Reproduce “EMA: Efficient Model Adaptation for Learning-based Systems” on OpenResearch

About this paper

Daiyang Yu, Xinyu Chen, Yihan Zhang, Yan Liang, Yaqi Qiao, Fan Lai

Researchers at the University of Illinois Urbana-Champaign and The Hong Kong University of Science and Technology developed EMA, a general model adaptation system that efficiently adjusts learning-based systems to changing environments. It reduced total adaptation costs by 14.9–42.4%, accelerated adaptation by 2.3–15.3 times, and improved post-adaptation system performance by 6.9–31.3% across various network applications.

Key insights

• Aligning input state distributions between new and historical environments through a model-agnostic State Transformer enables efficient knowledge transfer and warm-starts model fine-tuning.
• A Labeling Agent that prioritizes data points based on model prediction uncertainty and their heterogeneous labeling cost effectively reduces data acquisition expenses during adaptation.
• An Adaptation Orchestrator dynamically balances the trade-off between continued training on existing data and acquiring new, labeled data by assessing marginal utility per unit cost.

Summary by alphaXiv.