Quick and accurate knowledge adaptation in machine learning
Abstract
Neural networks excel when they have access to all data at once, requiring multiple passes through the data during training. However, standard deep-learning techniques are unable to continually adapt as the environment changes: either they forget old data, or they fail to sufficiently adapt to new data. This inability to adapt is a problem in many real-world applications. One example is called continual learning (or lifelong learning), where data examples arrive sequentially over time, and our machine learning model must adapt to the new data while not forgetting old data. Another example is in machine unlearning, where our model must ‘forget’ some of its training data, such as due to data deletion requests. We could retrain our model from scratch on all of our data every time new data (or a data deletion request) comes in, but can we be quicker and more efficient than this? In this talk, we show how such knowledge-adaptation tasks are closely related, and tackle them using a Bayesian approach. Our Knowledge-adaptation priors (K-priors) combine weight and function- space priors to accurately reconstruct the gradients of past data. This recovers and generalises many existing, but seemingly-unrelated, adaptation strategies. Empirical results show that adaptation with K- priors achieves performance similar to full retraining, but only requires training on a subset of past examples.
Bio:
Siddharth Swaroop is a postdoctoral fellow at the Data to Actionable Knowledge Lab at Harvard University, working with Professor Finale Doshi-Velez. His current research focusses on probabilistic deep learning for a broad range of knowledge adaptation problems (such as continual learning, federated learning, unlearning and knowledge distillation), as well as reinforcement learning problems that arise in human-AI settings. He has also worked on machine learning interpretability and privacy, as well as on applications of large language models to automated knowledge base construction at Microsoft Research UK. Siddharth obtained his PhD in machine learning at the University of Cambridge, supervised by Professor Richard Turner, where he was awarded a Microsoft Research EMEA PhD Award and an Honorary Vice-Chancellor’s Award from the Cambridge Trust.