Contents Introduction
Traditional communication systems often overlook the significance of the meaning behind information. They operate on the assumption that all symbols or bits are of equal importance and are handled as such. The primary objective of these systems is to ensure the accurate retrieval of transmitted sequences at receiving ends, prioritizing conformity in transmission. The design methods in this field have predominantly relied on principles from digital communication. Information theory establishes the maximum limits on the capacity of the system. While channel coding concentrates on developing strategies that can approach these limits with extremely low error probability, source coding (known as data compression) refers to the process of encoding information in a way that reduces the amount of data required with the objective of optimizing the length of the source encoded sequence. However, the latest generation of communication systems is being applied in ways that challenge the conventional design paradigm, particularly in terms of semantic and task-oriented aspects [1].
The semantic aspect pertains to the level of precision and accuracy with which transmitted symbols are able to convey the intended meaning and comprises the transmission of a notion or informational material from a source to a destination without delving into the intricacies. It entails the comparison of the inferred meaning by the recipient with the intended meaning by the sender while considering the content, requirements, and semantics in order to enhance the communicative system towards a state of intelligence. Furthermore, the task-oriented aspect prioritizes task completion and efficiency and examines the potential ramifications associated with the utility of the provided information. The efficacy of a task or performance metric is determined by how efficiently the acquired information aids in its accomplishment. The achievement of a shared aim within task-oriented limitations and requirements is facilitated by the utilization of available resources, including communication bandwidth, computing expense, and power consumption. The evaluation of system performance can be measured in relation to the extent to which a certain objective is achieved, taking into account the allocated resources, instead of considering all the transmission sequences that can typically be conveyed in the aforementioned information theory framework-based approach. Based on the stated task objectives and existing knowledge, it can be demonstrated that semantic source coding [2] has the potential to achieve greater reductions in redundancy expense and communication overhead compared to the Separate Source-Channel Coding (SSCC). This is mostly due to its ability to refine the most pertinent and concise information and then condense it. An intriguing aspect worth exploring is the extent of compression achieved in semantic source coding in relation to the original information [3]. The authors have obtained the theoretical boundaries for lossless and lossy compression based on this semantic source, together with the lower and upper limitations on the rate-distortion function.
Furthermore, semantic-information feature learning [4] is the key to addressing the utilization of cognitive techniques employed as a means to direct computational resources towards activities of higher importance. The process of achieving dynamic adaptation in semantic compression involves the utilization of a feature learning module and an attention feature module. These modules enable the source encoder to generate a variable number of symbols and modify the capability of the source encoder and channel encoder through the use of cognitive techniques. Moreover, task-oriented feature learning [5] addresses that the attainable precision of inference is contingent upon the amount of feature components collected and the extent to which they are distorted by detecting noise and quantization defects. The classification gain is only influenced by the distributions of classes in the feature space and represents the highest possible inference accuracy that can be attained theoretically. Therefore, this learning accuracy is dependent on not only the underlying design of the artificial intelligence models being used but also the classification distributions of the feature space. In order to facilitate the implementation of immediate intelligent services in future communication systems, it is advantageous to distill and communicate merely the information that is pertinent to the task at hand and precise semantics. The semantic and task-oriented approach effectively reduces the overall latency of the system. However, it should be noted that this method deviates from the optimal Shannon’s SSCC [6], which is typically applied in the communication of long block-length bit sequences. SSCC employs advanced compressing methods to eliminate all redundant sequences from source-encoded symbols, whereas Joint Source Channel Coding (JSCC) exploits the remaining redundant information that emerges following compressing with the error-correcting capability in order to minimize distortion within a certain limit of codelength. Unequal error protection (UEP) prevents errors from occurring by assigning encoded redundancies according to the significance of the information bits. Only some sets of bits are of equal significance when sending source-encoded information because of the varied degree of vulnerability of the source decoder. Therefore, the authors in [7] present a remarkable performance of the UEP JSCC code system by providing an innovative adjustable code rate for multiple semantic task classes. We summarize our contribution from the following perspectives:
This paper provides a brief exploration of the latest designs and methodologies in semantic and task-oriented communication, with a specific emphasis on the knowledge pertaining to the prototype of the JSCC scheme. The proposed fixed-point JSCC scheme serves as a practical solution not only for transmitting and receiving semantic features over noisy channels but also for UEP semantic importance applications. This study presents a novel approach for enhancing the semantic encoder/decoder by including the UEP capabilities of the quasi-cyclic Low-Density Parity-Check (QC-LDPC) JSCC decoder. The primary objective is to safeguard the semantic significance of 6G communication.
