Where Machine Intelligence Is Going Next

Making Algorithms Explain Themselves

MI systems today are largely opaque—there is no way, even for their creators, to understand how they arrive at the inferences they make. MI researchers are working to address the problem of "algorithmic black boxes" in a few ways. One method bypasses the problem of opening up algorithms completely by instead testing their outputs to detect bias. In 2016, researchers at Google and the Toyota Research Institute co-authored a paper detailing this approach, saying, "Our criteria does not look at the innards of the learning algorithm. It just looks at the predictions it makes."

Enabling Machines to Do More with Less Data

Researchers are working hard to reduce MI's massive requirements for data. In 2016, the International Conference on Machine Learning (ICML), a prestigious academic conference first held in 1980, hosted a workshop on data-efficient machine learning to tackle the problem of enabling machines to "learn in complex domains without large quantities of data." Participating researchers explored a variety of techniques for achieving this goal. One promising method is generalized learning, in which machines take knowledge from one domain space and apply it to another rather than receiving an entirely new data set to train on.