PhD student Bo Pang recently presented at the International Joint Conference on Artificial Intelligence (IJCAI) in Montreal, showcasing his work using the Our Voices data.
Bo’s paper, CABIN: Debiasing Vision-Language Models Using Backdoor Adjustments, addresses a critical challenge in AI: ensuring that vision-language models (VLMs) do not reproduce harmful stereotypes or unequal performance across demographic groups. This challenge is not only technical but also deeply social, as biased systems may perpetuate inequities in real-world applications.
This work not only advances the technical frontier of debiasing AI but also shows how causal methods can align machine learning with values of equity and inclusion, demonstrating how innovation in AI can actively support more just and inclusive digital systems.
The IJCAI conference was a great experience for Bo. “I had the opportunity to present my paper CABIN: Debiasing Vision-Language Models Using Backdoor Adjustments, and received insightful suggestions on possible future directions. I also had the chance to engage in in-depth discussions with other researchers about challenges such as stereotypical bias in vision-language models, which gave me new perspectives on my work. It was wonderful to experience Montréal by trying some of the local cuisine and enjoying the city’s atmosphere. I also met researchers and peers from all over the world, which made the trip even more enjoyable.”
