There are many different ways A.I. developers are trying to get intelligent machines to learn and absorb information and experiences — and these usually involve making programs dig through huge dumps of data. But a team of Stanford scientists are looking to a much more conventional form of teaching that humans have relied on since the dawn of the written word: Reading.
In a new study uploaded to the arXiv (pronounced “archive”) paper repository, a research team outlines how it created a program named Augur to access to an insanely large database of online fiction — and it has learned how to accurately predict different kinds of human responses to specific situations — based solely on what it has read.
Augur has basically learned about humans through 600,000 stories currently stored on the online writing community WattPad. It’s read descriptions of human behavior ranging from the mundane, like eating food or taking a selfie, to the much more extreme. Because of this, Augur can identify actions of individual humans in real-world situations and predict what the next step will be, “such as a phone that silences itself when the odds of you answering it are low,” write the researchers.
And it’s easy to see why fiction is such a useful learning tool. “‘While we tend to think about stories in terms of the dramatic and unusual events that shape their plots,” the researchers write in the paper, “stories are also filled with prosaic information about how we navigate and react to our everyday surroundings. Over many millions of words, these mundane patterns are far more common than their dramatic counterparts. Characters in modern fiction turn on the lights after entering rooms; they react to compliments by blushing; they do not answer their phones when they are in meetings.”
In the field tests conducted so far, participants were given a Augur-powered wearable camera to allow the system to identify objects and individuals in a given environment. The system was able to predict the next move with 71 percent accuracy. About 94 percent of those predictions were rated “sensible” — a pretty substantial feat when you remember that’s just a bunch of algorithmic code able trying to predict the future.
Of course, it’s not the first time A.I. researchers have turned to literature to teach machines. Facebook recently made 1.6 gigabytes of children’s story available to the research community with an eye to help A.I. distinguish realistic scenarios from the fantastical.