The primary aim of this study is to explore the potential of surpassing the traditional Shannon criteria, while also presenting the initial iteration of a prototype for a semantic JSCC communication system. The JSCC prototype significantly reduces the data width from a 32-bit floating-point to a 6-bit fixed-point, making practical FPGA implementation possible. Additionally, we revisit the design of semantic codec learning and task-oriented signal compression in order to explore how the semantic task-oriented techniques can be adapted to the proposed JSCC platform.
The JSCC system under consideration acts as a prototype to make it easier to modify communication protocols in the future. Through the use of deep learning techniques, this system is specifically created to be adaptable to a broad variety of semantic and task-oriented features. The proposed prototype, when compared to other state-of-the-art, can deliver better BER performance despite the reduction in fix-point implementation and is achieved by applying QC-LDPC codes with a reasonable code rate.
In this paper, we revisit the design methodologies of next-generation 6G communication systems with regard to semantic and task-oriented aspects and JSCC system following the demonstration of the state-of-the-art prototype designs in Section II. Second, we investigate the applications of a JSCC system based on QC-LDPC codes. By leveraging quasi-cyclic characteristics and optimizing the whole system, QC-LDPC codes are adopted to enable the feasibility of a JSCC system. We deploy the JSCC system in a Field-Programmable Gate Array (FPGA)-based platform so the users can configure this platform after manufacture. We also demonstrate that semantic feature transmission and reception are potential application scenarios. The JSCC system can maintain a high code rate (0.8) at a low level of
Design of Next Generation Communication and Joint Source Channel Coding
In this section, we examine the application of JSCC system in relation to the integration of AI for semantic and task-oriented considerations to explore the opportunities of future communications.
A. Semantic Codec Learning
The evolution of semantic communication [2] is traced back to the early 20 th century with its continuous growth into the realm of modern communications with regard to beyond 5G and 6G technologies. In light of this, there is a significant need to develop more intelligent and efficient communication protocols that can meet the diverse quality of service (QoS) needs. This must be done while addressing the challenge of limited communication bandwidth and computation. The development of an intelligent communication system is considered essential in both industry and academia. Such a system is not limited to memorizing data flows based on rigorous regulations but also aims to comprehend, analyze, and articulate the fundamental semantics. The ambition of “beyond Shannon” [8] surpasses the conventional Shannon paradigm, which focuses on ensuring the accurate receipt of individual sent bits, regardless of the conveyed meaning of these bits.
In the context of conveying meaning or achieving a goal through communication, the crucial factor lies in the influence exerted by the received sequences on the interpretation of the message meant by the sender or on the attainment of a shared objective. However, despite its growing popularity, research on semantic communication remains fragmented and encompasses a wide range of research interests. The evolution of human dialogues in relation to the semantics of everyday usage for the purpose of semantic communication is still in its infancy. The challenge of developing theoretical semantic models for actual multidimensional information has led to the adoption of JSCC technique in most extant semantic designing strategies. The current implementation of module architectures, similar to the classical Shannon paradigm, presents several issues. As the level of interest in this particular field continues to increase, there is a concerted effort being made to address and surmount the challenges associated with it. Hence, the semantic scheme based on JSCC emerges not only as a highly viable contender for next-generation 6G communication systems but also as a promising transit candidate for the optimal goal of semantic communication. The partnership between JSCC and deep learning in [9] reveals that the deep learning encoder and decoder, as presented, exhibit superior performance in terms of word error rate compared to the conventional technique, particularly when the computational resource allocated for syllable encoding is limited. A limitation of this approach is the utilization of a predetermined bit length for encoding words of varying lengths. In [10], a performance comparison of deep and JSCC-based semantic communication [9] to present the potential advantage and the design strategies has concluded the difference between semantic communication and traditional communication as follows:
There exist various domains of processing information. The first phase of SC involves the manipulation of information within the semantic realm, whereas the conventional focuses on compressing information within the realm of entropy.
Traditional communication methods prioritize the precise retrieval of information, whereas semantic communication systems are designed to facilitate decision-making or the achievement of specific transmission objectives.
Conventional systems primarily focus on designing and optimizing the information transmission modules found in standard transceivers. In contrast, semantic communication systems take a holistic approach by jointly designing the entire information processing framework, spanning from the source information to the ultimate goals of applications.
In Fig. 1, we summarize the design overview of the semantic aspect in [2] from semantic theory encircled semantic channel capacity and future communication channels and datasets, including text, audio, image, video, unmanned aerial vehicles (UAVs), and the Internet of Things (IoT), are based on semantic theory. In order to calculate the source semantic entropy, a model measurement is required. The semantic quantification determines the semantic compression and signal processing. Furthermore, the semantic framework and hierarchy are illustrated in Fig. 1. To effectively enable task execution, the effectiveness on the top of the design architecture is based on task or goal-oriented methodologies. Following the basics of theory and dataset, the second level brings the semantic aspect to an evolution of future communication. For the bottom level, this layer via physical transmission integrates with the semantic level and presents JSCC encoder/decoder regarding synchronization, precoding, antennas, and power amplify.
