Google’s new browser experiment lets you learn about basic AI
Just how does machine learning work? You’ve probably read a primer or two on the subject, but often the best way to understand a thing is to try it out for yourself. With that in mind, check out this little in-browser experiment from Google named Teachable Machine. It’s a perfect two-minute summary of what a lot of modern AI can — and more importantly can’t — do.
Teachable Machine lets you use your webcam to train an extremely basic AI program. Just hit the “train green/purple/orange” buttons, and the machine will record whatever it can see through your webcam. Once it’s “learned” enough, it’ll output whatever you like (a GIF or a sound effect or some speech) when it sees the object or activity you trained it with. I taught to recognize my houseplants and respond with relevant GIFs, but others have used it make their hands go moo or play air guitar on command.
All this is pretty fun, but it also demonstrates some fundamental aspects of machine learning. First, that programs like these learn by example. They look, they find patterns, and they remember them. Second, that they need a lot of examples to learn from. And third, and most importantly, that their understanding of the world is superficial and easily broken.
Earlier, for example, I said that I “taught” the machine to recognize my houseplants. The truth is that I only taught it to recognize a vaguely green and fuzzy array of pixels. It doesn’t see my asparagus fern and think (like I do): “Ah, this needs keeping out of sunlight and semi-frequent watering, and I wonder why millennials are drawn to houseplants in the first place? I’ve heard it’s because they can’t afford houses, but also hashtag urban jungle I guess.” All the machine knows is the pixels it can see, and any extra information has to be programmed in.
All this is worth remembering next time you’re reading about machine learning or artificial intelligence. Yes, the field has made huge, huge strides in recent years, but as we’re seeing more and more, the algorithms it’s creating are nowhere near as clever as we’d like them to be. In other words, they’re still learning.