0% found this document useful (0 votes)

32 views14 pages

Adaptive Modulation and Retransmission Scheme For Semantic Communication Systems

Uploaded by

Sharda Tripathi

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

32 views14 pages

Adaptive Modulation and Retransmission Scheme For Semantic Communication Systems

Uploaded by

Sharda Tripathi

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 14

This article has been accepted for publication in IEEE Transactions on Cognitive Communications and Networking.

This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCCN.2023.3315386

Adaptive Modulation and Retransmission Scheme

for Semantic Communication Systems
Huiguo Gao, Guanding Yu, and Yunlong Cai

Abstract—Traditional adaptive modulation scheme aims to Thus, semantic communication has recently emerged as
maximize the spectral efficiency by selecting the appropriate a new end-to-end communication paradigm that focuses on
modulation scheme under the premise of perfect bit-level data conveying the meaning behind a message rather than the
transmission. However, in task-oriented semantic communication
systems, imperfect transmission can still lead to a good inference accurate message bits. In the Weaver’s communication model
performance of semantic tasks due to the error correction capa- mentioned above, it works on a semantic and effective level.
bility of neural networks. In this paper, we propose novel adaptive To be more specific, in semantic communication systems, the
modulation and retransmission schemes to maximize the spectral meaning of a message is first extracted at the transmitter
efficiency while guaranteeing the performance of semantic tasks. (TX) and then sent to a receiver (RX). Due to the shared
Specifically, we first introduce the robustness verification problem
in semantic communication systems to analyze the robustness knowledge base (KB), the RX can interpret the meaning
of neural network inference. We then formulate and solve the of received messages. Since it is very difficult to extract
modulation scheme selection problem constrained by a robustness the intrinsic semantics of the source, neural networks (NNs)
probability threshold. Consequently, a novel adaptive modula- are typically utilized to automatically extract features and
tion scheme is developed to maximize the spectral efficiency complete intelligent tasks [4].
while guaranteeing the goal of semantic communication. We
also develop a retransmission scheme using existing combining
techniques to further increase the data rate under harsh channel
A. Literature Review
conditions. Extensive simulations are performed in unencoded
and encoded semantic communication systems to validate the In general, existing works on semantic communication fall
effectiveness of the proposed schemes. into two categories: data-oriented reconstruction and task-
Index Terms—Machine learning, semantic communication, oriented inference. For data-oriented reconstruction, semantic
adaptive modulation, retransmission, robustness. information is extracted from semantic sources, such as text
[5], [6], images [7]–[9], speech [10] or videos [11], before
I. I NTRODUCTION being utilized to rebuild the original data. In contrast, for task-
oriented inference, only the task-related semantic information
The aim of conventional communication systems is to
is extracted and directly applied to accomplish the task [12],
ensure flawless and high-speed data transmission, regardless
[13]. Most of these works consider a novel joint source channel
of the meaning of messages. As a result, the traditional trans-
coding (JSCC) scheme. Different from traditional separate
mission techniques are primarily developed to combat vari-
source-channel coding methods, it utilizes NN for source and
ous channel impairments. However, Weaver’s groundbreaking
channel encoding and requires joint-training under certain
work [1] asserts that communication can actually be divided
channel statistics. Existing research shows that it can improve
into three levels: technical, semantic, and effective levels, from
the transmission reliability of semantic information, especially
low to high. Traditional communication is limited to the most
at low signal-to-noise ratio (SNR) [14].
basic technical level, only considering precise bit transmission.
Though the above-mentioned semantic communication sys-
While at the intermediate semantic level and high effective
tems have shown excellent performance under specific channel
level, precise meaning and effective transmission are taken
conditions, more and more in-depth studies are still needed to
into account, respectively.
further improve the spectral efficiency of wireless transmission
On the other hand, as the data rate approaches the Shannon
over time-varying channels. Traditional adaptive modulation
limit, traditional communication generally encounters obsta-
techniques maximize spectral efficiency by selecting the max-
cles to further increase the transmission rate. However, high
imum modulation scheme that does not exceed a certain bit
throughput becomes essential to enable emerging intelligent
error rate (BER). It can increase the average throughput and
applications, such as virtual reality (VR) and autonomous
reduce the energy consumption by adapting to fading channels
vehicles, in beyond fifth-generation (B5G) communications
compared with those systems designed for the worst channel
[2]. Additionally, with the advent of the Internet of Everything
conditions [15], [16].
(IoE), the massive amount of data exchange may cause severe
network congestion [3]. However, traditional transmission schemes ignore the mean-
ing behind transmitted bits. To resolve this issue, several
H. Gao, G. Yu, and Y. Cai are with the College of Information Science works have been proposed to investigate how to adapt the
and Electronic Engineering, Zhejiang University, Hangzhou 310027, China transmission rate according to the changing channels in se-
(e-mail: {huiguogao, yuguanding, ylcai}@zju.edu.cn). The work of Y. Cai
was supported in part by the National Natural Science Foundation of China mantic communication systems. Recently, the variable-length
under Grants 61971376, U22A2004, and 61831004. variational feature encoding (VL-VFE) has been utilized as a

