Incorporating Fairness criteria in Computer Vision and Neuroimaging by Vikas Singh
Title: Incorporating Fairness criteria in Computer Vision and Neuroimaging
Abstract: AI algorithms underlie a broad range of modern systems that we depend on every day. While such algorithmic decision making continues to integrate closely with our lives, a number of high profile news stories have shown troubling blind spots, including discriminatory behavior in recommendations made by blackbox automated systems as well as biases against skin color in surveillance systems. The study of of mechanisms that provide guardrails against such behavior has grown from a niche pedagogical topic about a decade ago to informing procedures being frequently adopted when deploying machine learning systems. This talk will briefly review how the relevance of this topic evolved in the context of computer vision, and describe various strategies that are being developed here and elsewhere that offer practical solutions in a limited number of cases in deep learning, as well as offer benefits that go beyond fairness. We will then segue to results showing how these algorithms and ideas can be repurposed for analyzing multi-site neuroimaging (and other types of) datasets when commonly used deep neural network architectures are being utilized, for both supervised and unsupervised learning.
Based on joint work with Vishnu Lokhande, Aditya Akash, Sathya Ravi, Hao Zhou and Grace Wahba.