UNICEF and Tryolabs have partnered to understand how machine learning models perform by using explainable artificial intelligence (XAI) methods, improving these to better detect schools in satellite imagery to support foundational work to connect every school in the world.
UNICEF has been expending great effort to map schools in developing countries. Part of the United Nations (UN), UNICEF provides humanitarian and developmental aid to children worldwide. One of their goals is to facilitate children’s access to education. By combining high-resolution satellite imagery, UNICEF’s Office of Innovation created a model capable of recognizing schools. Distinctive features of schools as viewed from above allowed the model to distinguish them from other buildings.
However, UNICEF wanted to ensure that they avoid biases in their model as much as possible. Hence, UNICEF and Tryolabs partnered together to understand how UNICEF’s Machine Learning model was actually performing by employing explainable artificial intelligence (XAI) methods. Leveraging AI for social good was the essence of the project and our main goal.
Training a machine learning model to identify schools from satellite imagery of developing countries is a difficult task. Our collaboration brought together their experience and our AI skills to gain new insights into their model’s predictive behavior. Tryolabs believes this kind of partnership promotes innovative and unique results.
By employing Explainable AI methods on innovative ways we could understand how our AI model was working. We found hidden biases in the dataset that were translated to inequalities during the final AI model usage. Re-train an AI model with a new dataset based on landscape features will improve school detections on satellite imagery and should be the next step.
We learned that this AI4SG project required an open approach of open data, open source code and an open mind to consider different fields of knowledge.
By joining efforts, Giga and Tryolabs can better serve the purpose of providing internet to every school in the world.