Invited Speakers

​

Speaker: R Venkatesh Babu, Indian Institute of Science, Bangalore

Title: Tackling Bias in Deep Models and Enhancing Fine-Grained Control in Generative Models

 

Abstract: Deep models are biased since they were trained on datasets that are biased. Debiasing becomes much more challenging when we use features from pretrained black-box models for downstream applications. To address these issues, in our recent work, we propose a simple method to address the bias issue of a blackbox feature encoder, with no knowledge of the bias attribute.  The talk will further discuss our team’s efforts to investigate, reduce bias, and improve the precision of text-to-image diffusion models. Our first study examines whether these models accurately represent global environments or are biased toward specific regions. A crowdsourced study revealed that they often default to environments from the United States, Canada, and India, underrepresenting other regions. This underscores the need for greater geographical inclusivity in future models. The second part of our work focuses on addressing biases in face generation, where models tend to favor certain demographic groups. To combat this, we introduce a new technique called Distribution Guidance, which works within the model’s semantic space to ensure generated images reflect a balanced attribute distribution—without needing extra data or retraining. Lastly, we explore fine-grained image editing using text-to-image models, that offers precise attribute control for faces and learn human affordances for a given background environment. 

 

 

Biography: R. Venkatesh Babu is a professor and Chair of the Department of Computational and Data Sciences (CDS), Indian Institute of Science (IISc), Bangalore. He received his doctoral degree from the Dept. of Electrical Engineering, Indian Institute of Science, Bangalore. He held postdoctoral positions at NTNU, Norway, IRISA/INRIA, Rennes, France, and NTU, Singapore. He is the head of the 'Vision and AI Lab' (VAL) at IISc. His research interests include Computer Vision and Machine Learning. He is a recipient of the SERB Star (2020) and Sathish Dhawan Young Engineer awards (2019). He served as a Program Chair of the AIML-Systems 2023, NCVPRIPG’19 conferences and General Chair of ICVGIP'24 and SPCOM 2020. He is an associate editor of IEEE TPAMI, IEEE TIP, PR and CVIU journals. Prof. Venkatesh serves as an area chair for several top conferences including CVPR, AAAI, NeurIPS, ICLR, ICCV, ECCV, ACCV, WACV, ACML, and AISTATS. He is a senior member of IEEE and AAAI.