Terahertz communication is a pivotal candidate technology for future 6G networks. Deep learning (DL)-based transmission methods can utilize the real-time data to model channel statistics and device imperfections, presenting an effective manner to solve the modeling problems of non-ideality and channel in the terahertz band. However, there still exist several shortages hindering the development of DL-based terahertz communications, that is high system complexity, costly online retraining to fit changing non-ideality and channel conditions, and learned diagrams with high peak-to-average power ratio (PAPR). This paper proposes novel methods to address the above challenges. At first, a new regulated autoencoder (RAE) structure is proposed to fit changing conditions without online retraining. Secondly, a binary neural network (BNN) method is leveraged to reduce the receiver complexity and a lookup table based method to cut down the transmitter complexity. Lastly, a new maximum normalization method is proposed to reduce PAPR of the learned diagram. Extensive simulations are performed to verify the effectiveness of the proposed methods.

Semantic communications have the potential to improve transmission efficiency and support intelligent tasks. However, the commonly used global sampling-based pattern ignores the fact that the data processing ability of edge transmitters is strictly limited and only a small part of information is available for a single sample in some scenarios. This paper proposes a novel partial sampling-based semantic communication (PSSC) framework where an edge transmitter is guided by feedback from the receiver to locate and collect only part of content relevant to the target task. Taking the vision-based task as an example, the transmitter selectively samples a small patch of a large-size image until the intelligent task is successfully executed at the receiver. The selection of sampling location is modeled as a partially observable Markov decision process problem and an intelligent approach based on reinforcement learning is proposed to solve the problem. In addition, a recurrent neural network-based receiver is designed to fuse information received over multiple transmission rounds. Besides, we prove that the feedback does not increase the semantic channel capacity. Simulation results demonstrate that the proposed framework can locate the informative areas accurately and achieve competitive performance compared to the existing global sampling-based methods.

One primary challenge hindering the development of Terahertz communication is the significant non-ideal device attributes including phase noise (PN) and in-phase/quadrature-phase (I1Q) imbalance that is hard to model offline. Deep learning (DL)-based models can learn and fit channels and non-ideal characteristics by the actual transmission data. This paper proposes two novel constellation shaping methods to limit the distribution of the constellation points in the DL-based modulation for Terahertz communication, in order to resist non-ideal effects in both the training and testing stages. Simulation results show that, without any extra compensation modules, the DL-based modulation and normalization can greatly mitigate non-ideal effects such as PN and I1Q imbalance. In addition to restricted constellation distribution, the proposed normalization and corresponding training methods can learn to have diagrams with smaller peak-to-average power ratio (PAPR), and thus can potentially support a higher average transmission power.

More and more intelligent tasks rely on network edge nodes to collect data, which brings inevitable communication problems such as limited bandwidth and high data volume. As-sisted by deep learning, semantic communication technology is an effective means to improve the efficiency of joint communication and intelligent task execution. However, the commonly used global sampling-based pattern ignores the fact that the data processing ability of edge devices is strictly limited and only a small part of information is available at one time in some applications. This paper proposes a novel partial sampling based semantic communication framework where an edge device is guided by feedback from the receiver to locate and collect only part of the semantic content relevant to the target task at one time. Taking the vision-based task as an example, the transmitter selectively samples a small patch of a large-size image until the intelligent task is successfully executed at the receiver. Simulation results demonstrate that the proposed partial sampling-based framework can locate the informative areas accurately and achieve com-petitive performance compared to existing global sampling-based methods.

Edge devices in existing systems mainly rely on cloud-based services to conduct the semantic understanding, which brings in unnecessary communication overhead and slower response. This paper proposes a novel semantic communication framework for joint semantic transmission and understanding at the network edge, wherein the structured semantic information such as the user intent and slot values that can be understood by machines are extracted directly from the received features without recovering the explicit user input. To achieve this target, we propose to use the slot name as the initial information to generate slot values by a lightweight recurrent neural network. An additional attention mechanism is adopted to integrate the semantic information in the slot value generation process and a hybrid loss is proposed to train the intent prediction and the slot value generation in parallel. Experimental results show that, compared to the conventional separate bit transmission and se-mantic understanding technologies, the proposed system achieves a much better performance under low SNR channels, and greatly reduces the computing overhead at the receiver end and thus lowers the threshold for the local semantic understanding at the edge.

Image compression and bit transmission are con-ducted separately in most existing methods for image trans-mission, leading to possible transmission failure or a waste of communication resource for a time-varying channel condition. This paper proposes a neural network-based image transmission system trained by generative adversarial networks (GANs) aiming to achieve robust transmission. Specifically, the deep semantic of an input image is extracted and represented as bit streams at the transmitter, and the receiver reconstructs the original image based on possible bit error and the same background knowledge as the transmitter. Experimental results show that the proposed robust transmission system trained by GAN can adapt to the current communication condition, and achieve a high-quality reconstruction even with a high transmission error rate and a smaller transmission data size than engineered codecs such as JPEG.

