Tuesday, August 16, 2016

Deep Learning, General AI

François Chollet, Deep learning researcher at Google

[Answering the question "is deep learning overhyped?"] In many respects, it is. For sure, the recent successes of deep learning have been amazing: we went from being really terrible at supervised learning on perceptual problems (image classification, speech recognition) to being really good at it. Deep learning has been transformative for many subfields of machine learning. But here's the thing: lots of people, most of them not directly involved with deep learning research, tend to extrapolate too much from these recent successes. For instance, when we started achieving below 4% top-5 error on the ImageNet classification task, people started claiming that we had "solved" computer vision. We most certainly haven't solved computer vision at this point; it's still a tremendous challenge to generate accurate, precise descriptions of the contents of a picture or a video, or to get meaningful answers to basic visual queries (e.g. "get me a close-up of the handbag of the second lady from the left"), things that humans take for granted. Our successes, which while significant are still very limited in scope, have fueled a narrative about AI being almost solved, a narrative according to which machines can now "understand" images or language. The reality is that we are very, very far away from that.

In the pitches of startups that are attempting to cash in on deep learning, I see a lot of grossly unrealistic expectations. Some of them are just naively over-optimistic, but some others are essentially living in a fictional universe —I've seen at least 3 different startups state that they would solve "general artificial intelligence" in the next few years. Best of luck to them. Most of these companies have no issue getting generously funded, but quite a few of them will find it very difficult to get a decent exit. A lot of disappointment will follow, especially among VCs and corporate decision makers, and unless this is counter-balanced by a larger wave of successful value-producing applications of deep learning, then we might witness a new AI winter in the future.

Overall: deep learning has made us really good at turning large datasets of perceptual inputs (images, sounds, videos) and simple human-annotated targets (e.g. the list of objects present in a picture) into models that can automatically map the inputs to the targets. That's great, and it has a ton a transformative practical applications. But it's still the only thing we can do really well. Let's not mistake this fairly narrow success in supervised learning for having "solved" machine perception, or machine intelligence in general. The things about intelligence that we don't understand still massively outnumber the things that we do understand, and while we are standing one step closer to general AI than we did ten years ago, it's only by a small increment.