![FIGURE 1. - The design overview for semantic communications in [2].](/mediastore/IEEE/content/media/6287639/10380310/10416847/sham1-3360003-large.gif)
B. Task-Oriented Signal Compression
The tremendous growth seen in telecommunication technologies is yet to follow the goal of reliable, fast, and low latency communications or to improve the capacity at which a communication network can serve users. While the idea of reliable communication appears to be a very intuitive and obvious requirement for communication systems, it is arguably a human need. Be it communication for a voice call, download of a photo, or streaming a video, we always prefer to receive our desired content at the best perceivable quality. This need, however, is no longer in place when designing machine-to-machine communications. The machine-to-machine communications [1] occur since this can help the receiver to make more informed decisions [11], [12], [13] or more precise estimates or computations [14] or both [15]. Naturally, in this context, there is no need for the reliability of communications to be beyond serving the specific needs of the control, estimation, or computational task at hand. This calls for a fresh examination into the design of communication systems that have been engineered with reliability as one of their ultimate goals [16]. The emerging literature regarding SC [17] as well as goal/task-oriented communications [11] is attempting to take the first steps towards the above-mentioned goal, i.e., incorporating these semantics [18], [19], together with the goal of message exchange, into the design of communication systems. The ever-increasing growth of machine-to-machine communications is the major motivating factor behind the accelerating research interest in the task-oriented design of communications. As IoT networks and cloud-based applications become more commercialized, autonomous vehicles/UAVs become more mature, and industry 4.0 approaches maturity, a boom in machine-to-machine communications is fueled. To emulate a cyber-physical system composed of several inter-dependant devices or machines, this paper considers the mathematical framework of a generalized decentralized partially observable Markov decision process (Dec-POMDP). There is a significant body of literature behind the theoretical advancements for solving generalized forms of Markov Decision Processes [20], and their applications in telecommunication and cyber-physical systems [12], [21]. Their work departs from the literature on the instance design of the observation function for each agent in the Dec-POMDP. The challenge of jointly developing the observation function and control strategy for each agent in Dec-POMDPs was investigated in [11]. It is important to note that the agents’ observations come from a fundamental Markov decision process (MDP). While in classical Dec-POMDP problems [20], the observation function is considered to be a single fixed function, the framework in [11] offers more flexibility in designing the control policies for a multi-agent system. This approach specifically permits a restricted joint design of the observation and control policy, which is summed up as follows:
The bit-budgets for the inter-agent communication channels are respected;
The observation functions filter any non-useful observation information for each agent;
The removal of non-useful observation information by the observation functions is carried out such as minimization of any loss on the average return from multi-agent systems (MAS’s) due to bit-budgeted inter-agent communications.
Reduces the complexity of the clustering algorithm by transforming it from multi-dimensional observations to the one-dimensional output space of the value functions,
The observation points are linearly separable when being clustered according to the generalized data quantization problem
The effectiveness of the data for the task is considered for goal-oriented quantization.
The value of the observations begins to grow as the ultimate target of the task at hand becomes closer.
Selected reference of the machine-to-machine communications for task-oriented design.
C. Joint Source Channel Coding and its Prototype
In accordance with Shannon’s separation theorem, a typical cascade structure that improves the performance of source coding and channel coding independently may maintain the entire system at its optimum [6], such as the majority of contemporary systems for wireless image transmission, which compresses the picture using a source coding method (e.g., JPEG, WebP, BPG) before encoding the bit stream with a source-independent channel code (e.g., LDPC, Polar, Turbo, BCH etc.). Nevertheless, the theory is based on certain premise conditions, such as an unlimited code length, point-to-point transmission system, memory-less stationary source, etc. Since these requirements are seldom satisfied in real-world applications, these tandem coding schemes are often suboptimal, such as autonomous driving and the Internet of Things (IoT) that enforce low latency real-time communication and/or low computation complexity implementation. In addition, if the channel quality goes below a specific level, the channel coding may not offer enough error corrections, and the source coding will inevitably collapse catastrophically. Consequently, jointly optimizing source coding and channel coding for relatively short messages, also known as JSCC, gradually becomes advantageous and garners a great deal of interest.
JSCC was initially conceptualized more than four decades ago [36]. This has been explored further since the 1990s [37], [38] [39]. An iterative joint source-channel decoding algorithm was then proposed in the subsequent works, such as [40], and it was verified that these structures produce a large coding gain over separate coding in finite block-length transmission.
