1. Introduction

Massive multiple-input multiple-output (MIMO) technology plays a crucial role in delivering the potential of 5G and meeting its requirements. To fully exploit the potential of MIMO technology, it is crucial to obtain downlink channel state information (CSI) on the base station (BS) side by feeding the estimated downlink CSI from the user equipment (UE) back to the BS through the uplink channel, as shown in Fig. 1. The feedback of CSI leads to extra overhead, which escalates significantly with the increase in the number of service antennas. When the uplink channel resources are limited, the challenge of CSI feedback lies in maintaining the accuracy of the feedback while minimizing the feedback overhead.

Traditional methods for reducing the CSI feedback overhead include techniques based on codebooks or compressed sensing (CS). In the past few years, artificial intelligence (AI) has been utilized in the CSI feedback space to improve the precision of CSI reconstruction; see, for instance, the AI for CSI feedback enhancement in the study item of the Third Generation Partnership Project (3GPP) Release 18 [1].

In this article, we first introduce an AI-based one-sided CSI feedback method, wherein the DL model is only employed at the BS, and AI-based multi-module learning involving the CSI feedback. In addition, we articulate the prospects and challenges pertaining to AI-driven CSI feedback tasks, taking into account both the AI limitations and the real-world implementation of AI models.

2. AI-Based One-Sided CSI Feedback

For one-sided CSI feedback, the AI model is implemented solely on the BS, to reduce transmission overhead and improve performance. In addition, BS-sided model can avoid multi-vendors collaboration effort and preserve model confidentiality and data privacy. One-sided CSI feedback architectures, such as CS-CsiNet [2], CFnet [3], and CSI-PPPNet [4], improve CSI reconstruction using deep learning (DL). CSI-PPPNet [4] uses a single DL model for arbitrary compression rates, with CSI compressed at the UE and iteratively recovered at the BS, a feature not achieved by the other two methods. This approach decouples training from compression, simplifying model maintenance for the network vendor. Fig. 2 shows the normalized mean square errors (NMSE) performance of AI-based one-sided CSI feedback (CS-CsiNet, CSI-PPPNet) and two-sided CSI feedback (CsiNet) for indoor and urban macrocell (UMa) scenarios, with 256 subcarriers. The BS uses a ULA with 32 omnidirectional antennas, and the UE uses a single omnidirectional antenna.

3. Joint Multi-Module Learning with CSI Feedback

To fully exploit MIMO technology and improve throughput, it is crucial to jointly design multiple modules, including channel coding, CE, PD, and precoding, alongside the CSI feedback task.

3.1 Joint CSI Compression and Channel Coding

The CSI compression can be viewed as a source compression task, which can be designed independently of channel coding, following the paradigm of separate source-channel coding (SSCC).

Using AI-based source coding methods, the CSI is compressed by an encoder and reconstructed using the corresponding decoder. When the quality of the channel deteriorates to such an extent that it surpasses the processing capacity of the channel coding, the precision of the reconstructed CSI within the SSCC framework experiences a significant drop, a phenomenon referred to as the ````cliff effect''. The use of heavily distorted CSI for the creation of precoding vectors results in an undesirable reduction in the system throughput. While the hybrid automatic repeat request mechanism and other methods can alleviate this problem, they also cause a delay in the downlink CSI acquisition and an increase in the feedback overhead.

To address the ''cliff effect'', a new CSI feedback architecture, coined as DJSCC, was introduced in [5]. This approach leverages DL to integrate source and channel coding, training the system with a dataset that encompasses both the source and the wireless channel. The performance of the SSCC and DJSCC networks for CSI feedback with the same overhead is shown in Fig. 3. Both networks are trained and tested on the same dataset, which is generated by QuaDriGa in an FDD indoor scenario. The number of feedback symbols n is set to 16. We employ explicit feedback, where the system employs 256 subcarriers. The BS utilizes a ULA with half-wavelength antenna spacing and 32 omnidirectional antennas, while the UE deploys a single omnidirectional antenna.

