Jingyan Jiang

we tackle the dual challenge of computationbandwidth trade-off and cost-effectiveness by proposing A2, an efficient joint Adaptive model, and Adaptive data deep learning serving solution across the geo-datacenters. Inspired by the insight that a trade-off between computational cost and bandwidth cost in achieving the same accuracy, we design a real-time inference serving framework, which selectively places different “versions” of the deep learning models at different geolocations, and schedules different data sample versions to be sent to those model versions for inference. We deploy A2 on Amazon EC2 for experiments, which shows that A2 achieves 30%-50% serving cost reduction under the same required latency and accuracy as compared to baselines

Fast-DRD: Fast decentralized reinforcement distillation for deadline-aware edge computing

We everages DRL with a distillation module to drive learning efficiency for edge computing with partial observation. We formulate the deadline-aware offloading problem as a decentralized partially observable Markov decision process (Dec-POMDP) with distillation, called fast decentralized reinforcement distillation(Fast-DRD). compared with naive Policy Distillation, Fast-DRD’s two-stage distillation dramatically reduces the amount of exchanging data, and the learning time and data interaction cost decrease nearly 90%. In a complex environment of heterogeneous users with partial observation, offloading models learned by decentralized learning in Fast-DRD still maintain offloading efficiency.

BACombo—Bandwidth-aware decentralized federated learning

The initial verison of BACombo was accepted by IJCAI 19' FL workshop, which is the first try of segmented aggregation for FL . In the current version, we extend the worker selection scheme.

To avoid the drawback of network congestion in centralized parameter servers architecture, which is adopted in today’s FL systems, we explore the possibility of decentralized FL solution, called BACombo. Taking the insight that the peer-to-peer bandwidth is much smaller than the worker’s maximum network capacity, BACombo could fully utilize the bandwidth by saturating the network with segmented gossip aggregation. The experiments show that BACombo significantly reduces (up to 18×). the training time and maintains a good convergence performance.