Abstract: Focal loss has emerged as a de facto training loss in class-imbalanced classification problems, especially in computer vision. Despite its empirical success, its theoretical properties and benefits have not been well explored. In this talk, we investigate several information-theoretic aspects of focal loss, including its connection to relative entropy and the emergence of novel information measures it induces. These results, which are also experimentally validated, provide a theoretical foundation for understanding focal loss and help clarify the trade-offs it introduces when applied to imbalanced learning tasks. We also propose the focal loss as a distortion measure for lossy source coding and discuss some research directions that are the object of current investigation.

Bio: Martina Cardone received her Ph.D. degree in electronics and communications from Télécom ParisTech (with work done at Eurecom in Sophia Antipolis, France) in 2015. She is currently an Associate Professor with the Electrical and Computer Engineering Department, University of Minnesota (UMN). From July 2015 to August 2017, she was a Postdoctoral Research Fellow with the Electrical and Computer Engineering Department, UCLA Henry Samueli School. Her main research interests are in estimation theory, network information theory, and wireless networks. She is a recipient of the 2022 McKnight Land-Grant Professorship, the NSF CAREER Award in 2021, the NSF CRII Award in 2019, the Outstanding Ph.D. Award from Télécom ParisTech (Paris, France), and the Qualcomm Innovation Fellowship in 2014.