“We wanted to show that useful insight can be gained from images, the same way people do this for social networks or other textual-based data,” Timnit Gebru, one of the lead researchers on the paper, told Digital Trends. “Some of the car-politics or car-race associations were intuitive, but still surprising that we could capture from our data.”
The deep learning neural network was trained on a dataset of more than 50 million Google Street View images from a variety of cities. This data was then compared to ground census data to help the algorithm make the right connections between race, education, income and voter preferences, and the make, model and year of every car produced since 1990. The artificial intelligence uncovered a number of intriguing tidbits — such as the fact that if the number of sedans in a neighborhood is greater than the number of pickups, there is an 88 percent chance the precinct votes Democrat. More pickups than sedans on your street? That means there’s an 82 percent chance you’re in Republican territory.
While Google is unlikely to add “likely voter demographic” as a tool on Street View anytime soon, the research demonstrates how impressive modern A.I. is — not just in identifying objects, but also at drawing actionable conclusions from this information. As Gebru points out, similar research could be used for exploring things like the links between neighborhoods and health or pollution levels.
A paper describing the work, “Using deep learning and Google Street View to estimate the demographic makeup of neighborhoods across the United States,” was recently published in the journal Proceedings of the National Academy of Sciences.
