Computer learning systems are solving big problems. The programmer doesn't tell the computer how to proceed, but merely provides massive amounts of data for the computer to learn with. The computer sets about blindly looking for patterns in the data and sometimes finds patterns that people would never see.
- choosing a move in chess or go
- diagnosing a patient given a constellation of symptoms
- telling how old a person is by looking only at the microbes in his poop
On the one hand, it's great to to have someone who really "thinks different" looking at the data and making suggestions. On the other hand, it's a human being who is going to use this pattern or (formula or algorithm of diagnosis or plan) at the end of the day, and often the stakes are high. How is the human to know that this crazy idea the computer came up with isn't just an artifact of the data, or a programming error?
"Hal, explain yourself!"
Over at Google Brain, Been Kim has worked on this problem, and now has (the beginnings of) an interface to query the computer, so we might learn not only what the computer's solution is, but how it got there