Peer-to-Peer Learning + Consensus with Non-IID Data

The article discusses peer-to-peer deep learning algorithms that allow distributed devices to collaboratively train deep neural networks without exchanging raw data or relying on a central server. The authors identify factors that amplify performance oscillations and introduce a new approach, P2PL with Affinity, which reduces test performance oscillations in non-IID settings without incurring additional communication costs. The paper also highlights the importance of peer-to-peer deep learning approaches in the context of the proliferation of IoT devices and evolving 6G environments.

Publication date: 22 Dec 2023
Project Page: Not Provided
Paper: https://arxiv.org/pdf/2312.13602