To demonstrate that capability, Nvidia's engineers drove a vehicle equipped with the X1 around San Jose, Calif., for two weeks. That gave the computer an opportunity to see what different classes of objects -- pedestrians, cyclists, motorcycles, etc. -- looked like when they were partially hidden by obstacles.
The processor also learned how to distinguish different types of vehicles: ambulances vs. cars, for instance. "If you detected a car behind you, you might do nothing," Huang said. "But if you detected an ambulance with its lights flashing, you would want to pull over."
Such capability would represent a significant advance over the current generation of microprocessors. But Huang had another trick up his sleeve.
The vehicle's processor could transmit any unfamiliar image back to a central supercomputer, as if phoning a friend for help during a game show. After studying similar images sent by other cars, the supercomputer identifies the object and updates each vehicle's library of images over a wireless network. Ultimately, a driverless vehicle would rarely encounter an unknown object.
Nvidia didn't invent this approach to image recognition. Facebook, for example, uses it for its facial recognition software. But Nvidia claims to be the first to adapt this way of thinking to self-driving cars.
Nvidia's technology holds out the possibility that by learning from and teaching one another, self-driving vehicles could be ready for the unexpected twists of local or urban streets.
In years to come, we're likely to hear a lot more about the "hive mind" and how it works. Stay tuned.