From California Magazine
How deep-learning algorithms can help governments detect remote clusters of famine and crisis.
By Katherine Blesie
RESEARCHERS AT BERKELEY’S DATA-INTENSIVE development lab are using big data to deliver aid to the world’s chronically hungry—a group that has doubled in size from 135 million to more than a quarter billion during the pandemic.
Lab director and associate professor Joshua Blumenstock has spent much of the past six months developing an algorithm to identify at-risk regions in the West African nation of Togo, where almost 60 percent of the population lives in hard-to-reach rural areas and more than 90 percent has no work beyond informal employment. The Togolese government is now using the deep-learning algorithm, trained to identify patterns in roofing quality, road conditions, farm plot size, etc., to scan satellite data for clusters of extreme poverty. Once identified, direct payments are sent to people in those regions in the hope of fending off famine and other humanitarian crises.
But the Togolese government doesn’t just want to reach the poorest regions, they want to reach the poorest individuals in those regions. To make that happen, Blumenstock and his team developed a separate technology that uses machine learning to identify patterns in phone behavior that are indicative of poverty, such as infrequent phone credit top-ups and brief calls.
Joshua Blumenstock is an associate professor in the UC Berkeley School of Information.