MatchNAS: Optimizing Edge AI in Sparse-Label Data Contexts via Automating Deep Neural Network Porting for Mobile Deployment

Hongtao Huang, Lina Yao, Xiaojun Chang, Wen Hu

Keywords: Edge AI, mobile intelligence, deep neural network, AutoML

Abstract: Recent years have seen the explosion of edge intelligence with powerful deep learning models. As 5G technology becomes more widespread, it has opened up new possibilities for edge intelligence, where the cloud-edge scheme has emerged to overcome the limited computational capabilities of edge devices. Deep-learning models can be trained on powerful cloud servers and then ported to smart edge devices after model lightweight. However, porting models to match a variety of edge platforms with real-world data, especially in sparse-label data contexts, is a labour-intensive and resource-costing task. In this paper, we present MatchNAS, a neural network porting scheme, to automate network porting for mobile platforms in label-scarce contexts. Specifically, we employ neural architecture search schemes to reduce human effort in network fine-tuning and semi-supervised learning techniques to overcome the challenge of lacking labelled data. MatchNAS can serve as an intermediary that helps bridge the gap between cloud AI and edge AI, facilitating both porting efficiency and network performance.

Track: COI (submissions co-authored by SAC)

Submission Guidelines Scope: Yes

Submission Guidelines Blind: Yes

Submission Guidelines Format: Yes

Submission Guidelines Limit: Yes

Submission Guidelines Authorship: Yes

Student Author: Yes

Submission Number: 1301