Authorized licensed use limited to: BIRLA INSTITUTE OF TECHNOLOGY AND SCIENCE. Downloaded on December 27,2023 at 05:26:08 UTC from IEEE Xplore. Restrictions apply.
© 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See https://www.ieee.org/publications/rights/index.html for more information.
This article has been accepted for publication in IEEE Transactions on Cognitive Communications and Networking. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TCCN.2023.3315386

way to adjust the length of transmitted features under different even overwriting of the previously learned model. Thus, the
channel conditions [17]. Specifically, by utilizing information performance in the previous channel environment decreases
bottleneck (IB) framework, more transmitted features can be significantly [25], [26], which is known as the catastrophic
activated by dynamic NNs to combat the noise when the forgetting (CF) phenomenon. Moreover, the performance of
channel condition deteriorates. In [18], the traditional rate- transfer learning depends heavily on the choice of pre-trained
distortion theory is extended and the IB theory is also lever- models, which creates obstacles in the training and further
aged to achieve an improved performance at the same signal application.
compression degree. A similar adaptive design can be found Additionally, when faced with countless intelligent tasks,
in [19], where the policy network is trained to improve recon- traditional encoding methods can make use of larger storage
struction quality as well as decrease transmitted features using capacities provided by hard disks, memory, or graph memory.
the Gumbel-Softmax scheme. While in [20], the projected However, the JSCC scheme is constrained by resource-limited
gradient descent optimization method is utilized to shape baseband chips. To be specific, in both traditional and JSCC
the signal constellation to adapt the rate. To mimic human schemes, different NNs are required to perform different
communications, a semantic communication with memory intelligent tasks, but their operational layers differ based on the
is introduced in [21]. Two dynamic transmission methods, open systems interconnection (OSI) model. In JSCC schemes,
importance mask and consecutive mask, are proposed based NNs directly encode semantic sources into modulated signals
on the relationship between the signal length and channel and decode them at the physical layer, utilizing resource-
noise. The dynamic transmission module is optimized by limited field-programmable gate arrays (FPGAs), application-
training the lower bound of mutual information to learn to specific integrated circuits (ASICs), or other baseband chips.
identify and mask unimportant semantic elements. However, On the other hand, in traditional systems, NNs extract semantic
SNR acquisition is indirect and can be erroneous in practice. data sources into features at the application layer, which are
Therefore, the pilot and received signals instead of accurate then subjected to channel encoding and modulation at the
SNR are jointly input to the decoder to help estimate the cur- physical layer. Since the NNs work at the application layer,
rent SNR and decode the signal in [22]. To further improve the they would consume far larger storage space of hard disk,
denoising capability, an adaptive bit-rate control mechanism memory or graphic memory than baseband chips.
based on self-attention denoising has been proposed in [23]. Thus, in this paper, we exploit the robustness verification
Specifically, the policy network is introduced to determine problem (RVP) to establish the connection between wireless
the appropriate coding rate and progressive hybrid automatic communication and performance of semantic tasks. The RVP
repeat request (HARQ) is introduced to enhance transmission aims to find the maximum input perturbation that keeps the
reliability. The HARQ technique is also utilized in [24]. output of NNs unchanged. Then, the robustness of NN can be
However, it focuses on combining the semantic coding (SC) described as how constant its output is, when the input data
with conventional Reed-Solomon (RS) channel coding and is perturbed. An NN is referred to as robust if it remains its
HARQ. The SC-RS-HARQ architecture is introduced and the output unchanged given specific input perturbations.
similarity detection network called Sim32 is also explored for In our scheme, the input perturbation of the inference net-
error detection. work is caused by channel imperfections. Since the robustness
verification focuses on the inference stage of NNs, we can
establish this connection independently of the training stage
B. Motivation of NNs. There are basically two approaches to solve the RVP.
However, the above-mentioned solutions for adaptive rate The first one uses branch and bound [27], mixed integer linear
control fail to be applied to current communication systems. programming [28], or satisfiability modulo theory [29] to
Essentially, it is inherently incompatible with the current obtain an exact and optimal solution. However, these methods
3rd Generation Partnership Project (3GPP) standard. First, consume a lot of time for computation and therefore cannot
traditional source encoding methods generalize to one or more be used even for small NNs. The second approach chooses
types of data. However, the NN-based source encoding method to relax the nonlinear activation in NNs to efficiently obtain
depends on both datasets and specific learning tasks. More- suboptimal solutions, such as duality method [30], abstract
over, traditional channel encoding methods can be applied domains [31], and local Lipschitz constants analysis [32].
to any channel environment. In contrast, the current JSCC However, due to the complex structure of models in semantic
schemes in semantic communication have poor generalization communication systems, these aforementioned methods can-
capabilities. It is true that JSCC schemes are capable of not solve the corresponding robustness verification problems
adapting to similar channel models. Transfer learning can also in a limited time. For this reason, we choose to leverage
enhance the ability of JSCC schemes to generalize to different linear bound propagation [33], which can handle complex
channel models. However, in real-world scenarios, the channel architectures of NNs while obtaining near-optimal solutions.
environments can vary significantly due to different path losses Specifically, for neurons in the current layer, the method first
and unique fading patterns. The encoder and decoder trained represents their bounds as linear functions of neurons in a
in only one channel model can be hardly applied to another previous layer. Then, it propagates the bounds to the previous
channel model without retraining. Besides, once the encoder sub-layers recursively and finally obtains the output bounds as
and decoder have adapted to a new channel environment linear functions of the perturbed input embeddings. Since the
through transfer learning, it can lead to a degradation or method considers how to bound basic operations in NNs, it is

