Understanding Gender and Racial Biases in Computer Vision: Model, Data, and Human Biases

When:
Friday, May 22, 2020, 12:30 pm - 1:30 pm PDTiCal
Where:
VTC Only, see link below
This event is open to the public.
Type:
AI Seminar
Speaker:
Jungseock Joo (UCLA)
Video Recording:
https://usc.zoom.us/rec/share/3cxVEpfu2E9IZtbD5RqHBfMaIJj8T6a81HUervFey0miIcsPfsn4GwCY60OLOJ4_
Description:

 

Abstract: 
Automated computer vision systems have been applied in many domains including security, law enforcement, and personal devices, but recent reports indicate that these systems may produce biased results, discriminating against people in certain demographic groups. Identification and diagnosis of such model bias, however, are challenging tasks because modern computer vision systems rely on complex, black-box models whose behaviors are hard to decode. In this talk, I will first introduce our new dataset, FairFace, which can be used to measure the biases of computer vision models for face attribute classification. The dataset contains 108,501 images which is balanced on race in contrast to existing public face image datasets, which are dominated by White faces. I will also discuss our new framework which allows us to measure counterfactual fairness in computer vision models. By using a generative model for face attribute manipulation, our method can synthesize new facial images varying in the dimensions of gender and race, while keeping other information intact. Such images can then be used to measure the sensitivity of a computer vision model to gender or race related cues. Using this new dataset and method, I will demonstrate the biases of several public datasets and commercial services commonly used by researchers. 

Bio:
Jungseock Joo is an assistant professor in Communication at University of California, Los Angeles. His research primarily focuses on understanding multimodal human communication with computer vision and machine learning. His research employs various types of large scale multimodal media data such as TV news or online social media and examines how multimodal cues in these domains relate to public opinions and real world events. His research has been supported by National Science Foundation, Hellman Foundation, Samsung, and UCLA Faculty Career Development Award. He received Ph.D. in Computer Science from UCLA, M.S. in Computer Science from Columbia University, and B.S. in Computer Science and Engineering from Seoul National University. He was a research scientist at Facebook prior to joining UCLA in 2016.

« Return to Upcoming Events