An effective method to extract lightweight specialist subnetworks from large foundation models - "Keep what you need: extracting efficient subnetworks from large audio representation models" https://www.arxiv.org/abs/2502.12925