a universal framework that can tackle different NNs. A. Classification Problem

Considering the trade-off between communication and com-

C. Main Work and Contributions putation, we choose device-edge co-inference for fast edge
In this paper, we focus on the design of adaptive modu- inference as an example [34]. As shown in Fig. 1, we consider
lation and retransmission schemes to maximize the spectral an end-to-end semantic communication system composed of
efficiency while guaranteeing the performance of semantic a TX and a RX, both of which contain the NN for feature
tasks. The main novelty of our design is that both the meaning extraction and inference, respectively. To obtain these NNs, we
of data samples and channel conditions are taken into account first train an NN for certain semantic task and then split it into
in the selection of appropriate modulation scheme. To be two parts, the first part for TX and the other part for RX. Note
more specific, a robustness verification problem in semantic that the NNs in the TX and the RX are trained separately in our
communication systems is first introduced to measure the im- proposal. Without loss of generality, we take text classification
pact of channel perturbations on the performance of semantic as an example and denote the initial text as the semantic source
tasks. The robustness probability threshold is then proposed S. The TX utilizes an NN, 𝑔(·; 𝜙), to extract sample S ∈ S into
to ensure semantic task performance. After that, we formulate the feature sequence X = [x1 ; x2 ; . . . ; x 𝑁 ] ∈ R 𝑁 ×𝐷 , where R
and solve the modulation scheme selection problem to find is the field of real numbers, 𝑁 is the sequence length, and 𝐷
the maximum BER subject to the constraint that robustness is the feature dimension, i.e., 𝑔(S; 𝜙) = X. The 𝑛-th feature
probability exceeds a given threshold. By mapping from BER x𝑛 = [𝑥 𝑛1 , 𝑥 𝑛2 , . . . , 𝑥 𝑛𝐷 ] ∈ R𝐷 from X is a 𝐷-dimension vector.
to modulation scheme, we can then obtain the appropriate
modulation scheme. Finally, we design novel adaptive modula-
tion and retransmission schemes for semantic communication.
To sum up, our main contributions are summarized as follows.
• We propose the RVP for semantic communication sys-
tems to establish the connection between wireless trans-
mission and semantic task performance. Then, based on
the proposed robustness probability threshold, we formu- Fig. 1. System model.
late and solve the modulation scheme selection problem.
• We develop new adaptive transmission schemes to im-
For transmission through the wireless channel, we first
prove transmission rate while guaranteeing the perfor-
quantize each feature element 𝑥 𝑛𝑑 into a 𝐾-bit fixed-point num-
mance of semantic tasks. Compared to the existing joint-
training methods, they can generalize to a wild variety ber 𝑥 𝑛𝑑, (𝑞) , whose bit sequence is o𝑛𝑑 = [𝑜1,𝑑 2,𝑑 𝐾 ,𝑑
𝑛 , 𝑜 𝑛 , . . . , 𝑜 𝑛 ],
𝑘,𝑑
of semantic sources, dynamic channel statistics, and also where 𝑜 𝑛 = {0, 1}, ∀𝑘, 𝑑, 𝑛. The choice of quantization num-
various semantic tasks. ber 𝐾 here follows the principle of ensuring acceptable com-
• We conduct extensive simulations in both uncoded and
putational complexity and storage capacity, while quantization
coded semantic communication systems to verify the loss can be negligible. The 𝑛-th feature is then flattened into a
effectiveness of the proposed schemes. new bit sequence, o𝑛 = [o1𝑛 , o2𝑛 , . . . , o𝑛𝐷 ] = [𝑜1,1
𝑛 , . . . , 𝑜𝑛
𝐾 ,𝐷
]∈
𝐾 𝐷
R . The corresponding bit sequence of feature sequence can
be denoted as O = [o1 ; o2 ; . . . ; o 𝑁 ] ∈ R 𝑁 ×𝐾 𝐷 . Then the TX
