Autonomous Car - what goes into sensing for autonomous driving?

Listen to Mobileye’s Co-founder, CTO and Chairman, Prof. Amnon Shashua share his thoughts as he talks about autonomous car and about whether an End-to-End deep learning architecture can succeed in the context of autonomous driving.

[@3dsideviewface]

I agree we need to understand more about NNs (especially w.r.t. specific problems) of what they can do and what they are doing, instead of treating them as powerful black box that can do everything end to end. And I guess a lot of researchers indeed follow this path to improve NNs and propose new architectures. The essential problem with black box end2end is that it requires much more data to train (not sure it is exp growth or not in autonomous driving though), but with domain priors *appropriately* encoded (in architecture or learning), it may be trained more efficiently and reliable. Think about human, we learn by summarizing rules and learn from rules, not just guessing what to do by only intuition.

[@aswad7368]

Where Can I find this presentation?

[@heltok]

Interesting talk! I am not sure if I agree with the conclusions though. While the DNN might not recognize the rare vehicles as rare vehicles, it seems likely that it will activate some neurons of similar concepts, with enough of these I would guess that it should be enough for it to take a correct action.

[@usacc8400]

!!! love this guy

[@DingshengLi]

I thought you guys are against "hands-free" driving with the current level of technology as early as 2015 based on recent statements? Why it seems to me you're OK with it in this 2016 video?

[@RiderVSRoad]

Don't worry, they just forgot to delete this video, it will be gone soon.

[@nogtname123]

XZ

[@AdamWood]

Guy who works with computers doesn't know that computer processing power grows exponentially over time. Amazing.

[@etiennetiennetienne]

What a smart talk! But I still have doubts. I'm not sure to buy the first argument. I mean, why do we bother to do end-to-end for object detection? Don't you think we could argue the exact same thing? "it's rare to see a hat on a lady older than 50 years old so it is likely I should teach my system to recognize a hat and a lady before I teach him to recognize a lady with a hat". Isn't he missing the concept of feature invariance, generalization & possibly transfer learning? Or Am I completely trapped in the "deep learning hype"?

[@carlesgelada2043]

+Etienne Perot You are right, the first argument is basically bullshit. The second is completely true tough.

[@etiennetiennetienne]

Thanks! Feels good to not be the only one :-) . I just googled nbit parity w nn : www.sciencedirect.com/science/article/pii/S0893608099000696 ? & www.eng.auburn.edu/~wilambm/pap/2003/IJCNNparity.pdf. && multiply seems to be there also : www.lcc.uma.es/~lfranco/A1-Franco+Cannas-1998.pdf ?

[@carlesgelada2043]

+Etienne Perot Well the papers you cite do not use a learning mechanism. They use hardwired connections to implement those algorithms. I am not sure if someone has tried to use RNN with memory to this problems, but i am pretty sure they would perform well.

[@etiennetiennetienne]

+Carles Gelada Ya... interestingly i don't find any modern paper on that. and old sources only talk about feed forward fully connected nets...I think you are right, intuitively RNN should be suited to the carry issue

[@krimdelko]

End to End works, but it's too risky since rare events can cause false negatives or false positives. Take algorithmic trading for example, typical example for end to end approach to stock trading. It works until it doesn't. Then all hell breaks loose. Autonomous driving cannot take such risks according to Amnon Shashua. Of course Mobileye is also talking its own book here since they have a lead in semantic segmentation of the problem by dominating vision through cameras.

[@davidrhallsr]

Air plane

[@SuriyaNarayanan987]

Is end to end learning for self-driving cars really impossible? this link is worth checking out: arxiv.org/pdf/1604.07316v1.pdf

[@carlesgelada2043]

+Suriya Narayanan This paper does not propose a driving system, it proposes a steering system. And end to end net for driving will require several major breakthroughs. And i am pretty sure that the first batch of self driving cars won't use them.

[@SuriyaNarayanan987]

+Carles Gelada yeah true.. only steering is shown to work using end to end architecture.. but isn't that itself disproving what the author in this video claims? I think it's only a matter of time to incorporate the other controls in end to end arch for self driving

[@carlesgelada2043]

+Suriya Narayanan Some of the things that the author claims are plainly wrong, as evidenced by this paper. And i do believe that end to end is the future. But you have to realize that complexity of driving. It requires the ability of reading the drivers manual, interpret a set of rules and be able to apply them while driving. And it would also need to do path planning, and natural interaction with people etc. It is a huge problem, so big that we don't even have to tools to tackle it.

[@SuriyaNarayanan987]

+Carles Gelada haha.. I don't agree

[@carlesgelada2043]

Suriya Narayanan If you can do it, go ahead. You will be one of the most important AI researchers of the century.

[@SuriyaNarayanan987]

DNN architecture is such that it mimics how the human brain works.. Of course we humans are not great at multiplication of n-bit parity or whatever example is quoted in this video while our brain is good at object recognition.. I feel the outcome of this talk is not directed towards the right direction

[@abdul11235]

+Suriya Narayanan He was hinting that the black box approach certain companies (comma.ai) are taking doesnt do well for the last 0.1 or 1 % required for completely autonomous driving

[@SuriyaNarayanan987]

+Hello world in fact mobileye's path planning solution is based on end to end architecture and not that of component based.. there is a paper by mobileye on this.. you may Google it

[@KangZhang]

Can you provide the name of the paper ? For the DNN part, I think the key point for building autonomous system is explainable. The system can fail but you must explain when will it fails and why it fails. However, DNN is still very hard to explain.

[@computerex]

The example mentioned is simple multiplication. I'm terrible at math, but even I know that 4x2 = 8. Also DNN do not mimic the human brain. Neural networks were inspired by neurons, but that's where their similarities end.

[@bobinorlando]

Loading up a car with all this ridiculous tech is akin to creating an autonomous finger - you know disconnecting your finger and putting a brain and eyeballs in it and oh while you're at it how about a couple of ears and legs too. What's missing is the higher level of intelligence in the system it is part of and must be connected to. A smart finger is nothing without the rest of the body which already has eyes ears and legs. And the smart car - really a dumb car by the time you're done because the ultimate function - the hidden agenda - of all these eyes and sensors is to do nothing more than essentially drive in lock step with the cars around it thus diminishing it to the primordial state of a non-autonomous train car. Meanwhile, the traffic control system lingers about running and ruining everything in the same old way because these geniuses never thought about improving the meta-level that governs all vehicles. If they spent even 10% of their time improving the meta level they could solve 90% of the problem.