Massive connectivity is a key requirement for the Internet of Things (IoT). In practice, the network should be capable of accommodating thousands of devices and meeting their traffic demands. In this article, we consider the access phase for IoT in a mixed-analog-to-digital converter distributed massive multiple-input-multiple-output system, in which users are classified into light-load users and heavy-load users depending on their traffic load requirements. To meet the low-latency and low-cost demands in IoT, the access scheme for both types of users are designed in a grant-free fashion. For users with light-load traffic demands, by formulating the user activity detection (UAD) and channel estimation (CE) into a compressed sensing problem, we provide a low-complexity algorithm solver which requires no prior information. The simulation results verify that the proposed algorithm can effectively detect user activity and estimate channel state information (CSI) between the users and access points (APs). To satisfy the throughput requirements of heavy-load users, after UAD and CE, a two-step dynamic clustering is proposed for coordinated multipoint transmission using the large-scale fading (LSF) information. The impact of quantization noise on LSF estimation is investigated, as well as, a corresponding compensation method and accuracy bound. By detecting the clustering behavior among users in the first step, the complexity of the joint user and AP clustering is substantially reduced. The numerical results reveal that the proposed algorithm can offer significant performance gains in various scenarios with fast convergence.

In this paper, we consider the channel estimation and multi-user detection problems in fog radio access networks (F-RANs). Based on block coordinate descent algorithm, we propose two methods to solve a mixed ℓ2,1-regularization functional which exploits both the sparsity of user activities and the spatial sparsity of user signals in F-RAN. Both of our methods split the computation and corresponding data into multiple units of a cluster and solve the problem in a distributed manner. Hence they can be deployed flexibly at the distributed logical edges as well as the cloud baseband unit pool in F-RAN. The differences between the two methods are that the first one operates in a serial manner and is guaranteed to converge, while the second one works in parallel and under empirical guidance. Deployment details are also provided. Numerical results demonstrate the effectiveness of the proposed methods.

This paper considers the channel estimation (CE) and multi-user detection (MUD) problems in cloud radio access network (C-RAN). By taking into account of the sparsity of user activities in C-RAN, we solve the CE and MUD problems with compressed sensing to greatly reduce the large pilot overhead. A mixed ℓ2,1-regularization penalty functional is proposed to exploit the inherent sparsity existing in both the user activities and remote radio heads with which active users are associated. An iteratively re-weighted strategy is adopted to further enhance the estimation accuracy, and empirical and theoretical guidelines are also provided to assist in choosing tuning parameters. To speed up the optimization procedure, three low-complexity methods under different computing setups are proposed to provide differentiated services. With a centralized setting at the baseband unit pool, we propose a sequential method based on block coordinate descent (BCD). With a modern distributed computing setup, we propose two parallel methods based on alternating direction method of multipliers (ADMM) and hybrid BCD (HBCD), respectively. Specifically, the ADMM is guaranteed to converge but has a high computational complexity, while the HBCD has low complexity but works under empirical guidance. Numerical results are provided to verify the effectiveness of the proposed functional and methods.

This paper considers the channel estimation (CE) and multi-user detection (MUD) problems in cloud radio access network (C-RAN). Assuming that active users are sparse in the network, we solve CE and MUD problems with compressed sensing (CS) technology to greatly reduce the long identification pilot overhead. A mixed ℓ2.1-regularization functional for extended sparse group-sparsity recovery is proposed to exploit the inherently sparse property existing both in user activities and remote radio heads (RRHs) that active users are attached to. Empirical and theoretical guidelines are provided to help choosing tuning parameters which have critical effect on the performance of the penalty functional. To speed up the processing procedure, based on alternating direction method of multipliers and variable splitting strategy, an efficient algorithm is formulated which is guaranteed to be convergent. Numerical results are provided to illustrate the effectiveness of the proposed functional and efficient algorithm.

In this paper, we investigate robust energy efficiency design for multicast downlink cloud radio access network (C- RAN) for both data-sharing and compression strategies. The two strategies mainly differ in whether the baseband units pool compresses the user messages before transmitting them to remote radio heads (RRH). The performance of worst-case energy efficiency for both strategies is compared by formulating the robust energy efficiency maximization problem subjected to finite RRH power budgets, limited fronthaul link capacity, and minimum quality-of-service constraints when only imperfect channel state information (CSI) is available at the transmitter. To solve these problems, we cast them into semidefinite program problems and solve them iteratively. Simulation results demon- strate effectiveness of our proposed algorithms and two strategies are compared under imperfect CSI.

In this paper, a new iteratively reweighted least squares method is proposed for recovery of block-sparse signals with unknown cluster patterns. In many practical applications, sparse signals have block-sparse structures with nonzero coefficients occurring in clusters, while the prior information of the cluster pattern is usually unavailable. To address this issue, we propose an element-overlapping log-sum functional to encourage the sparseness and the cluster pattern simultaneously. The algorithm is developed by iteratively minimizing a convex surrogate function that majorizes the original objective function, which results in an iteratively reweighted process that alternates between estimating the sparse signal and refining the weights of the surrogate function. Convergence of the iterations to a local minimum of the penalty function is also guaranteed. Numerical results are provided to illustrate the effectiveness of the proposed method.