学进去

By the age of seven months, most children have learned that objects still exist even when they are out of sight. Put a toy under a blanket and a child that old will know it is still there, and that he can reach underneath the blanket to get it back. This understanding, of “object permanence”, is a normal developmental milestone, as well as a basic tenet of reality. It is also something that self-driving cars do not have. And that is a problem. For a self-driving car, a bicycle that is momentarily hidden by a passing van is a bicycle that has ceased to exist.

This failing is basic to the now-widespread computing discipline that has arrogated to itself the slightly misleading moniker of artificial intelligence (AI). Current AI, based on the idea of machine learning, works by building up complex statistical models of the world, but it lacks a deeper understanding of reality. Similar techniques are used to train self-driving cars to operate in traffic. Cars thus learn how to obey lane markings, avoid other vehicles, hit the brakes at a red light and so on. But they do not understand many things a human driver takes for granted—that other cars on the road have engines and four wheels, or that they obey traffic regulations (usually) and the laws of physics (always). And they do not understand object permanence.

In a recent paper in Artificial Intelligence, Mehul Bhatt of Orebro University, in Sweden, describes a different approach. He and his colleagues took some existing AI programs which are used by self-driving cars and bolted onto them a piece of software called a symbolic-reasoning engine.

Instead of approaching the world probabilistically, as machine learning does, this software was programmed to apply basic physical concepts to the output of the programs that process signals from an autonomous vehicle's sensors. This modified output was then fed to the software which drives the vehicle. The concepts involved included the ideas that discrete objects continue to exist over time, that they have spatial relationships with one another-such as “in-front-of” and “behind”—and that they can be fully or partly visible, or completely hidden by another object. The improvement was not huge, but it proved the principle. And it also yielded something else. For, unlike a machine-learning algorithm, a reasoning engine can tell you the reason why it did what it did. A machine-learning program cannot do that. Besides helping improve program design, such information will, Dr Bhatt reckons, help regulators and insurance companies. It may thus speed up public acceptance of autonomous vehicles.