Jagmohan Chauhan
Associate Professor, UCL
Bio
Jagmohan Chauhan is an Associate Professor at UCL, where he leads the IntellEcT Systems group. His research focuses on efficient, adaptive, and trustworthy machine learning systems for Generative AI, Agentic AI, world models, and reasoning, with applications in robotics and healthcare. He has secured around 10.7 million in research funding from EPSRC and industry, and is a member of the UCL AI Centre, UCLIC, and the UCL Generative AI Hub. Previously, he was a Lecturer at the University of Southampton (2021–2025) and a postdoctoral researcher in the Mobile Systems Group at the University of Cambridge. He has also worked with researchers at Samsung AI Center, Huawei, and Nokia Bell Labs. He received his PhD from UNSW, Australia, and Data61 (CSIRO), and his MSc in Computer Science from the University of Saskatchewan, Canada.
Efficient Video-Language Understanding: From Scalable Inference to Intelligent Search
Abstract
Understanding long videos efficiently remains a fundamental challenge for video-language models (VLMs): processing visual information at scale is computationally expensive, while finding the right evidence across extended temporal sequences remains difficult. This talk explores two complementary approaches to making video-language understanding more efficient, scalable, and intelligent. First, SURGE tackles inference inefficiency by identifying and retaining informative visual tokens, reducing computational overhead while preserving task-relevant information. Second, BeliefSearch rethinks video navigation, treating VLMs as noisy sensors and using Bayesian beliefs to guide where to look, what to observe, and when to stop. Together, these works point towards a shift from processing more information to processing the right information. The talk will conclude with emerging challenges and opportunities in long-video understanding.