ELLIS Delft Talk by Merve Gürel: How to make your AI survive Craigslist: Enhancing the Robustness of Machine Learning with Knowledge and Reasoning

Abstract

Despite the remarkable success achieved by deep learning models, recent studies show that they are vulnerable to adversarial attacks or minor perturbations at the inference time. This talk dives into my research on building AI systems we can trust, through "certifiably robust" machine learning pipelines that blend data-driven learning with knowledge and reasoning. In the first half of the talk, I will introduce our Learning-Reasoning pipeline, which leverages semantic information extracted from data (learning) and encodes them through probabilistic models (reasoning). I will show that the Learning-Reasoning pipeline offers not only a certifiably robust prediction accuracy, but also empirical robustness against real-world attacks and perturbations. In the second half, I will particularly focus on Large Language Models (LLMs) and present our certified Retrieval Augmented Language Models framework, the first certified framework of its kind. By utilizing an external knowledge base, our framework can significantly and provably reduce the generation risks of LLMs, even under inference-time distribution shifts. I will demonstrate the soundness and tightness of our certification guarantees across several datasets and retrieval models. Join me at this talk to discuss the exciting possibilities of knowledge-enhanced data-driven learning. Where do we go from here?