A Multi-Modal Learning Framework for MIMO Channel Information Acquisition

Multiple-Input Multiple-Output (MIMO) has been proven to be a successful technology in 5G systems. Applying Artificial Intelligence (AI) and machine learning to MIMO is expected to further boost the spectrum efficiency and capacity of MIMO for 5G-advanced and 6G, because more accurate Channel State Information (CSI) will be enabled via intelligent channel estimation and prediction. Low overhead CSI feedback is also possible when AI is utilized in the channel compression and recovery in frequency division duplex systems. However, designing separate AI models for channel estimation, prediction, and compression/recovery requires huge efforts in model training and large memory to store the model parameters. It is highly motivated to explore how to achieve accurate MIMO CSI with minimum AI model training and storage overhead for all the possible CSI related use cases. In this paper, a multi-modal learning framework is proposed for MIMO channel estimation, prediction, and compression. The framework consists of a single common feature extraction module used in all the use cases, followed by alight weight adaptation modules designed for different use cases. This proposed design is able to avoid the supernumerary extraction of shareable features for different use cases and has been demonstrated to significantly reduce the complexity of the AI models.