Elijah Cole: Learning from real-world data

Humans are collecting vast amounts of data to tackle societal challenges like preserving biodiversity, improving healthcare outcomes, and accelerating scientific discovery. However, in many domains we are collecting far too much data to analyze manually. To meet these challenges, we need machine learning algorithms that can automatically extract knowledge from this growing influx of raw data. Unfortunately, many machine learning algorithms are not compatible with the limited, noisy, and weak supervision often found in real-world settings.

This talk focuses on the task of learning from real-world data. I will present progress on the problem of learning from minimal labels in contexts like multi-label image classification and self-supervised representation learning. I will also describe how we can leverage small amounts of domain context to improve machine learning algorithms in a modular and flexible way, with examples from image classification and object localization. Finally, I will show how studying real-world data opens new frontiers in machine learning research and provides greater opportunities for impact in important application domains.

Elijah Cole (https://elijahcole.me/) is a Ph.D. candidate in the Computing and Mathematical Sciences department at Caltech, advised by Pietro Perona. He received a B.S.E. from Duke University in 2017 with a double major in electrical engineering and mathematics. During his Ph.D. he completed internships at Google Research, Microsoft AI for Earth, and the Air Force Research Laboratory. His research focuses on deep learning and computer vision, with an emphasis on learning from limited, noisy, and weak supervision. He works with ecologists, physicians, and other domain experts to develop new benchmarks that test algorithms under realistic conditions and challenge traditional machine learning paradigms. Elijah’s work is funded by an NSF Graduate Research Fellowship and the Resnick Sustainability Institute.