Causal Circuit Audits for Multimodal Medical Foundation Models

Create open-source methods to find and remove the exact internal cues that make image+text medical models rely on shortcuts or memorized data, improving robustness and privacy without retraining the whole model.

Biography

George Kaissis focuses on foundational research in machine learning, including differential privacy, fairness, interpretability, model editing and efficiency, to the creation of large-scale, human-in-the-loop systems based on multimodal large reasoning models designed to empower clinicians, researchers and patients.