Towards a Unified Framework for Split Learning

Split Learning (SL) is a principled approach for training models on data distributed across multiple devices without sharing training data. While SL emerged as an alternative to federated learning to reduce the compute burden on devices, it also enables a more fair redistribution of work between edge devices and the server. Despite its potential, there is no unified framework for implementing and deploying SL algorithms, leaving several research questions underexplored. To address this gap, we introduce SplitBud, a versatile framework to implement virtually any SL algorithm. By supporting various variants of SL, SplitBud facilitates research and development in the field. In this paper, we demonstrate its flexibility by implementing and evaluating multiple SL algorithms, and we discuss future directions for the field.