Krish Kabra, Ph.D. student in electrical and computer engineering (ECE), and Guha Balakrishnan, assistant professor of ECE, recently won the Best Paper Award at the Responsible Generative AI (ReGenAI) Workshop in Seattle, Washington.
The ReGenAI workshop brought together researchers, practitioners, and industry leaders to discuss concerns around data, ethics, privacy and regulation in generative AI. One concern is bias in data used to train computer vision models, which could lead to unfair and potentially harmful outcomes.
Kabra and Balakrishnan address this concern in their paper titled, "GELDA: A generative language annotation framework to reveal visual biases in image generators,” which was co-authored by Kathleen M. Lewis from MIT.
In this paper, they propose GELDA (GEnerative Language-based Dataset Annotation), a nearly automatic framework that leverages large generative language models, as a complement to human analysis of vision datasets.
“By ensuring datasets are diverse, representative, and free from biases, we can develop AI systems that avoid reinforcing harmful stereotypes,” said Kabra. “This is what GELDA helps accomplish—it is a tool that helps developers discover biases in image datasets in an automatic and flexible manner.”
The conversations on responsible generative AI at ReGenAI happened as the global computer vision industry is forecasted to grow by more than 80% over the next six years.
“With the rapid development of generative AI technologies, there is an even greater need to design tools to ensure their safe and responsible deployment,” said Kabra. “We thank the organizers for recognizing our work as an important contribution to the field of responsible generative AI.”
They were formally presented with the award at the conclusion of the workshop on June 18, 2024, at the Seattle Convention Center.