Reliable Saliency Explanations for Computer Vision and Medical Imaging -- Xingjian Li

​Saliency maps are widely used to explain deep vision models by highlighting image regions that influence model predictions. However, existing explanation methods can be unreliable when they depend too heavily on decision boundaries, gradients, out-of-distribution perturbations, or unstable local sampling. In this talk, we will discuss two recent works, DiffCAM and MedLIME, that aim to make saliency explanations more reliable and flexible.
DiffCAM introduces a data-driven perspective for visual explanations by identifying how a target image differs from representative reference examples, rather than simply following the fastest direction to change a model’s decision. This enables more flexible explanations, including ‘why is’ and ‘why not’ questions. MedLIME further extends reliable saliency explanations to medical imaging based on LIME. Instead of being faithful to the model, it aims to produce more robust abnormality localization maps under model-agnostic settings.
This session will be particularly interesting for those working in explainable AI, medical imaging, trustworthy computer vision and reliable deployment of AI systems in high-stakes domains. Join us for an insightful discussion!

Speaker Bio: Xingjian Li is a Postdoctoral Researcher in the Computational Biology Department at Carnegie Mellon University, working with Prof. Min Xu. His research focuses on reliable and trustworthy AI for computer vision and high-stakes applications such as biomedical imaging.

//

Join our sessions to catch up on all things  ML Healthcare research, led by Leema Krishna, Anas Zafar and Oumayma Essarhi. Learn more and check out recent events from this group on their minisite.

​For the full experience, connect with the community on Discord. Learn more about the Cohere Labs ML research open science community, and apply to join.