The article [41] presented a novel JSCC scheme in which double LDPC codes [42] were applied as the source and channel codes. It was also found that for fixed-size blocks, the source (encoded or unencoded) is redundant, and this redundancy may be exploited on the decoder side. For instance, the channel encoder uses information from the source to lower its frame error rate (FER) while maintaining a very low signal-to-noise ratio (SNR).
Image transmission and reception in a JSCC system were soon proposed as a feasible application. The authors of [43] explored the possibility of transmitting images in a JSCC system through a deep-space communication channel. The authors of [44] proposed a joint source-channel coding scheme using BCH codes in a binary symmetric channel (BSC), and it reduced the distortion of satellite images better than a classical tandem source-channel coding scheme based on BCH codes.
In [45] and [46], the authors optimized the JSCC system or proposed new codes to enhance systemic performance, such as the BER. The research on JSCC systems has become more popular [47], [48], [49], [50], [51], [52], but the feasibility of implementing a JSCC system at the circuit level is rarely mentioned. As a sub-class of LDPC codes [42], [53], [54], [55], [56], QC-LDPC, where the parity check matrix is composed of permutation matrices (CPMs), can contribute to effective partial-parallel processing, due to the regularity of their parity check matrices H. For telecommunication, there have been comprehensive studies [57], [58], [59] on improving the complexity and accuracy of the LDPC decoders. They applied appropriate calculations [60], [61], [62] and various structures of LDPC codes [63], [64], [65], [66]. Other than communications, LDPC code decoders have also been popular in other areas such as storage [67], [68], [69] or biometric systems [70], [71], [72]. The prototype of the JSCC decoder can be remarked on and summarized in [30] that trace back to the early 20th century the initiative of JSCC implemented in Variable Length Error Correction (VLEC) code [73] exhibits a considerable level of intricacy due to the utilization of a vast alphabet for the selection of encoded sources. With the advent of the UEC-URC code [74], Unary Error Correction (UEC) code combined with a Unity Rate Convolutional (URC) code. It provides a nearly optimal performance at a reduced computational cost. This holds true even when dealing with extensive encoded sources, such as those seen in source coding. Although the Log-BCJR algorithm used in UEC-URC decoding is hindered by the presence of sequential information dependencies, which negatively impact processing latency, a solution known as the Fully Parallel Turbo Decoder (FPTD) [75] overcomes these limitations. By eliminating the aforementioned dependencies inherent in the traditional Log-BCJR approach, the FPTD attains the performance of the first high throughput near-capacity JSCC decoder and its prototype architecture [30]. Furthermore, the Decode-or-Compress-and-Forward (DoCF) [29] is proposed for the mechanism that the JSCC decoding step is responsible for processing the demodulated sequences from a relay, followed by an additional stage to retrieve the original message from the source, taking into account the fluctuating circumstances of the channel. The method of the decoder at the point of arrival involves a two-phase decoding procedure that utilizes the standard BCJR algorithm. The experimental verification of the proposed scheme’s superiority is conducted by implementing it on Software-Defined Radios (SDRs), with a focus on addressing system-level difficulties in provisioning. The analog circuit prototype [31], [32], [33] of JSCC can get a compression technique for lightweight devices and the performance of the system was assessed through the utilization of SPICE simulations, in addition to the construction and testing of PCB prototypes. The variant of deep learning-based JSCC (DJSCC) [76], [77], [78], [79] is to attain notable levels of reliability despite constraints such as restricted resources and poor SNRs. The reconstruction work involves the utilization of DJSCC, which may be seen as a variant of an auto-encoder. In DJSCC, the SNRs are introduced into the intermediary segment of the encoder, resulting in a distorted version of the original auto-encoder. The DJSCC encoder employs semantic signals or analog waveform sequences instead of digital transmission, unlike traditional wireless transmission. This enables its operation in challenging channel circumstances while still achieving satisfactory restoration. Nevertheless, the efficiency enhancement obtained by current DJSCC techniques is only discernible through simulations in academic studies, which serve as the knowledge sharing by auxiliary transmission for semantic networks [5]. These simulations typically assume ideal synchronization, precoding, antenna, and power amplifier conditions. Therefore, the study [34] focuses on an SDR-based DJSCC platform. The capability of this system is evaluated by taking into account two important factors: synchronization error and non-linear distortion. Moreover, drawing inspiration from the impressive resilience demonstrated by Vision Transformers (ViTs) [80] in effectively addressing various challenges associated with picture nuisances, the authors in [35] provide a novel approach that utilizes a ViT-based framework for the purpose of SC. The methodology employed in our study demonstrates an increase in peak signal-to-noise ratio (PSNR) by a satisfactory level when compared to several variations of convolutional neural networks. Finally, we summarize the state-of-the-art prototype of JSCC and DJSCC in Table 1. Our work demonstrates the inclusive prototype of an edge semantic device to inspire further advanced hardware design for JSCC-based semantic task-oriented communication.
