EdgeSync: Accelerating Edge-Model Updates for Data Drift through Adaptive Continuous Learning
Runchu Donga, Peng Zhao, Guiqin Wang, Nan Qi, Jie Lin
TL;DR
EdgeSync tackles data drift in real-time edge video analytics by introducing an edge-cloud update loop that filters training samples with timeliness-aware entropy-based scoring and orchestrates adaptive retraining via a continuous training manager. The method combines a sample filtering module, which computes $E(x;\theta) = -f_{\theta}(y|x) \log f_{\theta}(y|x)$, $T(x) = \frac{1}{1+\exp(-t_x/T_w)}$, and $Q(x) = \alpha E(x;\theta) + \beta T(x)$ to select top-$k$ samples for cloud retraining, with a training manager that prioritizes edge models using a history-window signal and a two-phase offline/online hyperparameter strategy. The offline phase uses Bayesian Hyperparameter Optimization with a Gaussian Process surrogate and Expected Improvement to establish $h_0$, followed by mini-batch refinement; the online phase fixes hyperparameters while adaptively controlling training epochs via early stopping and time budgets, freezing backbone layers to speed updates. Empirical results on diverse, real-world video data show EdgeSync achieving up to approximately 8% higher accuracy than No Adaptation and more efficient updates than competing baselines, while maintaining robustness as the number of cameras increases; offline hyper-parameter profiling further reduces computational overhead and speeds convergence. Overall, EdgeSync offers a practical, scalable solution for timely, accurate edge-model updates under data drift, with potential impact on live video analytics deployments across security, traffic, and augmented reality applications.
Abstract
Real-time video analytics systems typically deploy lightweight models on edge devices to reduce latency. However, the distribution of data features may change over time due to various factors such as changing lighting and weather conditions, leading to decreased model accuracy. Recent frameworks try to address this issue by leveraging remote servers to continuously train and adapt lightweight edge models using more complex models in the cloud. Despite these advancements, existing methods face two key challenges: first, the retraining process is compute-intensive, causing significant delays in model updates; second, the new model may not align well with the evolving data distribution of the current video stream. To address these challenges, we introduce EdgeSync, an efficient edge-model updating approach that enhances sample filtering by incorporating timeliness and inference results, thus ensuring training samples are more relevant to the current video content while reducing update delays. Additionally, EdgeSync features a dynamic training management module that optimizes the timing and sequencing of model updates to improve their timeliness. Evaluations on diverse and complex real-world datasets demonstrate that EdgeSync improves accuracy by approximately 3.4% compared to existing methods and by about 10% compared to traditional approaches.