Gaining Experience Using Deep Learning for Social Good in Global Competition
May 06, 2020
AI has demonstrated its versatility in recent years, extending far beyond enterprise use into the realm of social good.
AI has demonstrated its versatility in recent years, extending far beyond enterprise use into the realm of social good. Today, machine learning and deep learning are used to solve some of the most pressing dilemmas facing humanity, ranging from diagnosing and treating disease to getting smarter about energy consumption.
To help move this social progress along, organizations such as DrivenData host online challenges, combining the latest technologies in data science and crowdsourcing to help solve large-scale social problems. In partnership with companies like MathWorks, DrivenData regularly runs data science competitions, like its OpenAI Caribbean Challenge, aimed at solving critical problems in areas like healthcare, education, and public services by challenging engineers, scientists, and students to develop the best algorithms for social good.
The OpenAI Challenge tasked competitors with developing AI models that can identify areas in the Caribbean most vulnerable to natural disasters. The AI models needed to analyze aerial drone imagery data and identify high-risk areas based on roofing materials…a critical determining factor for the degree of damage a structure might suffer during an earthquake or flood.
The competitors’ algorithms can help scale disaster resilience efforts by prioritizing areas or specific buildings for greater protection in advance of disasters. This approach will save tremendous time and risk management resources, both of which are in short supply when natural disasters hit. “Previously, identifying at-risk areas and buildings required personnel to go door to door by foot,” said cofounder and principal at DrivenData Greg Lipstein. “This approach was taxing from both a budgetary and time perspective, especially with citizens’ safety on the line.”
In addition to benefitting the Caribbean people, DrivenData’s challenge also gave engineers, scientists, and students the opportunity to work with a unique, real-world dataset and develop AI models with life-saving potential. The aerial drone imagery data for the challenge was provided and prepared by WeRobotics, a non-governmental organization dedicated to bringing robotics technology to developing countries, and the World Bank Global Program for Resilient Housing. “The competitors gain invaluable experience working with real-world data, and the competition platform gives us the ability to engage and connect thousands of individuals around the world who are passionate and motivated to use data science for social good,” Lipstein said.
Ning Xuan, a biomedical engineer with a master’s degree from Columbia University, entered the competition to enhance his deep learning skills. The competition gave Xuan the opportunity to work with aerial imagery and extract geolocation coordinates for the first time. A key learning for him was understanding that geolocation coordinates are mapped on a 3D surface to simulate the Earth’s geography, something he did not know before the competition.
“It gave me hands-on experience to work on a project with real-life data I had not worked with before,” Xuan said. “The competition gave me a broader perspective on how data science can be applied to multiple fields of work.” He also valued the discussion panel accompanying the competition to both teach and learn from others in the data science community.
As part of its sponsorship, MathWorks provided competitors, like Xuan, access to MATLAB for the competition. Xuan used MATLAB and Deep Learning Toolbox to process geographic and aerial imagery data of buildings in St. Lucia, Guatemala, and Colombia; classify images of roofing materials; and develop and train the deep learning model. Xuan’s model correctly identified the roof construction material with over 80% accuracy, an impressive result given five different roof categories to choose from. This was more than double the accuracy rate of the pre-competition benchmark model. His successful model was recognized by DrivenData, earning Xuan the top MATLAB user award.
Companies, organizations, and governments have always looked to use technology to solve the world’s problems. As engineers and scientists are increasingly called upon to apply their expertise to deliver social good, they will look for opportunities, like DrivenData’s competitions, to gain AI skills and experience working with real-world data.