How to Train Your Algorithm
Can a well-programmed machine do anything a human can — only better? Complex algorithms are choosing our music, picking our partners, and driving our investments. They can navigate more data than a doctor or lawyer and act with greater precision. For many years we’ve taken solace in the notion that they can’t create. But now that algorithms can learn and adapt, does the future of creativity belong to machines, too? It is hard to imagine a better guide to the bewildering world of artificial intelligence than Marcus du Sautoy, a celebrated Oxford mathematician whose work on symmetry in the ninth dimension has taken him to the vertiginous edge of mathematical understanding. In The Creativity Code: Art and Innovation in the Age of AI he considers what machine learning means for the future of creativity. Here is a brief excerpt from the book looking at how we might train algorithms to better our lives.
You may feel there is something scary about an algorithm deciding what you might like. Could it mean that, if computers conclude you won’t like something, you will never get the chance to see it? Personally, I really enjoy being directed toward new music that I might not have found by myself. I can quickly get stuck in a rut where I put on the same songs over and over. That’s why I’ve always enjoyed the radio. But the algorithms that are now pushing and pulling me through the music library are perfectly suited to finding gems that I’ll like. My worry originally about such algorithms was that they might corral everyone into certain parts of the library, leaving others bereft of listeners. Would they cause a convergence of tastes? But thanks to the nonlinear and chaotic mathematics usually behind them, this doesn’t happen. A small divergence in my likes compared to yours can send us off into different far corners of the library.
I listen to a lot of algorithm-recommended pieces when I am out running. It’s a great time to navigate the new. But I made a big mistake a few weeks ago. My wife asked for my help putting together a playlist for her birthday party. She wanted dancing. She wanted the eighties. So we spent a couple of evenings listening to lots of possibilities. It’s not my choice of music, but we put together a great list of songs that got all our guests up and moving. The problem came when I went out for my first run following the party. My usual choice to let the player surprise me took me deep into the library aisles stocked with eighties dance music. I pressed “skip” as I ran on, but I couldn’t find my way out. It took several weeks of retraining the algorithm on Shostakovich and Messiaen before I got things back on track.
Another context in which we teach algorithms trying to serve us has to do with the spam filters on email applications. A good filter begins by training on a whole swath of emails, some marked as spam, the rest considered legitimate. These are emails that aren’t particular to you yet. By analyzing the words that appear in these emails it starts to build up a profile of spam emails. It learns to treat 100 percent of the emails using the word “Viagra” as spam, along with 99 percent of the emails with the word “refinance.” One hundred percent of the emails with the combination “hot Russian” are spam. The word “diabetes” is more problematic. A lot of spam emails promise cures for diabetes, but it is also a word that crops up legitimately in people’s correspondence. The algorithm simply counts the split in its training data. Perhaps one in twenty messages containing the word turns out not to be spam, so it learns to score an email with “diabetes” as 95 percent likely to be spam.
Your email filter can be set at different levels of filtering. You might specify that only if it’s 95 percent sure should an email go into the junk folder. But now comes the cool bit. While the algorithm initially trained on a generic set of emails, your ongoing actions teach it to recognize the sorts of things you are interested in. Suppose that, in fact, you do suffer from diabetes. At first, all emails with the word “diabetes” will go into your junk folder. But gradually, as you mark emails including the word “diabetes” as legitimate, the algorithm recalibrates the probability of spam to some level below 95 percent and the email arrives in your inbox.
These algorithms are also built to spot other keywords that mark out the junk diabetes emails from the legitimate ones. The inclusion of the word “cure” could well distinguish the duds. Machine learning means that the algorithm will go through every email that comes in, trying to find patterns and links, until it ends up producing an algorithm highly customized to your own individual lifestyle.
This updating of probabilities is also how driverless cars work. It’s really just a more sophisticated version of controlling the paddle in the Atari game Breakout: move the steering wheel right or left according to the pixel data the machine is currently receiving. Does the score go up or down as a consequence?