Justin Solomon, an Associate Professor in the MIT Department of Electrical Engineering and Computer Science (EECS), is using geometric techniques to solve complex computing problems. Solomon says this method is ideally suited to finding solutions in data science, as it can enable a deeper understanding of the distances, similarities, curvature and shape data.
About half his team at the Geometric Data Processing Group are working on 2D and 3D geometric data issues, such as aligning 3D organ scans in medical imaging and helping autonomous vehicles differentiate pedestrians using spatial data from LiDAR sensors. Meanwhile, the rest of his team use geometric tools to carry out high-dimensional statistical research to develop superior AI models.
Solomon started his career with a fascination with computer graphics which led him to intern at a research lab where he helped develop algorithms for 3D face recognition. Later, at Stanford University, he majored in math and computer science and took part in summer internships at Pixar Animation Studios, where he worked on improving the realism of animated films through physical simulation of cloth and fluids, as well as enhancing rendering techniques.
As a graduate student at Stanford, Solomon focused on optimal transport – finding the most efficient way to move a distribution of some item to another distribution. This work initially focused on computer graphics but later expanded to encompass a wider range of applications.
At MIT, Solomon started the Summer Geometry Initiative, a six-week paid research program focusing on bringing geometric research to underserved students and those from underrepresented backgrounds. He sees a need for a more diverse field of researchers as geometric techniques are increasingly used in machine learning and statistics.
Currently, Solomon is working on improving unsupervised machine learning models with geometry, enabling these models to recognize patterns without having labeled training data. The challenge is clear – a vast majority of 3D data isn’t labeled and having humans label 3D objects can be very costly. However, Solomon believes that by incorporating sophisticated models and geometric wisdom, computers will be able to understand complex, unlabeled 3D scenes, paving the way for more effective learning models.