Proposed Prototype of JSCC System for Semantic Feature Transmission and Reception
A. Overview
The proposed JSCC system is designed to accommodate various types of semantic communications and task-oriented communications through the application of deep learning techniques. A deep learning model recovers the transmitted data with the semantic or task-oriented feature in accordance with side information derived from knowledge-based and task-effectiveness metrics. As depicted in Fig. 3, the architecture employs a specialized semantic encoder to transform the raw data into a source sequence based on knowledge-based side information, denoted as
B. QC-LDPC Codes Construction and Encoding
The JSCC system outlined in this paper is designed for implementation on hardware such as FPGAs, requiring a shift from floating-point to fixed-point arithmetic. This change inevitably leads to a predictable decrease in computational accuracy. To counteract this performance loss, additional parity bits are incorporated into the JSCC QC-LDPC encoder. This adjustment ensures that the system maintains a reasonable performance level when deployed on hardware. Furthermore, in our previous work [81], we provide a UEP LDPC code construction that shows the best irregular node distribution in accordance with the design of the large-degree variable node, which has a better capability for error correction than the small-degree variable node. From a hardware perspective, a variable node unit with a larger degree and stronger error-correcting capabilities has a higher level of computational complexity than a smaller degree. In a tanner graph, the degree of the check/variable node can be represented as its connecting edge. The authors derive the theoretical analysis to demonstrate the decoding capability of UEP LDPC code. This result can be applied to UEP QC-LDPC code design for semantic communication and apply the following QC-LDPC code construction by determining variable node degree to achieve UEP for the JSCC system.
The construction of QC-LDPC codes is twofold based on Protograph LDPC (P-LDPC) codes and CPMs replacement for the quasi-cyclic characteristic. Our aim is to construct a QC-LDPC matrix based on 
\begin{align*} \mathbf {H}_{50\times 90} = \begin{pmatrix} \mathbf {H}_{s(20\times 40)} & \mathbf {H}_{L(20\times 50)}\\ \mathbf {0}_{s(30\times 40)} & \mathbf {H}_{c(30\times 50)} \end{pmatrix} \tag{1}\end{align*}
The next step is to replace traditional LDPC codes with QC-LDPC codes. The parity check matrix of the QC version of the selected P-LDPC codes can be obtained by replacing “1”s with CPMs of appropriate shift values and “0”s with all-zero matrices based on
By lifting \begin{align*} \mathbf {H}_{QC(50\times 90)} = \begin{pmatrix} \mathbf {H}_{sQC(20\times 40)} & \mathbf {H}_{LQC(20\times 50)}\\ \mathbf {0}_{sQC(30\times 40)} & \mathbf {H}_{cQC(30\times 50)} \end{pmatrix} \tag{2}\end{align*}
Referring to Eq. 2, the parity check matrices for the JSCC encoding are indicated as
The encoding method can be simply achieved by \begin{equation*} \mathbf {c} = \mathbf {G}^{T}\mathbf {s} \tag{3}\end{equation*}
Source Compression: The first step can be regarded as compressing data, which is given by \begin{equation*} \mathbf {b} = \mathbf {H_{sQC}}. \tag{4}\end{equation*}
Based on the size of parity check matrices, the source information vector, denoted as
Channel Parity-bit Generation: the subsequent manner is based on the property of LDPC parity check matrix in Eq. 5:\begin{equation*} \mathbf {H}_{QC} \: \mathbf {c} = [\mathbf {H_{1(4800\times 4800)})} \: \mathbf {H_{2(4800\times 3200)}}] \: \mathbf {c} = 0. \tag{5}\end{equation*}
As codeword \begin{equation*} \mathbf {c} = [\mathbf {p}\: \mathbf {b}]^{T} \tag{6}\end{equation*}
\begin{equation*} \mathbf {p} = \mathbf {H_{1(4800\times 4800)}^{-1}} \mathbf {H_{2(4800\times 3200)}} \mathbf {b}. \tag{7}\end{equation*}
The second half of the proposed JSCC encoder, which behaves like a channel encoder and is fed with a 3200-bit long
It is noted that all data involved in the JSCC encoding scheme are represented as binaries, corresponding operations depicted in the aforementioned equations are merely bitwise operators (ANDs and XORs), which are hardware-friendly to implement.
C. JSCC Layered Decoding Algorithm
The JSCC decoder using QC-LDPC codes can be visualized through a Tanner graph. This graph is essentially divided into two interconnected subgraphs: one representing the source and the other representing the channel, depicted in Fig. 4. As the major component of the proposed system, the QC-LDPC code-based JSCC decoder can be achieved by applying a QC-LDPC layered sum-product decoding algorithm [62], [86], although some new calculations, such as message exchanges between the source side and the channel side, need to be included. The partial parallelism of the layered decoding algorithm, which enables full parallel operations amongst all sub-matrices within one check node or “layer”, simplifies the hardware implementation of the source and channel decoders. The following parameters and values should be defined and assumed before explaining the JSCC decoding process.
