Reproduce “Mixture of Experts-augmented Deep Unfolding for Activity Detection in IRS-aided Systems” on OpenResearch

About this paper

Zeyi Ren, Qingfeng Lin, Jingreng Lei, Yang Li, Yik-Chung Wu

Researchers at the University of Hong Kong and Great Bay University developed a Mixture of Experts (MoE)-augmented deep unfolding network for activity detection in IRS-aided massive machine-type communications. This approach accurately identifies active devices in complex mixed channel environments without needing prior knowledge of individual device channel types, demonstrating improved detection performance and lower inference latency compared to established baselines.

Key insights

• Adapting model-driven deep unfolding to account for IRS composite channels through modified gradient computations effectively addresses the non-trivial optimization problem in these systems.
• A Mixture of Experts architecture allows deep learning models to autonomously adapt to diverse and mixed channel conditions without explicit prior knowledge of individual device channel types.
• The integration of domain knowledge via deep unfolding with data-driven adaptability via MoE leads to a robust and efficient solution that outperforms purely analytical or black-box deep learning methods.

Summary by alphaXiv.