Suddenly, autonomous vehicles are real. Companies like Google, Baidu, Tesla, Audi and Ford are on course to transform the automotive industry sooner and more profoundly than seemed possible as little as three years ago. The self-driving car’s transition from science fiction to fact is linked with the meteoric rise of artificial intelligence. From healthcare to finance to retail, there is no industry that isn’t being touched by this new era of computing. But, like most ‘overnight’ successes, this intelligent industrial revolution has been many years in the making. The concept of artificial intelligence has been around for decades; Alan Turing first speculated that machines could one day think like humans back in the 1950s. But it’s the combination of research breakthroughs, the wider availability of big data and advances in graphics processing unit (GPU) technology that has ignited the AI explosion taking place today. Deep learning, a form of AI inspired by our understanding of the biology of our brains, is being used to solve extremely complex problems, including developing a self-driving car. There are a staggering number of variables involved in the ‘simple’ act of driving a car. Programming a car to drive itself based on ‘if, then’ algorithms is simply not practical – there are too many possible scenarios. Deep learning is a new kind of computing which offers a powerful solution to this problem. The artificial neural networks used in deep learning have discrete layers, connections and directions of data propagation. Information passes through the network, layer by layer. Each neuron assigns a weighting to its input — how correct or incorrect it is relative to the task being performed. The final output is then determined by the total of those weightings. For example, attributes of a stop sign image are chopped up and ‘examined’ by the neurons — its shape, colour, size and so on. The neural network’s task is to conclude whether this is a stop sign or not. It comes up with a ‘probability vector’ based on the weighting. There are two phases to deep learning: training and inferencing. Returning to our example, the more images of stop signs you use to train the network, the better it gets at identifying them correctly. However, processing the vast amount of training data required to create a sufficiently accurate neural network was, until recently, too computationally intensive to be practical. To the rescue is a piece of silicon originally designed to run 3D video games. GPUs, like artificial neural networks and the human brains on which they’re modelled, can handle multiple tasks simultaneously. They are used in the datacentre to accelerate neural network training. The same GPU architecture is then deployed in the vehicle to run the trained model. This phase is known as inferencing. The model deployed in inferencing is much more streamlined than the network trained in the datacentre. However, because driving is such a complex task, inferencing in an autonomous vehicle still requires a powerful on-board computer. The consequences of losing connectivity while in transit are too grave to rely on a cloud-based solution. With a brand promise focused on safety, it’s significant that Volvo will become one of the first to put autonomous vehicles onto real roads with consumers behind the wheel. Its Drive Me project is expected to begin next spring, putting 100 self-driving XC90s on the roads of Gothenburg, Sweden. The brain of these vehicles is a powerful GPU-based deep learning computer the size of a lunchbox called DRIVE PX 2. Meanwhile Tesla Motors has recently announced and begun production of its newest vehicles that are equipped with the hardware capable of driving autonomously. In addition to a broad suite of sensors, the DRIVE PX 2 AI supercomputer will be software updateable to enables AutoPilot and fully autonomous capabilities in the future. The system will use deep learning to perceive and understand the car’s surroundings. Deep learning has ignited the big bang of modern artificial intelligence. It’s an approach that has proved incredibly effective at solving some of the most complex problems in computer science, and self-driving cars are no exception.
Danny Shapiro, senior director of automotive at NVIDIA