Gadgetory

welcome we're going to talk about self-driving cars today last year I said that in the future when you build cars it's going to be a lot more like a computer computer vision technology that had evolved up to this point are going to be difficult to continue to advance to the point where we could have cars that drives itself all of this computation capability is going to have to be done in real time all of this we've been working on the last year ladies and gentlemen the Nvidia drive px 2 this is the world's first in-car AI supercomputer and it's designed to make it possible for us to realize the vision of self-driving cars the computational capability of px 2 is equivalent to a hundred and fifty MacBook Pros all together this entire supercomputer fits in your trunk very nicely the size of a lunch box our vision is to make it possible for us to finally realize a self-driving car humans are the least reliable part of the car we represent almost all of the fatalities that are caused around the world over a million deaths each year or replacing the human altogether self-driving car technology is surely going to make a great contribution to society the biggest problem is perception first of all what is happening around me what are things that I should be concerned about and how should the car deal with it the folks at Nvidia research work together to make it possible to take advantage of the CUDA GPU that we invented we were able to accelerate the training by 30 to 40 times deep learning is able to achieve super human perception capability but it is now possible for us to train these incredibly complex networks to recognize objects of all kinds just to show it to you let's now see what it can do it took a month to train the original network with the imagenet data set without GPU acceleration that month would have been a couple of years this is an upcoming data set will be publicly available soon the cityscape data set there's more training images and they're very very finely segmented and detailed it's a very modern data set not one feature detector was coded by hand it's basically like holding up millions and millions of flash cards to the computer and telling it to learn and basically nudging in the right direction when it gets things wrong so this gives you the next level of perception saying well you know what can I Drive on what is this thing at this pixel that I'm looking at so it's a much more robust way to handle perception in a car but there's so much more to do there's so much more to do now we do we want to be able to recognize objects we want to be able to recognize circumstances every bus is not the same their passenger buses public transit buses where it's okay to just drive by school buses you should stop every truck is not a truck some trucks or ambulances and you should pull aside all of these different scenarios not only do you have to recognize what they are but understand the circumstances the special circumstances these are all trainable things so whereas today we're focused on training detection perception very shortly we're going to move towards training for recognizing circumstances