Nvidia GPU-Powered Autonomous Car Teaches Itself To See And Steer (networkworld.com)

An anonymous reader quotes a report from Network World discussing Nvidia's project called DAVE2, where their engineering team built a self-driving car with one camera, one Drive-PX embedded computer and only 72 hours of training data: Neural networks and image recognition applications such as self-driving cars have exploded recently for two reasons. First, Graphical Processing Units (GPU) used to render graphics in mobile phones became powerful and inexpensive. GPUs densely packed onto board-level supercomputers are very good at solving massively parallel neural network problems and are inexpensive enough for every AI researcher and software developer to buy. Second, large, labeled image datasets have become available to train massively parallel neural networks implemented on GPUs to see and perceive the world of objects captured by cameras. The Nvidia team trained a convolutional neural network (CNN) to map raw pixels from a single front-facing camera directly to steering commands. Nvidia's breakthrough is the autonomous vehicle automatically taught itself by watching how a human drove, the internal representations of the processing steps of seeing the road ahead and steering the autonomous vehicle without explicitly training it to detect features such as roads and lanes.

Comments

Inherent shortcomings

[fyngyrz]

In situations that do not resemble the training data, the network's response is essentially undefined, as well as unknown (it's all unknown... an NN results in behaviors that are not deterministic in the sense that anyone planned them out -- they are what they are, that's all.)

Nice experiment, though. :)

Re:Inherent shortcomings

[fyngyrz]

No, not like people. An NN of the type described doesn't learn further; won't generalize from other experiences (FI, this thing won't know what to do with a pedestrian unless it's observed the human driver stopping for some number of them. Otherwise it has no training that places value on a pedestrian. Same for cat, chunks of some trucks blown-off retread, potholes, etc.) A person will know from their far, far deeper experience that it's inherently a bad idea to run down the three year old that walked in front of the car, will usually at least try not to run down little Susie's pet cat, and will choose to go around the pothole if it is reasonable to do so.

An NN is a low-dimensional approach to solving problems. But most non-trivial problems -- and driving a vehicle is an entirely representative proxy for such problems -- tend not to be uniformly low-dimensional. Many aspects come into play suddenly and unpredictably. Some might not be seen for years, or ever. But then again, they might. In order to deal with such things, more than low-dimensional problem solving is required. NN's can't do it. They're inherently limited.

It's going to require very high level software. Not intelligent, but damned well informed and replete with a huge rulespace that can solve all of the types of reasonably solvable problems in the space.

That's all IMHO, of course. But I'm right, so there's that. :)

Re:Inherent shortcomings

[laing]

"Otherwise it has no training that places value on a pedestrian." /p It's been years since I played Death Race 2000 (the video game, not the movie). How many points is a pedestrian worth these days?

Re:Inherent shortcomings

[lorinc]

An NN is a low-dimensional approach to solving problems. But most non-trivial problems -- and driving a vehicle is an entirely representative proxy for such problems -- tend not to be uniformly low-dimensional. Many aspects come into play suddenly and unpredictably. Some might not be seen for years, or ever. But then again, they might. In order to deal with such things, more than low-dimensional problem solving is required. NN's can't do it. They're inherently limited.

No, they're not. The lack of proper response to an unseen before event (which is called bad generalization) can come from at least 2 things: a bad sampling for the training set (no pedestrian in the set), and the absence of transfer learning (use of another system trained on pedestrian to improve the first one). The first one is really easy to solve but time consuming, a lot of very bright people are working on the second.

Honestly, automatic driving by visual cues is not longer a challenging computer vision problem. It is still a challenging engineering problem, but all the scientific tools needed to solve it are already there, and they are being actively used to solve it.

There are a lot of computer vision tasks that are now to be considered as solved from the research point of view. Very few people seem to acknowledge the enormous progress that have been made in the field in the last 10 years. Even people working in the field. But it's there, and now the engineers have to use it.

Now there's a Rick and Morty line

[the_skywise]

"What is my purpose?"
"You drive me places." ... "Oh God!"
"Welcome to the club, pal!"

https://www.youtube.com/watch?...

Re:Nvidia

[Anonymous Coward]

Last I checked NVIDIA's market cap was 20.3B$. You'd think they have the ability to do more than one thing at a time, especially if it returns a profit on investment in the future.

This sort of AI usage is only going to increase, it would be dumb of NVIDIA not to at least divert a miniscule amount of manpower and resources to try and get their feet wet in this field. Anything they can patent or turn into a product will outweigh the paycheck for the engineers on this "worthless" venture.

Heh

[MobileTatsu-NJG]

We all know this car's running Windows cos Linux ain't got no good nVidia drivers!

Re:Heh

[Trogre]

Funny, but wrong.

nVidia hardware performs at least as good under Linux as Windows, including CUDA processing.

Re:Computers don't learn

[slew]

Sorry, there is no algorithm that makes algorithms..

Although there might not be an algorithm that makes algorithms, there are algorithms to configure a meta-algorithm implementations. And example meta-algorithm implementation would be a deep neural net, or a human brain. I don't think it is a stretch to call the algorithm used to configure meta-algorithm implementation "learning" (although commonly this is called training)...

But this is merely a semantic point.

Not a Luddite fear, but a valid fear

[GPS Pilot]

Here's an excellent article by Ashlee Vance about a self-driving neural net-based system created by one man, George Hotz: http://www.bloomberg.com/featu...

It contains one passage that sums up my fears:

Hotz hadn't programmed any of these behaviors into the vehicle. He can't really explain all the reasons it does what it does. It's started making decisions on its own.

This will come back to bite us. More than once. Systems can exhibit unexpected behavior even when the inventor has an excellent understanding of the invention; here, a very bright inventor seems to have no hope of fully understanding things.

Unexpected behavior from the control system of an object that has a lot of kinetic energy is usually bad. And a car is not the most dangerous thing you can put a neural net in charge of.