There are two approaches to making a machine intelligent. Experts can teach the machine what they know, by imparting knowledge about a particular field and giving it rules to perform a set of functions; this method is sometimes termed knowledge-based. Or engineers can design a machine that has the capacity to learn for itself, so that when it is trained with the right data it can figure out its own rules for how to accomplish a task. That process is at work in machine learning. Humans integrate both types of intelligence so seamlessly that we hardly distinguish between them. You don’t need to think about how to ride a bicycle, for example, once you’ve mastered balancing and steering; however, you do need to think about how to avoid a pedestrian in the bike lane. But a machine that can learn through both methods would require nearly opposite kinds of systems: one that can operate deductively, by following hard-coded procedures; and one that can work inductively, by recognizing patterns in the data and computing the statistical probabilities of when they occur. Today’s A.I. systems are good at one or the other, but it’s hard for them to put the two kinds of learning together the way brains do.