Data-Efficient Stream-Based Active Distillation for Scalable Edge Model Deployment

2025-09-24

AI Summary

To address the high transmission overhead and low data utilization efficiency in model continual updating for edge camera systems, this paper proposes an efficient active learning framework tailored for streaming video data. Methodologically, it integrates high-confidence sample filtering with diversity-aware sampling, coupled with server-side large-model-guided lightweight knowledge distillation to enable high-quality model updates at the edge under low query rates. Its key innovation lies in establishing a “selection–distillation” co-optimization paradigm, substantially reducing both image transmission and manual annotation costs. Experimental results demonstrate that, with comparable training iterations, the proposed method achieves baseline model performance using only ~10% of the queried data, improves model accuracy by 2.3–4.7%, and enhances deployment scalability by 3.2×.

Technology Category

• Machine Learning: Learning on the Edge & Model Compression
• Computer Vision: Video Understanding & Activity Analysis
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs

Abstract

Edge camera-based systems are continuously expanding, facing ever-evolving environments that require regular model updates. In practice, complex teacher models are run on a central server to annotate data, which is then used to train smaller models tailored to the edge devices with limited computational power. This work explores how to select the most useful images for training to maximize model quality while keeping transmission costs low. Our work shows that, for a similar training load (i.e., iterations), a high-confidence stream-based strategy coupled with a diversity-based approach produces a high-quality model with minimal dataset queries.

Problem

• Selecting most useful images for training edge models efficiently
• Maximizing model quality while minimizing transmission costs
• Developing data-efficient strategies for scalable edge deployment

Innovation

• Stream-based active distillation for edge model deployment
• High-confidence strategy with diversity-based image selection
• Minimizes dataset queries while maximizing model quality