Spurred by a plethora of Internet of things (IoT) and smart devices, the network edge has become a major source of data generation. Exploiting the sheer amount of these user-generated private data is instrumental in training high-accuracy machine learning (ML) models in various domains, ranging from medical diagnosis and disaster/epidemic forecast to ultra-reliable and low latency communication (URLLC) and control systems. However, local data is often privacy sensitive (e.g., health records, location history), which prohibits the exchange of raw data samples. In view of this, privacy-preserving distributed ML has recently attracted significant attention. A notable example is federated learning (FL), in which edge devices, i.e., workers, locally train their own ML models that are periodically aggregated and averaged at a parameter server, without exchanging raw data samples. In contrast to classical cloud-based ML, distributed ML hinges on wireless communication and network dynamics, whereby communication may hinder its performance. To obviate this limitation, the communication cost of distributed ML can be decreased by reducing the number of communication rounds until convergence, communication links per round, and/or the payload size per link.

​

Some related publications:

​

Elgabli, Anis, Jihong Park, Amrit S. Bedi, Mehdi Bennis, and Vaneet Aggarwal. "Q-GADMM: Quantized Group ADMM for Communication Efficient Decentralized Machine Learning." In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 8876-8880. IEEE, 2020.

​

Elgabli, Anis, Jihong Park, Amrit S. Bedi, Mehdi Bennis, and Vaneet Aggarwal. "GADMM: Fast and communication efficient framework for distributed machine learning." arXiv preprint arXiv:1909.00047 (2019).

​

Park, Jihong, Shiqiang Wang, Anis Elgabli, Seungeun Oh, Eunjeong Jeong, Han Cha, Hyesung Kim, Seong-Lyun Kim, and Mehdi Bennis. "Distilling on-device intelligence at the network edge." arXiv preprint arXiv:1908.05895 (2019).