Research

Most of my work was about doing machine learning when labels are scarce. The best-known project used satellite imagery to estimate poverty in Africa, where reliable survey data is rare and expensive to collect. Nighttime lights are a rough proxy for wealth, so we trained a convolutional neural network to predict nighttime lights from daytime images, then used what it learned to estimate household consumption and assets in Nigeria, Tanzania, Uganda, Malawi, and Rwanda. The model explained up to 75% of the variation in local economic outcomes, using only public data.

That project was published in Science in 2016, with code and data on GitHub. Later work extended the same idea: learning useful representations of places without labels (Tile2Vec), and semi-supervised regression that makes use of unlabeled data (semi-supervised deep kernel learning).