Klemens Flöge
All projects

Meta-Learning within the PAC-Bayesian Framework

University of Cambridge·thesis

Meta-learning on the sinusoids example, meta-learning a prior over Bayesian neural networks.
Meta-learning on the sinusoids example, meta-learning a prior over Bayesian neural networks.

Meta-learning research is overwhelmingly empirical: algorithms that work, with little said about why or under what conditions. This thesis approaches the field through a PAC-Bayesian lens, which turns “it works on these benchmarks” into bounds you can state.

The centrepiece is the PACOH algorithm, treated as a class of meta-learners with probabilistic performance guarantees. From there the essay reads optimisation-based meta-learners — MAML and REPTILE among them — through the same framework, showing how existing algorithms can be contextualised and theoretically supported rather than merely benchmarked. Numerical experiments accompany the theory throughout.