Efficient Neural Architecture Search for Split Computing via Supernet Fine-Tuning

Split Computing (SC) is a method that partitions a deep neural network into a head model running on resource-constrained edge devices and a tail model executing on powerful servers. The edge device computes intermediate feature maps, which are sent to the server, significantly lowering local computational demands and reducing communication compared to sending raw input data. This results in improved latency-accuracy trade-offs, enabling real-time inference of large models on constrained edge hardware. While previous research on SC primarily focused on improving training methods, model optimization through automated design approaches such as Neural Architecture Search (NAS) remains limited due to substantial supernet training time. This paper proposes a novel method that efficiently applies NAS to SC by fine-tuning a pre-trained supernet from a general Two-stage NAS framework into one tailored for SC. Using publicly available pre-trained supernets significantly reduces training time. Experiments on ImageNet-1K demonstrate that our proposed method achieves better latency-accuracy trade-offs compared to state-of-the-art NAS and manual SC designs, requiring only about 13 to 15% of original supernet training time. Additionally, for a practical case study on object detection using the COCO 2017 dataset, models optimized with our proposed method as backbones achieved a 69% latency reduction and 0.4% improved average precision over the EfficientDet baseline.