An AWGN channel with zero mean and variance
\delta ^{2} M_{s} M_{c} N_{s} N_{c}
The key decoding procedures for the source and channel are as follows.
Variable node processor (VNP) in source decoding: the variable node to check node (V2C) messages,
\beta _{jk}^{sc} L_{sc}^{k} k M(k)\setminus j k j \begin{equation*} \beta _{jk}^{sc} = L^{sc}_{k}+\sum _{j'\in M(k)\setminus j}^{}\alpha _{jk}^{sc} \;\;\;\;\;\;\;\;\;\forall k\in N(j) \tag{8}\end{equation*} View Source\begin{equation*} \beta _{jk}^{sc} = L^{sc}_{k}+\sum _{j'\in M(k)\setminus j}^{}\alpha _{jk}^{sc} \;\;\;\;\;\;\;\;\;\forall k\in N(j) \tag{8}\end{equation*}
Check node processor (CNP) in source decoding: Two operations are performed in two steps.
Update the check node to variable node (C2V) messages
\alpha _{jk}^{sc} Updating message from the source decoder to the channel decoder,
I_{\hat {k}}^{sc\_{}cc} \begin{align*} tanh(\alpha _{jk}^{sc}/2) &= tanh(I^{cc\_{}sc}_{\hat {k}}/2) \times \prod _{k'\in N(j)\setminus k}^{}tanh(\beta _{jk}^{sc}/2) \\ &\quad \forall k\in N(j) \tag{9} \\& \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! tanh(I_{\hat {k}}^{sc\_{}cc}/2) = \prod _{k'\in N(j)}^{}tanh(\beta _{j{k}'}^{sc}/2) \tag{10}\end{align*} View Source\begin{align*} tanh(\alpha _{jk}^{sc}/2) &= tanh(I^{cc\_{}sc}_{\hat {k}}/2) \times \prod _{k'\in N(j)\setminus k}^{}tanh(\beta _{jk}^{sc}/2) \\ &\quad \forall k\in N(j) \tag{9} \\& \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! tanh(I_{\hat {k}}^{sc\_{}cc}/2) = \prod _{k'\in N(j)}^{}tanh(\beta _{j{k}'}^{sc}/2) \tag{10}\end{align*}
Calculating Posteriori LLR \begin{equation*} l_{k}^{sc}=L_{k}^{sc}+\sum _{j'\in M(k)}^{}\alpha _{j'k}^{sc}. \tag{11}\end{equation*}
For channel decoding, the procedure for updating LLR messages is very similar to source decoding, as depicted on the right side of Fig. 4.
VNP in channel decoding: owing to the crossed message-passing mechanism between the source decoder and the channel decoder, VNP in channel decoding must be computed separately.
\widetilde {N}^{cc\_{}sc} \begin{equation*} \beta _{jk}^{cc} = L_{k}^{cc} + \sum _{j'\in M(k)\setminus j}^{}\alpha _{j'k}^{cc} \;\;\;\;\;\;\;\;\;\forall k\in N(j)\cap \widetilde {N}^{cc\_{}sc} \tag{12}\end{equation*} View Source\begin{equation*} \beta _{jk}^{cc} = L_{k}^{cc} + \sum _{j'\in M(k)\setminus j}^{}\alpha _{j'k}^{cc} \;\;\;\;\;\;\;\;\;\forall k\in N(j)\cap \widetilde {N}^{cc\_{}sc} \tag{12}\end{equation*}
In Eq. (12),
\widetilde {N}^{cc\_{}sc} N^{cc\_{}sc} \begin{align*} \beta _{jk}^{cc} = I_{\hat {k}}^{sc\_{}cc} + L_{k}^{cc} + \sum _{j'\in M(k)\setminus j}^{}\alpha _{j'k}^{cc} \;\;\;\;\;\;\;\forall k\in N(j)\cap N^{cc\_{}sc} \\{}\tag{13}\end{align*} View Source\begin{align*} \beta _{jk}^{cc} = I_{\hat {k}}^{sc\_{}cc} + L_{k}^{cc} + \sum _{j'\in M(k)\setminus j}^{}\alpha _{j'k}^{cc} \;\;\;\;\;\;\;\forall k\in N(j)\cap N^{cc\_{}sc} \\{}\tag{13}\end{align*}
CNP in the channel decoding:
\begin{equation*} tanh(\alpha _{jk}^{cc}/2) = \prod _{k' \in N(j)\setminus k}^{}tanh(\beta _{jk}^{cc}/2)\;\;\;\;\;\;\;\;\;\forall k\in N(j) \tag{14}\end{equation*} View Source\begin{equation*} tanh(\alpha _{jk}^{cc}/2) = \prod _{k' \in N(j)\setminus k}^{}tanh(\beta _{jk}^{cc}/2)\;\;\;\;\;\;\;\;\;\forall k\in N(j) \tag{14}\end{equation*}
Message from the channel decoder to the source decoder:
\begin{equation*} I_{\hat {k}}^{cc\_{}sc} = L_{\hat {k}}^{cc} + \sum _{j' \in M(\hat {k})}^{}\alpha _{j'\hat {k}}^{cc}. \tag{15}\end{equation*} View Source\begin{equation*} I_{\hat {k}}^{cc\_{}sc} = L_{\hat {k}}^{cc} + \sum _{j' \in M(\hat {k})}^{}\alpha _{j'\hat {k}}^{cc}. \tag{15}\end{equation*}
Posteriori LLR update without VNs connected with CNs in the source decoder:
\begin{equation*} l_{k}^{cc}=L_{K}^{cc} + \sum _{j'\in M(k)}^{} \alpha _{j'k}^{cc} \;\;\;\;\;\;\;\;\;\forall k\in N(j)\cap \widetilde {N}^{cc\_{}sc} \tag{16}\end{equation*} View Source\begin{equation*} l_{k}^{cc}=L_{K}^{cc} + \sum _{j'\in M(k)}^{} \alpha _{j'k}^{cc} \;\;\;\;\;\;\;\;\;\forall k\in N(j)\cap \widetilde {N}^{cc\_{}sc} \tag{16}\end{equation*}
