmt AI be very susceptible to purposely crafted optical illusion’s? For example, consider the flat painting on a road that is meant to make a human think that there is a little girl running across the street. Would AI be able to determine that what appeared to be little girls running across the street we’re actually just optical illusion’s? Would an artificial intelligence driver ever have enough common sense to realize that someone was purposely trying to mislead them? Considering that there have already been instances where people have attacked self driving cars, I think it’s a real possibility that people will attempt to craft optical illusions that cause artificial intelligence drivers to crash.
Couldn’t we drive those billion miles with millions of cars in one day using I simulated reality? So with CPU’s running billions of processes per second, we should be able to run billions of trial runs a day.
I'm sorry... I lost you at 0:-01
Electromorphous aw man!
Most miles driven are super boring. That also means that learning systems don't need as much training from them.
If we can isolate the "hard" parts where humans fail (or machines fail,) we need much less data for training than pure statistics say we would.
This is mainly because the statistical assumptions are wrong, because the problem is not homogenous or well distributed.
Wtf. Tuberculosis takes so many lives each year
Стас Бицько word!
3:14 - I think you're missing a few zeros there.
i'm studying computer science right now and soon I ma help yall build dem selfdriving cars
and I will make ai and take over the world
Musical backgrounds might be unnecessary here, slightly distracting.
Interesting content, good video, gave it a thumb up, thank you!
:o)