Collate Framework Optimizes Federated Learning for Edge Latency

Researchers have introduced Collate, a novel training framework designed to enhance federated learning (FL) for real-time, latency-critical edge systems. While FL is effective for collaborative model training and data privacy, device heterogeneity in edge environments often degrades inference performance, especially under strict latency requirements. Existing FL optimizations primarily focus on training efficiency rather than inference speed. Collate addresses this by enabling the collaborative learning of heterogeneous models, allowing each model to meet the specific latency constraints of its respective edge system. The framework incorporates a dynamic zeroizing-recovering method, which intelligently adjusts local model architectures to achieve high accuracy within their latency budgets. Furthermore, Collate features a proto-corrected federated aggregation scheme. This scheme efficiently aggregates diverse local models, ensuring that all systems' latency constraints are met through a single training process while preserving high accuracy. Experimental results demonstrate that Collate significantly improves accuracy for both extended and shrunk models compared to state-of-the-art methods, with negligible additional training overhead.