AI Digs Deeper

May 8, 2017

Terry Costlow

Valeo’s CES 2017 booth let humans pit their skills against an AI-based autonomous vehicle on a simulated course (image: Valeo).

Artificial Intelligence has been around since the 1950s, but in many ways it’s still an emerging technology. Large systems for years have utilized various forms of AI, but the technology only now is becoming viable in mass markets like automotive.

AI’s evolution has snowballed in recent years, gaining momentum as more sophisticated techniques could be deployed on more-powerful microcontrollers designed for tasks like graphics processing and image recognition. AI has morphed into variants like machine learning and deep learning, which make it easier for systems to basically program themselves. AI remains something of a catch-all term for the newer variants.

Deep learning is evolving quickly now that microprocessor power and parallel architectures let engineers create systems that better understand inputs from sensors like cameras and Lidar and develop strategies to quickly deal with detected objects. It adds more levels during processing, which helps ensure accuracy while providing more depth of understanding for objects and related decisions. Deep learning, which has been used to improve voice recognition, helps automotive systems learn that hexagonal red signs require a vehicle to stop, for example—and eventually help drive more like a human.

That’s important because no programming team can write all the code needed to identify pedestrians, bicycles, road signs and all the other objects and situations autonomous vehicles need to understand. AI-based systems can train themselves to recognize objects in a range of variables such as weather, lighting and obstructed views. They also can learn how to navigate based on parameters like the location of fixed objects, which can be relatively simply mapped.

Often, AI systems are taught by observing human drivers. But as developers push toward autonomy, they’re devising more ways to create learning programs that live up to their name, with little input from humans.