Description
Date depot: 24 mars 2024
Titre: Graphic spiking neural network (GSNN) based edge computing for MIMO system linearization through over-the-air (OTA)
Directeur de thèse:
Aziz BENLARBI-DELAÏ (GeePs (EDITE))
Encadrant :
Siqi WANG (GeePs (EDITE))
Domaine scientifique: Sciences pour l'ingénieur
Thématique CNRS : Intelligence artificielle
Resumé: Context: This work is part of the research activities carried out within GeePs Laboratory (UMR CNRS 8507) on low-power neuromorphic circuits and their extension to the artificial intelligence network. The work carried out on this theme aims to analyze information processing by low-power circuits inspired by biological neurons and to exploit the constraint edges on the performance of the neural networks.
Research on artificial intelligence (AI) systems has recently become very popular in academia and industry. Most of these AI systems are made with ANN and operated on conventional digital circuits, which give impressive performance but consume much energy. Knowing that the density of the cerebral cortex of a human being is around 0.01W/cm² and that of a microprocessor of today's technique is 100W/cm², an intuitive idea is to propose new techniques that emulate neurons and synapses, which gives us neuromorphic circuits and SNN networks.
Graphic neural networks, an emerging technology for artificial intelligence (AI), are designed to process and analyze data represented as graphs or networks. Unlike traditional neural networks, which operate on grid-structured data like images or sequences, GNNs excel at handling non-grid data with intricate relationships and dependencies. The network iteratively aggregates information from neighboring nodes, which allows it to capture complex dependencies and patterns within the graph structure, such as in radio telecommunication spectra processing. The graphic SNN (GSNN) may share both the advantage of SNN on power consumption and the advantage of GNN on accuracy [Ruan2024].
The linearization of RF PA in a base station commonly employs a DPD technique that needs the PA's characterization [Wang2023]. However, in 5G and beyond telecommunication systems, a massive, phased antenna array is used for beamforming and multi-input multi-output (MIMO), which brings challenges to characterizing the PA and the channel information. Classical methods based on feedback loops using a coupler at each PA’s output become fastidious and impractical in this case. To solve the DPD model identification problem, obtaining the PA characteristics and the channel information using UE through the over-the-air (OTA) method becomes promising.
This research aims to investigate and develop an innovative approach for linearizing MIMO systems by leveraging GSNN within the edge computing framework. The focus will be on achieving real-time and adaptive linearization through OTA feedback, allowing for efficient and dynamic compensation of non-linearities inherent in MIMO systems. The proposed study will investigate the potential of utilizing GSNNs at the user equipment (UE), such as mobile phones, to mitigate nonlinearities inherent in MIMO systems, employing real-time OTA adjustments for enhanced performance and efficiency. The research will focus on training GSNN models with OTA-collected samples to adaptively correct non-linear distortions, contributing to optimizing wireless communication networks. This study's outcomes can potentially revolutionize MIMO system design, offering improvements in energy consumption, reliability, and overall network efficiency.