Toward auction-based edge AI: Orchestrating and incentivizing online transfer learning in edge networks

The edge AI service can procure external pre-trained models to “strengthen” its local edge models by conducting online transfer learning from the former to the latter to serve inference requests continuously and resiliently. In this paper, we present a modeling and algorithmic study of designing an auction-based mechanism to realize this scenario, while overcoming unique challenges of the interdependency between consecutive auctions, the intertwinement between system overhead and models’ inference performance, and the need of preserving desired economic properties in each auction. We first formulate a long-term social cost minimization problem, which is unsurprisingly intractable. We then design a group of polynomial-time online approximation algorithms that decouple the interdependency, solve each relaxed auction individually, and round the fractional decisions into integers to procure and deploy pre-trained models. Our algorithms also control online transfer learning on each edge by adapting the weights associated to different models and updating the edge model itself through streaming data. We rigorously prove the competitive ratio of our approach, the upper bound on the number of the inference mistakes, and the truthfulness and the individual rationality of each auction. Via extensive evaluations, we have validated the superior performance of our approach over multiple alternative methods.