Shortly after the end of World War II, the British math prodigy Freeman Dyson and the Indian physics phenom Harish-Chandra, both at Cambridge, had a conversation about their plans to change fields. “I am leaving physics for mathematics; I find physics messy, unrigorous, elusive,” Harish-Chandra told his colleague. Dyson supposedly replied: “I am leaving mathematics for physics for exactly the same reasons.”
After a semester in Stanford’s online machine learning class, I find the discipline messy and elusive, if not easy, when I expected it to be more elegant and theoretical. I wrote earlier that the math gets tough at some point in any sort science, but that doesn’t mean the end result, if you survive the equations, has to lose its elegance as well. In the final lectures of this class, which ended last week, we saw examples of how machines can learn to decipher text and faces in photos with surprising accuracy. (Imagine being blind and wearing a hat that narrated your surroundings.) But the process is rife with error correction and heavily dependent on the subjective decisions that the researchers make when deciding how to take these complex problems piece by piece.
I’m glad that very smart researchers are working on these problems, and I’m relieved I’m not one of them. Some people’s imaginations light up at these sorts of problems, in which the task is to extract meaning from a noisy, disorganized bunch of data. Throughout the course, professor Andrew Ng would mention work he and colleagues had done teaming up with researchers from other disciplines, from tumor detection to self-driving cars. This is not a science that should be limited to improved Google results and smarter phone applications. The hope, among machine-learning enthusiasts, is that more graduate students devote themselves to the field and then go forth into the world to mingle with the many researchers who are in need of smart, adaptive learning algorithms to make sense of their data.
The computing pioneer Charles Babbage supposedly once said that he’d give up the rest of his life to return in 500 years for a three-day tour of the scientific new age. I’m not quite at that level of curiosity about machine learning, but I look forward to checking in periodically over the next decade to see how it has advanced. It’s quite possible that, like computing itself, it will have become a staple in a wide variety of industries. We can already safely predict that it will be important to medicine, for example. Less obvious and more tantalizing is whether it will become useful to the applied social sciences.
More than 10 years ago, Slate ran an article outlining why you can’t build a model to predict presidential elections. The logic there still holds: You can predict any outcome you like depending which data you choose to use. One of the most powerful aspects of the algorithms we studied in this class was their ability to compute tremendous amounts of data and scale it down to the most important factors. Researchers also have ways of preventing their models from “overfitting” the data. In other words, one could probably create an algorithm that predicted past elections with great accuracy, but it might be so finely tuned to the quirks and turns of that historical data that it would be meaningless in 2012. It may seem preposterous to think that any mountainous pile of polls, unemployment rates, approval ratings, the stock markets, and a million other factors could, if correctly parsed by a smart machine, predict human behavior. But by and large, people behave rationally, and machines, if we learned anything, can model rational systems very well. And they’re getting better every day.