Programmers trying to write all the possible scenarios needed for truly autonomous vehicles face a daunting challenge. Deep learning, in which computers learn a desired behavior using artificial intelligence and neural network concepts, could be a viable solution.
Speakers at the recent Nvidia GPU Technology Conference suggested that it would be more efficient to have machines teach themselves how to respond to the infinite changes that occur as autonomous vehicles drive from site to site. Deep learning has been used by companies such as Google and Netflix to help predict what consumers are looking for. Some observers feel various forms of machine learning will be a disruptive technology that changes many fields, possibly including automotive.
“Deep learning is arguably as exciting as the creation of the Internet,” said Jen-Hsun Huang, Nvidia’s CEO. “Humans can’t code if-then-else statements for all the situations vehicles face. We want to augment today’s advanced driver-assistance systems with deep learning systems that will learn the behavior of drivers over time.”
Deep learning lets machines program themselves to classify objects. In automotive systems, deep learning can be used to create a database of images and actions to be taken given their position in relation to the vehicle.
The U.S. Department of Defense's Defense Advanced Research Projects Agency (DARPA) used the concepts on Project Dave, in which a small autonomous vehicle learned how to plot a route to avoid obstacles that would prevent it from getting to its destination. Other researchers are starting to study the technology.
“We are starting an exploration phase; we need to have a better understanding of how deep learning can be used,” said Uwe Higgen, Head of the BMW Group Technology Office USA. “It has much potential, but we need to study it to learn how drivers can benefit.”
Analysts note that it will take a few years to commercialize the technology. Deep learning relies on multiple layers of analysis and pattern matching, so it requires powerful parallel processing chips. Semiconductor capabilities will have to advance along with the software before systems can meet the real-time demands of vehicle safety systems.
“Deep learning is a good solution given the many different objects that vision systems will have to analyze, but it’s a long way from meeting the low latency demands of automotive safety systems,” said Ian Riches, Global Automotive Practice Director at Strategy Analytics.
If and when deep learning systems are included in cars, they will probably continue to learn. Some speakers noted that these systems can shadow human drivers, comparing what the control system selects to what human drivers do. Systems can also share their experience with other systems and human overseers.
“When a system encounters a situation it’s never seen before, it can record it and upload the data to the cloud,” said Danny Shapiro, Nvidia’s Senior Director, Automotive. “Later on, it can download the solution over the air. Initially, I think we’ll see a lot of information shared within fleets, whether the fleet is for an automaker or a taxi cab or trucking company.”
Tesla CEO Elon Musk said that many of the challenges of autonomous driving have been solved, but noted that in-city driving remains a challenge. That’s because there are far more variations at medium speeds in cities than in traffic jams or freeway driving.
“Where (autonomous driving) gets tricky is in open environments where you’re going say, 20-40 mph,” Musk said. “You can handle 5-10 mph driving with ultrasonic sensors, but driving at 10-50 mph in urban environments, where a lot of unexpected things happen, is difficult. Above 50 mph, in a freeway environment, it gets easier again.”
Musk, questioned by fellow CEO Huang, declined to comment on deep learning’s potential for autonomous driving. When Huang noted that Musk had been quoted as saying artificial intelligence was potentially more dangerous than nuclear weapons, Musk sidestepped the question.