Chapter 3. To Supervise or Not to Supervise in AI?
One of the truisms of modern AI is that the next big step is to move from supervised to unsupervised learning. In the last few years, we’ve made tremendous progress in supervised learning: photo classification, speech recognition, even playing Go (which represents a partial, but only partial, transition to unsupervised learning). Unsupervised learning is still an unsolved problem. As Yann LeCun says, “We need to solve the unsupervised learning problem before we can even think of getting to true AI.”
I only partially agree. Although AI and human intelligence aren’t the same, LeCun appears to be assuming that unsupervised learning is central to human learning. I don’t think that’s true, or at least, it isn’t true in the superficial sense. Unsupervised learning is critical to us, at a few very important stages in our lives. But if you look carefully at how humans learn, you see surprisingly little unsupervised learning.
It’s possible that the the first few steps in language learning are unsupervised, though it would be hard to argue that point rigorously. It’s clear, though, that once a baby has made the first few steps—once it’s uttered its first ma-ma-ma and da-da-da—the learning process takes place in the context of constant support from parents, from siblings, even from other babies. There’s constant feedback: praise for new words, attempts to communicate, and even preschool teachers saying, “Use your words.” Our ...