用於多任務邊緣-雲端部署的架構與切割點聯合搜尋方法

Abstract

為了滿足邊緣裝置上即時且低延遲執行電腦視覺任務的需求，如何高效部署多任務學習（MTL）模型於邊雲協作架構中，成為當前的重要挑戰。由於邊緣裝置的運算資源有限，若僅在邊緣端執行複雜的多任務模型，或完全依賴雲端推論，往往難以在準確率與延遲之間取得理想的平衡。為此，本文提出 SplitNAS，一個統一式架構搜尋框架，可聯合探索任務專屬的分支結構與模型切割點，以支援協同式的邊雲部署。 SplitNAS 採用可微分的神經架構搜尋（Neural Architecture Search）流程，聯合優化任務分支結構與模型切割位置。為了平衡預測效能與系統延遲，SplitNAS 設計了延遲感知的損失函數，綜合考量邊緣運算、雲端推論與資料傳輸的成本。在此基礎上，模型切割點進一步結合 autoencoder 壓縮模組 以降低中間特徵的傳輸負擔，並應用 知識蒸餾（KD）技術 以提升模型準確率。 在 PASCAL-Context 與 NYUD-v2 兩個資料集上的實驗結果顯示，SplitNAS能在不同頻寬與硬體條件下取得優異的準確率與延遲權衡。與純邊緣端或純雲端推論相比，SplitNAS 在 MobileNetV2 與 ResNet34 架構下皆可減少超過 50% 的總體延遲，展現其於實際應用場景中的效能與實用價值。

To meet the demand for real-time and low-latency execution of computer vision tasks on edge devices, efficient deployment of multi-task learning (MTL) models in edge-cloud collaborative settings has become a critical challenge. Due to the limited computational capacity of edge devices, deploying complex multi-task models solely on the edge or fully on the cloud often results in suboptimal tradeoffs between accuracy and latency. To address this, we propose SplitNAS, a unified framework that jointly searches for task-specific branching architectures and optimal partition points for collaborative edge-cloud deployment. SplitNAS adopts a differentiable neural architecture search (NAS) process to jointly optimize task-specific branching structures and model partition points. To balance prediction performance and system latency, it integrates latency-aware loss functions that consider both edge and cloud execution costs, as well as transmission overhead. On top of this, an autoencoder-based compression module is introduced at the partition point to reduce feature transmission cost, and knowledge distillation (KD) is applied to improve the accuracy of the model. Experimental results on the PASCAL-Context and NYUD-v2 datasets demonstrate that SplitNAS achieves superior accuracy-latency trade-offs under various bandwidth and hardware conditions. Compared to pure edge or cloud inference, SplitNAS reduces total latency by over 50% on both MobileNetV2 and ResNet34 backbones, highlighting its effectiveness and practical value for real-world deployment.