I met with a data scientist so that I might get a better understanding of what it actually means to teach code to learn. We sat in a cabin without internet, surrounded by stands of pine, eating french toast from a cast iron skillet: in short the least likely place to start a conversation with “does creating intelligence make you a god?” I did not ask this; at times I can possess shockingly normal levels of chill. Instead, I asked him to explain in digestible terms why he needs artificial intelligence. The idea behind learning programs in his line of work is simple: the amount of data and the ways in which it can be manipulated are sometimes so ugly that it’s easier to let a program learn from experience with that data, and to accept a certain margin of expected error, than to lose the immense time for a human to do the same thing perfectly.
To illustrate better for me, he drew this example. The Post Office, as you might imagine, has access to an absurd amount of handwriting, a truly astronomical selection. The Post Office thinks, “if we can get a computer to read numbers, we can sort mail by zip code a thousand times faster.” So they give a programmer tens of thousands of handwritten digits, which the programmer splits into two halves: training and testing. Through the first part, a programmer says “this is what a seven looks like, this is how this arrangement of pixels translates to you as values,” since the computer doesn’t have a real visual component. After a while, the programmer might ask it to identify on its own. It can learn from being told its mistakes, over and over and over until the program has absorbed enough to test itself on the second set of data and see how well it did. It’s the teach a man to fish principle: by teaching the program to discover number imaging on its own rather than just giving it parameters, the brain has a better chance of accurate classification as a result of its catalog of successes and failures. In the beginning it looks like a dad setting up the fishing line for the child to cast, as the coder holds the code’s hand through one small, specific part of the task, but that brain is stronger and more capable having learned by doing and can build upon that task. These are the kinds of little problems scientists are giving to learning computers to solve. Mundane tasks and games with a learning element are very popular among scientists now, because they are extremely helpful and time-saving to humans but also ever-so-slowly nudge forward the study and efficiency of learning on the whole.
I address envelopes more meaningfully with this in mind, and regret that phase in grammar school when I put different smiley faces in all my zeros. I wonder if there is a line of frustrated code for deciphering them, one of millions, just for me. I formally apologize, technology. I couldn’t have known.