From the experimental results shown in Fig. 3, we can observe that the SSCC network suffers from a severe ''cliff effect''. DJSCC addresses this issue by yielding only a gradual rise in NMSE with the reduction of the SNR in the feedback channel, while enhancing the performance across all SNR levels. To address the generalization problem, an attention mechanism-based DJSCC network (ADJSCC) was proposed in [5] to dynamically adjust the ratio of source coding and channel coding outputs based on the uplink SNR.

3.2 Joint CSI Compression, Channel Coding and Precoding

In a multiuser MIMO scenario, the precoding vector for one UE affects interference to/from other UEs. To maximize the downlink sum-rate, precoding vectors are jointly designed after the BS aggregates CSI from all UEs. Joint feedback and precoding networks (JFPNet) have been proposed to optimize both feedback and precoding design [6]. Notably, [6] employs implicit feedback, where the feedback source is the eigenvector matrix rather than full CSI.

Fig. 4 shows the downlink SE of different methods. Initially, the downlink CSI undergoes preprocessing to generate the eigenvalues and the eigenvector matrix, and the DJSCC network is then used to compress the eigenvector matrix for feedback. The uplink channel SNR during the training phase spans a range from -10 dB to 10 dB, following a uniform distribution and the feedback overhead is fixed to n=32 symbols. The BS and UE use ULA omnidirectional arrays with 32 and 4 elements, respectively. The full CSI consists of 624 subcarriers, i.e., 13 subbands. The JFPNet, which jointly implements CSI compression, channel coding and precoding, achieves the highest SE at all SNRs. The label ''SFPNet'' indicates that the precoding module is trained separately from the DJSCC module for CSI feedback, and ''JFPNet'' refers to the joint training of the CSI feedback and precoding modules. ''JFPNet’’ outperforms ''SFPNet'', highlighting the benefits of multi-module learning for extracting task-related semantic information. ''DJSCC_BD_WF'' and ''PF_BD_WF'' refer to DJSCC CSI feedback and perfect CSI feedback, respectively, using traditional non-AI precoding methods (block diagonalization (BD) and water-filling (WF) algorithms).

4. Opportunities and Challenges

DL-based CSI feedback in FDD massive MIMO systems reduces overhead and complexity. Using a one-sided model and joint multi-module learning enhances practical deployment and end-to-end performance. However, challenges remain in both AI technology and the real-world implementation of AI models for CSI feedback. Current datasets are mostly statistical, generated by simulators under the assumption that channels across cells are independent and identically distributed. However, this assumption doesn’t reflect real-world conditions, leading to mismatched data and reduced model performance, which can degrade system throughput.

To support AI/ML for CSI feedback in current 5G or future 6G systems, for standardization, data collection, performance monitoring, and feedback of inference results are the main considerations. Data collection is needed for model training and performance monitoring, including data for AI inputs and ground truth for labeling. Data omission or compression may be needed to reduce the overhead. Performance monitoring is another unique aspect compared to traditional non-AI communication systems. Metrics and procedures for performance monitoring also need to be standardized. Recall that new signaling for model inference results in feedback. For example, for joint multi-module learning for CSI feedback, and since channel coding and potential modulation are already included in AI encoders and the output of the encoder may not fit in the existing signaling processes, new CSI feedback types need to be defined in standards. Besides, two-sided models, wherein the encoder and decoder need to be pairwise trained, will require inter-vendor calibration. To avoid offline calibrations between different vendors, some proposals are under discussion in 3GPP in Release 19. For example, the training data set and the model structure can be standardized. Alternatively, a well-trained model or model parameters can be delivered over the air interface to improve the performance.

5. Conclusion

In this article, we investigated the recent developments of DL-based CSI feedback from the perspectives of one-sided model and joint multi-module learning. Firstly, we introduced a one-sided CSI feedback architecture that only replaces the traditional decoding module with a DL module. Subsequently, we delved into CSI feedback architectures, jointly learned with various modules, e.g., channel coding, and precoding. For example, the joint design of the CSI feedback and channel coding via DJSCC overcomes the “cliff effect” observed in the traditional SSCC-based CSI feedback methods. Unlike traditional approaches of designing each module independently, multi-module joint design leverages DL techniques to jointly optimize the various modules and exploit the interdependencies between different modules, and can consequently enhance the overall system performance and training efficiency.

Link to the paper

https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=10812964