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

TL;DR

MatchNAS tackles cloud-to-edge DNN porting under label scarcity by coupling semi-supervised learning with neural architecture search within a single supernet. The largest subnet generates high-quality pseudo-labels to guide smaller subnets, while a zero-shot NAS scorer selects platform-specific architectures under resource constraints, formalized as $\min_W \mathbb{E}_{\alpha \in \mathcal{A}}[ \mathcal{L}^l_A + \mathcal{L}^u_A + \sum_{i=1}^{n-1}( \mathcal{L}^l_{\alpha_i} + \mathcal{L}^u_{\alpha_i} ) ]$ and $\alpha^* = \arg\max_{\alpha \in \mathcal{A}} \mathcal{S}(\alpha, R)$. Empirically, MatchNAS yields notable improvements in small/medium subnets across four image datasets and delivers favorable on-device latency-accuracy trade-offs on multiple smartphones, while reducing training cost by training a single supernet rather than per-platform models. The approach remains robust to varying labelled data and can be further boosted by narrowing the search space or integrating with other NAS frameworks. These results illuminate a practical path to scalable, label-efficient edge AI deployment.

Abstract

Recent years have seen the explosion of edge intelligence with powerful Deep Neural Networks (DNNs). One popular scheme is training DNNs on powerful cloud servers and subsequently porting them to mobile devices after being lightweight. Conventional approaches manually specialized DNNs for various edge platforms and retrain them with real-world data. However, as the number of platforms increases, these approaches become labour-intensive and computationally prohibitive. Additionally, real-world data tends to be sparse-label, further increasing the difficulty of lightweight models. In this paper, we propose MatchNAS, a novel scheme for porting DNNs to mobile devices. Specifically, we simultaneously optimise a large network family using both labelled and unlabelled data and then automatically search for tailored networks for different hardware platforms. MatchNAS acts as an intermediary that bridges the gap between cloud-based DNNs and edge-based DNNs.

Figures (7)

• Figure 1: MatchNAS bridges the Cloud AI and Edge AI.
• Figure 2: A workflow for MatchNAS. Given a pre-trained cloud-based DNN and a set of resource constraints, MatchNAS first transforms the DNN to a supernet and inherits its network weights. Then, MatchNAS conducts a semi-supervised-NAS training, which is a combination of semi-supervised learning and one-shot NAS, to transfer the supernet to a label-scarce dataset. After training, MatchNAS leverages the zero-shot NAS techniques to efficiently sample high-quality subnets from the supernet according to the resource constraints without further training and build a network family for efficient network mobile porting.
• Figure 3: The semi-supervised-NAS training in MatchNAS
• Figure 4: Latency-accuracy trade-off on four mobile devices.
• Figure 5: (a): Network performance of a set of subnets on Cifar-100; (b): Comparison of different unlabelled loss types.