As demonstrated from the polling failures from our last elections, human-driven statistical analysis is prone to bias, overconfidence, and errors. Perhaps computer algorithms can do a better job of accurately predicting political bias?
One unique approach, from renowned computer vision expert Fei Fei Li, applies deep learning to Google Street View data and identifies the make, model, and year of every motor vehicle encountered. In a recently published paper, Li and her team analyzed 50 million images in 200 American cities to label 22 million automobiles which constitute 8% of the total number of vehicles in the country.
Google Street View Can Reveal A Lot About You
Turns out the cars in your neighborhood are a reliable predictor for your socioeconomic status and political leanings. According to Li, et al, “if the number of sedans encountered during a 15-minute drive through a city is higher than the number of pickup trucks, the city is likely to vote for a Democrat during the next Presidential election (88% chance); otherwise, it is likely to vote Republican (82%).”
The algorithms also use the vehicle data to predict demographics. A neighborhood with a strong presence of Hondas and Toyotas typically indicates an Asian population, since Asian drivers often purchase Asian cars. African American neighborhoods disproportionately feature Chrysler, Buick, and Oldsmobile vehicles, and while pickup trucks, Volkswagens, and Aston Martins are indicative of mostly Caucasian presence. These demographic results discovered independently by the algorithm confirms previous independent studies of vehicle preferences.
The U.S. Census Bureau spent an estimated $14.7 billion to conduct the 2010 Census. The labor-intensive collections process is not only expensive, but also causes the data to be inconsistently dated across locales and lag behind reality by up to half a decade.
Automated computation methods driven by deep learning can be used in combination with manual, on-the-ground methods to enhance accuracy, improve speed, and dramatically bring down costs. Beyond just assessing political leanings, such computer vision based approaches can be used for the greater good.
Automated Analysis Can Inform Policy & Governance
Deep learning neural network approaches have been successfully applied to socioeconomic predictions before. Stefano Ermon is a computer science professor at Stanford University committed to addressing global issues like poverty with an approach called “computational sustainability.”
By applying deep learning to satellite imagery, Ermon and his interdisciplinary team at the Sustainability & AI Lab at Stanford can predict poverty in regions where surveys are inaccurate or even impossible. Surveyors can’t be sent safely into violence-fueled regions like Somalia without exposing them to kidnapping risk and other dangers.
Even with low-resolution and publicly available satellite imagery, the algorithms can identify features such as whether a roof is made of grass, thatch, or metal, or if a neighborhood has a swimming pool. These features can be used to compute estimates of economic development and pinpoint regions in need of greatest aid. While daytime satellite images can identify volume of agricultural production and proximity of water sources, nighttime images can be used to map out electrical infrastructure which is a strong predictor of the divide between rich and poor.
Just as Li’s automated methods could complement the work done by the U.S. Census Bureau, Ermon’s work with satellite data has been shown to be comparably accurate to manual approaches. In crop yield analysis, his team’s remote sensing approaches driven by deep learning can outperform USDA’s crop predictions which are garnered from sending surveyors into the fields.
Given the right data and insights, policymakers both in the United States and abroad can improve decision-making, resource-allocation, and ultimately quality of life for their constituents. If deep learning approaches are used wisely and for the better good, their potential positive impact to society could be transformative.