Responsive DNN Adaptation for Video Analytics against Environment Shift via Hierarchical Mobile-Cloud Collaborations

TL;DR

Environment shifts in mobile video analytics demand rapid adaptation, yet cloud-centric solutions incur prohibitive latency. MOCHA presents a mobile-cloud hierarchical framework that distributes lightweight on-device reuse and single-layer LoRA fine-tuning with cloud-backed end-to-end retraining, guided by a semantic taxonomy for fast model retrieval. Key contributions include an on-device adaptation hierarchy, a domain-semantics taxonomy for real-time model selection, and a cache-enabled mobile layer that enables proactive prefetching and fast switching, all validated on real-world video tasks with substantial improvements in recovery accuracy and response latency. The framework promises enhanced responsiveness and scalability for autonomous driving and other mobile video analytics applications under varying environments.

Abstract

Mobile video analysis systems often encounter various deploying environments, where environment shifts present greater demands for responsiveness in adaptations of deployed "expert DNN models". Existing model adaptation frameworks primarily operate in a cloud-centric way, exhibiting degraded performance during adaptation and delayed reactions to environment shifts. Instead, this paper proposes MOCHA, a novel framework optimizing the responsiveness of continuous model adaptation through hierarchical collaborations between mobile and cloud resources. Specifically, MOCHA (1) reduces adaptation response delays by performing on-device model reuse and fast fine-tuning before requesting cloud model retrieval and end-to-end retraining; (2) accelerates history expert model retrieval by organizing them into a structured taxonomy utilizing domain semantics analyzed by a cloud foundation model as indices; (3) enables efficient local model reuse by maintaining onboard expert model caches for frequent scenes, which proactively prefetch model weights from the cloud model database. Extensive evaluations with real-world videos on three DNN tasks show MOCHA improves the model accuracy during adaptation by up to 6.8% while saving the response delay and retraining time by up to 35.5x and 3.0x respectively.

Figures (25)

• Figure 1: Teacher Model vs. Expert Model. The change ratio denotes the ratio of accuracy drop in the current domain over accuracy in the optimal domain, reflecting model generalizability in different environments.
• Figure 2: Example frames of the domains in Figure \ref{['fig:golden_vs_expert']}.
• Figure 3: Communication overhead ratio comparison. The system first calculates the model to reuse and then transmits model weights to the mobile device.
• Figure 4: Onboard Reuse v.s. Cloud Reuse.
• Figure 5: Reuse on similar domains. Pearson correlation coefficient (PCC) is calculated from the distance and mAP for each validation domain.