A Posteriori LLR update with only VNs connected with CNs in the source decoder:
\begin{equation*} l_{\hat {k}}^{cc}=I_{\hat {k}}^{cc}+L_{\hat {k}}^{sc\_{}cc}+\sum _{j' \in M(\hat {k})}^{}{\alpha _{j'\hat {k}}^{cc}}\;\;\;\; \forall k\in N(j)\cap N^{cc\_{}sc} \tag{17}\end{equation*} View Source\begin{equation*} l_{\hat {k}}^{cc}=I_{\hat {k}}^{cc}+L_{\hat {k}}^{sc\_{}cc}+\sum _{j' \in M(\hat {k})}^{}{\alpha _{j'\hat {k}}^{cc}}\;\;\;\; \forall k\in N(j)\cap N^{cc\_{}sc} \tag{17}\end{equation*}
The last step is the hard decision and the stopping criterion for the decoding iteration. Let us consider the decoded source-side sequence and channel-side sequence with
\hat {c}_{k} l_{k}^{cc} \geq0 \hat {c}_{k} 1~\forall k \hat {s}_{k} l_{k}^{sc} \geq0 \hat {s}_{k} 1~\forall k
Stopping the decoding process is subject to if the decoder reaches the maximum number of decoding iterations, which is preset before decoding.
It should be emphasized that in our case, the number of layers is:
The proposed architecture shown in Fig. 5 is different from a normal layered LDPC decoding architecture, as
A brief design architecture of the QC-LDPC decoder implementation for either the Source or Channel side.
D. Interleaver and De-Interleaver for UEP Installation
The interleaving technique is used to spread burst errors and average the distribution of “0”s and “1”s in the source vector. Especially for semantic feature images, the importance is likely to be centered and continuous. Interleaving the information in these images can improve the error correction capability of LDPC codes. To recover the order of each binary on the decoding side, the de-interleaver is applied. In [87], the proposed regular interleaver and de-interleaver are shown in Fig. 6 with the assumption of semantic importance having a higher occurrence probability on the even position using the signal processing technique or semantic model. This technique can be designed for the allocation of semantic importance to invulnerable or stronger error-correction capability as illustrated in the red codeword segment in contrast to the vulnerable codeword in the green segment. Hence, the semantic encoder/decoder possesses information on the distribution of UEP in the transmitted coded sequence. Consequently, it may achieve enhanced semantic compression while semantic importance is protected by an invulnerable codeword segment.
![FIGURE 6. - The proposed regular interleaver in [87] for UEP intallation.](/mediastore/IEEE/content/media/6287639/10380310/10416847/sham6-3360003-large.gif)
Performance Evaluation
A. Experimental Platform
The experimental setup utilizes a Virtex Ultrascale + FPGA VCU118 evaluation kit as the receiving platform, shown in Fig. 7. The system’s processor is built around an open-source 32-bit RISC-V CPU core, deriving from [89] and connecting to a flexible interconnection bus. Essential modules such as DMA controllers, NAND flash controllers, and main memory controllers for DDR RAMs are integrated, along with an Ethernet IP. The core components of the system include a specialized JSCC decoder based on QC-LDPC codes and a deep learning accelerator for semantic and task-oriented processes. This accelerator is designed to handle matrix multiplication, convolution, and activation functions, which are fundamental to neural network computations based on YOLO [90] and auto-encoder. To simulate transmission modules (in the first row depicted in Fig. 3) and model an AWGN channel, a separate computer is employed. Consequently, the combined computer-FPGA setup functions as two distinct JSCC systems. The first system encodes semantic feature images and transmits them as BPSK-modulated data through an AWGN channel. The second system receives these image data from the channel and reconstructs the original image information using demodulation and JSCC decoding techniques.
