Masters student Adam Gronowski received the Canadian Society for Information Theory (CSIT) Paper Award for the best paper accepted at the Seventeenth IEEE Canadian Workshop on Information Theory, Ottawa, June 5-8, 2022.
The paper is entitled R茅nyi Fair Information Bottleneck for Image Classification by A. Gronowski, W. Paul (APL, Johns Hopkins Univ.), F. Alajaji, B. Gharesifard and P. Burlina (APL, Johns Hopkins Univ.). The paper consists of a certificate and a $1000 prize.
In spite of the vast benefits of machine learning algorithms used in a variety of high stake applications, ranging from loan approvals to disease diagnosis, the use of automated algorithms that are not designed to address bias can lead to harm and exacerbating social inequities. The problem of developing algorithms that are both accurate and fair, i.e., not discriminating against individuals because of their gender, race, age, or other sensitive attributes, are hence of paramount importance.
This work develops a novel method for ensuring fairness and examines its applications in image classification problems. The mathematical idea is to leverage a variational information bottleneck approach that incorporates constraints for utility, fairness, and compactness of representation, taking into account the mutual information between the data and its representation using the so-called R茅nyi divergence. The work captures a number trade-offs between utility and fairness, providing a better understanding of the complexities involved in developing fair production-grade AI systems.
The award鈥檚 reference terms can be found .