On Sep 10, 2026, the 92nd Zhi Xin Forum was held in Lecture Hall 117 of the Zhi Xin Building. Dr. Oya Celiktutan from King’s College London, UK, was invited to deliver a talk entitled “Fairness for Human-centred Deep Learning Models”.
In this talk, Dr. Celiktutan focused on fairness in human-centred deep learning models. She first introduced a feature-level framework for evaluating demographic biases in facial expression recognition models, and then systematically discussed model biases across different demographic groups and their statistical assessment. In response to group bias in vision-language models, Dr. Celiktutan proposed an LLM-guided method for group prototype construction and feature-space debiasing, which reduces model bias while preserving model utility. At the application and evaluation level, the talk demonstrated the application of these methods in zero-shot image classification, text-to-image retrieval, and text-to-image generation, and further discussed the trade-off between model utility and fairness.
After the talk, Dr.Celiktutanengaged in a lively exchange and discussion with the attending faculty and students on topics such as how to improve model fairness. Drawing on her own research experience, she encouraged students to actively broaden their horizons, explore diligently, and discover and solve new scientific problems. This talk significantly broadened the perspectives of our university’s faculty and students and deepened their understanding and knowledge of fairness in artificial intelligence.