B. Implementation Results
A 6-bit quantization scheme denoted as “Proposed-Q6” in Fig. 3, is adopted in the proposed JSCC system. The synthesized results comparing resource utilization with the state-of-the-art in [30] are shown in Table 2. It should be noted that the logic (FFs and LUTs) to implement the decoder occupies around 33% of the entire FPGA and around 1.4% of block memories. The resource utilization of Intel Altera, employing ALUT, is demonstrated. The registers comprise logic components and adaptive logic modules, with each module housing two ALUTs. For 50 decoding iterations, 31 microseconds are spent by the FPGA driven by a 100 MHz clock. In Table 3, the feature comparison reveals that our FPGA platform for JSCC semantic communication encompasses both floating point analysis and fixed point (6)-bit) quantization. The resource consumption, when compared to the state-of-the-art technique, demonstrates a suitable cost to achieve improved BER performance. Additionally, the inclusion of supplemental parity bits enhances the BER performance when employing a code rate of 0.8 to compensate for the quantization loss, in comparison to the other platforms utilizing a code rate of 1.
Owing to the distinct architecture of the proposed scheme and the dimensions of the LDPC matrix, a direct performance comparison with the existing literature is not feasible. As a result, a simulated system employing 32-bit floating-point precision, labeled as “Proposed-FP32” in Table 3, serves as a benchmark against the 6-bit hardware implementation and six additional relevant studies. As illustrated in Fig. 8, the Bit Error Rates (BER) for the proposed codes, both under floating-point simulation and 6-bit hardware implementation, were scrutinized in juxtaposition with other works that solely investigated JSCC systems in a simulated setting, as indicated in the final column of Table 3. The BER is plotted on the y-axis, while the
As the proposed UEP QC-LDPC codes in this design are optimized on the condition that
Conclusion and Future Directions
By facilitating the exchange of highly informational, up-to-date, and efficient data. SC has the potential to enhance the effective use of resources, improve information accuracy and efficacy in task completion, and serve as a model and technical foundation for future generations of communication systems. The utilization of task-oriented communication has been widely regarded as a novel approach in the development of communication methods for multi-agent systems. In this paper, a novel JSCC system based on QC-LDPC codes is proposed as a promising candidate for semantic task-oriented communication systems. As the proposed irregular QC-LDPC code construction with the nature of UEP capability, the semantic importance can be dynamically assigned to the variable node with stronger error-correction capability by the proposed interleaver. After semantic source coding passes the information to the simple structure of the QC-LDPC codes, the JSCC system is implemented on the hardware device. Significantly, the operations of the JSCC decoder are layered and are then executed in parallel both on the source and channel sides. The fixed-point system also maintains fair BER performance compared to the simulated one. Moreover, the design with its optimized QC-LDPC codes is further investigated by compressing image data and protecting encoded data via an AWGN channel. This application is the semantic feature of image transmission and reception. In many cases, in practice, sources are transmitted uncoded, while state-of-the-art block channel encoders are used to protect the transmitted data against channel errors. If the JSCC scheme is used in such cases, the throughput can be improved by compressing the source data (
The design concepts entail the utilization of collaborative techniques with an edge AI server throughout the process of semantic transmission in order to enhance adaptability. The objective of this collaboration is to provide guidelines for configuring the JSCC decoder in the context of feature and federated learning.
Due to the nature of the applicable implementation in [93], the JSCC framework is used to interface with binarized neural networks to optimize resource utilization in edge devices with limited computational capabilities. This approach is specifically applied in the context of task-oriented communication and goal-oriented quantization while considering the application of the RISC-V low-power revolution.
Although there is a tendency for DJSCC to supplant the role of JSCC, the future possibility of collaboration with machine learning remains a focal point not only in the study of theoretical analysis but also in practical implementation. The importance of including semantic and task-oriented design aspects in the development of UEP JSCC prototype cannot be overstated, particularly in the context of edge AI for future-generation 6G communications.
In the cybersecurity aspect, the analysis of the secrecy rate transition between edge servers and devices utilizing JSCC can be explored within the framework of differential privacy [94], specifically in the presence of eavesdropping or the other assault scenario.
