Search papers, labs, and topics across Lattice.
This paper introduces an adaptive channel pruning scheme for split learning (ACP-SL) to reduce communication overhead by compressing intermediate feature representations (smashed data). A key component is the label-aware channel importance scoring (LCIS) module, which distinguishes important channels from less important ones based on label information. The adaptive channel pruning (ACP) module then prunes less important channels based on these scores. Experiments demonstrate that ACP-SL achieves higher test accuracy and reaches target accuracy faster than benchmark schemes, thus reducing communication overhead.
Achieve higher accuracy and faster convergence in split learning by intelligently pruning communication channels based on label awareness.
Split learning (SL) transfers most of the training workload to the server, which alleviates computational burden on client devices. However, the transmission of intermediate feature representations, referred to as smashed data, incurs significant communication overhead, particularly when a large number of client devices are involved. To address this challenge, we propose an adaptive channel pruning-aided SL (ACP-SL) scheme. In ACP-SL, a label-aware channel importance scoring (LCIS) module is designed to generate channel importance scores, distinguishing important channels from less important ones. Based on these scores, an adaptive channel pruning (ACP) module is developed to prune less important channels, thereby compressing the corresponding smashed data and reducing the communication overhead. Experimental results show that ACP-SL consistently outperforms benchmark schemes in test accuracy. Furthermore, it reaches a target test accuracy in fewer training rounds, thereby reducing communication overhead.
