DUET: A Tuning-Free Device-Cloud Collaborative Parameters Generation Framework for Efficient Device Model Generalization

TL;DR

DUET tackles the challenge of device model generalization under device–cloud constraints by shifting personalization from on-device fine-tuning to cloud-generated, device-specific parameters. It combines a Universal Meta Network to establish a global backbone and classifier, a Personalized Parameters Generator that uses device real-time data to produce dynamic layer weights, and a Stable Weight Adapter to reduce oscillations and improve robustness. The approach achieves faster adaptation with negligible on-device computation and minimal communication, outperforming fine-tuning baselines across sequential recommendation and vision benchmarks. The work demonstrates practical DMG with real-time latency, generalization across devices, and privacy-preserving data handling via embedding-based transmission.

Abstract

Device Model Generalization (DMG) is a practical yet under-investigated research topic for on-device machine learning applications. It aims to improve the generalization ability of pre-trained models when deployed on resource-constrained devices, such as improving the performance of pre-trained cloud models on smart mobiles. While quite a lot of works have investigated the data distribution shift across clouds and devices, most of them focus on model fine-tuning on personalized data for individual devices to facilitate DMG. Despite their promising, these approaches require on-device re-training, which is practically infeasible due to the overfitting problem and high time delay when performing gradient calculation on real-time data. In this paper, we argue that the computational cost brought by fine-tuning can be rather unnecessary. We consequently present a novel perspective to improving DMG without increasing computational cost, i.e., device-specific parameter generation which directly maps data distribution to parameters. Specifically, we propose an efficient Device-cloUd collaborative parametErs generaTion framework DUET. DUET is deployed on a powerful cloud server that only requires the low cost of forwarding propagation and low time delay of data transmission between the device and the cloud. By doing so, DUET can rehearse the device-specific model weight realizations conditioned on the personalized real-time data for an individual device. Importantly, our DUET elegantly connects the cloud and device as a 'duet' collaboration, frees the DMG from fine-tuning, and enables a faster and more accurate DMG paradigm. We conduct an extensive experimental study of DUET on three public datasets, and the experimental results confirm our framework's effectiveness and generalisability for different DMG tasks.

Figures (6)

• Figure 1: (a) describes the device model generalization in device-cloud collaboration, $\Delta S$ indicates the data distribution shift of global and local data. (b) and (c) are overviews of fine-tuning based approaches and our DUET, respectively. $\Omega_b$ and $\Omega_c$ respectively denoted the parameters of backbone and classifier. (d) is the comparison of fine-tuning and DUET (Time Delay: 0.34ms (DUET) $\ll$ 60,000ms (Fine-tuning)), AUC: 0.9230 (DUET) $>$ 0.9080 (Fine-tuning)).
• Figure 2: Overview of the proposed DUET. The UMN is trained on the cloud that contains a backbone with parameters $\Theta_b$ and a classifier with parameters $\Theta_c$. The PPG is deployed on the cloud, which generates and delivers the personalized parameters ${\Theta}^c_l$ of dynamic layers for the device classifier based on the distribution of the real-time samples uploaded from the device. The SWA aims to reduce the performance oscillation of single PPG, accelerating the convergence and improving the prediction stability.
• Figure 3: Effects of the number of SWAs in training.
• Figure 4: Performance comparison on each session.
• Figure 5: Training visualization of DUET and baselines.