AC²P²SL framework cuts split learning latency with pipeline parallelism
Overlapping communication and computation reduces training time across heterogeneous edge devices.
Split learning (SL) lets base stations train models using distributed data from user equipments (UEs) without exposing raw data. But its sequential computation and communication cause long training times. The new AC²P²SL framework (from Chenyu Liu, Zhaoyang Zhang, et al.) overcomes this by introducing fine-grained pipeline parallelism across multiple micro-batches, effectively overlapping communication and computation to slash overall training latency.
Beyond parallelism, the authors formulate a joint optimization problem that accounts for constraints in communication, computation, and storage, as well as UE heterogeneity. They propose a split and pre-allocation algorithm to boost pipeline efficiency, plus an adaptive re-allocation strategy that responds to dynamic edge environments. Extensive experiments confirm AC²P²SL significantly reduces training time while maintaining data privacy, making it a robust solution for real-world edge AI deployments.
- AC²P²SL introduces fine-grained pipeline parallelism across micro-batches to overlap communication and computation
- Joint optimization considers communication, computation, and storage constraints plus heterogeneous UE capabilities
- Adaptive re-allocation strategy handles dynamic edge environments while preserving local data privacy
Why It Matters
Faster, privacy-preserving distributed training on edge devices enables real-time AI applications in wireless networks.