Adapting knowledge in machine learning models

About the Project

This project will develop machine learning algorithms to cheaply and efficiently adapt knowledge in machine learning / AI models. This includes scenarios such as: sequentially updating models with new data without forgetting previous knowledge (continual learning); training a model on data split over many locations such as across hospitals (federated learning); and removing undesired data or concepts from a model (unlearning). The focus will be on developing algorithms that can be applied across all such knowledge adaptation tasks. These algorithms will be designed using probabilistic methods, such as Bayesian-inspired methods to incrementally update models: these provide a sound theoretical framework to unify knowledge adaptation tasks. Overall, the goal of this research is to develop cheaper, more robust, and theoretically-sound algorithms across knowledge adaptation tasks. There will also be opportunities to work at the intersection of knowledge adaptation and (i) AI policy, and/or (ii) human-AI interaction.

Essential: degree in Mathematics, Computer Science, or similar; course(s) in machine learning; coding skills (such as Python).

Desired: course on probabilistic machine learning or similar; experience working with neural networks or foundation models. Note that these skills can be replaced by the willingness and capacity to learn them.

Please contact the advisor (Dr Siddharth Swaroop) directly with questions or if interested. Applications from underrepresented groups in computer science are strongly encouraged.