Reproduce “Decentralized Low-Latency Collaborative Inference via Ensembles on the Edge” on OpenResearch

About this paper

May Malka, Erez Farhan, Hai Morgenstern, Nir Shlezinger

A framework called "Edge Ensembles" enables decentralized, low-latency collaborative deep learning inference among edge devices, leveraging compact models and communication-efficient feature exchange via trainable vector quantization. This approach achieves substantial accuracy improvements over individual compact models, in some cases outperforming larger centralized models, while preserving device autonomy.

Key insights

• Deep ensembles can be effectively adapted to decentralized edge environments, leveraging collective intelligence to improve inference accuracy.
• Trainable vector quantization provides a practical method to enable low-latency, communication-efficient feature exchange between edge devices.
• An ensemble of diverse compact models can collectively achieve higher accuracy than a single centralized model with a greater total number of parameters.

Summary by alphaXiv.