D. Organization chooses a certain modulation scheme, such as 𝑀-PSK or 𝑀-
The rest of the paper is organized as follows. We first QAM in uncoded systems. While in coded systems, the TX
introduce the system model in Section II. Next, we introduce chooses a modulation coding scheme (MCS) to further protect
robustness probability and then formulate the modulation bit transmission. Then the bit sequence is therefore mapped to
scheme selection problem in Section III. We then solve this the channel input alphabet C = {𝑧1 , . . . , 𝑧 𝑀 } ⊂ C, where C is
problem and propose the adaptive modulation and retransmis- the field of complex numbers.
sion scheme in Section IV. After that, we show the perfor- We denote the transmitted symbol sequence for X as Z =
mance of proposed schemes to validate the effectiveness of [z1 ; z2 ; . . . ; z 𝑁 ] ∈ C 𝑁 × 𝐽 , where z𝑛 represents a 𝐽-dimension
our design in Section V. Finally, we conclude the whole paper symbol vector of x𝑛 . Due to the channel imperfection, the
in Section VI. RX receives perturbed symbols Z̃ = [z̃1 ; z̃2 ; . . . ; z̃ 𝑁 ] ∈ C 𝑁 ×𝐽 .
Then the RX decodes the received signals, reconstructs the
bit sequence as Õ, and further dequantizes it into the recon-
II. S YSTEM MODEL
structed features X̃. Finally, the RX inputs the features into
Without loss of generality, we concentrate on the classifica- the inference network 𝑓 (·; 𝜃) to get the predicted label 𝑐, ˜ i.e.,
tion task in semantic communication systems in this study. We 𝑓 ( X̃; 𝜃) = 𝑐. ˜ Note that 𝑐˜ is not necessarily equal to the ground
begin with the description of the classification problem in end- truth 𝑐 = 𝑓 (X; 𝜃) due to wireless channel imperfection, i.e.,
to-end semantic communication systems. Then, we present our path loss, shadowing, and noise. Therefore, we need to design
wireless communication model. Finally, we show the basic a transmission protocol to attain the desirable performance of
idea of the proposed adaptive modulation and retransmission semantic tasks while achieving a high transmission rate at the
schemes. same time.

B. Wireless Communication Model a retransmission. However, in practice, the RX usually keeps

As for the quantization, we first normalize the features using decoded erroneous data instead of discarding them, and com-
the maximum absolute value, via scaling them into [0, 1], and bines it with the retransmitted data blocks [39]. For simplicity,
then perform the 𝐾-bit fixed point quantization, i.e., we assume that each retransmission has the same data. The
RX utilizes the soft combining technique [40] to add up the
𝐾 −1 received bits from previous transmissions. It can be seen as
∑︁ 𝑥 𝑛𝑑 1ª 𝐾
𝑜 𝑛𝑘,𝑑 2 𝑘 = + 2 − 1 (1)
©
® . increasing SNR by adding more energy to the received signal
2 max{|𝑥 𝑛𝑑 |} 2
𝑘=0 « 𝑛,𝑑 ¬ transmission per retransmission.
The TX sends the error-free normalized element to the RX for
each sample to help the RX dequantize the signal, which can III. P ROBLEM FORMULATION
be achieved by low code rate transmission.
We assume that the RX can accurately estimate the uplink As shown above, the reconstructed features X̃ are not the
SNR [35], [36] and then feed it back to the TX for modulation same as the extracted features X due to imperfect wireless
scheme selection using transmission scheme with low code transmission. Therefore, the predicted label 𝑐˜ at the RX is
rate. The TX transmits samples sequentially frame by frame, probably not the same as the original label 𝑐. Thus, we
and the channel gain, denoted by ℎ ∈ C, remains stable during need to first justify whether the predicted labels are robust,
one frame and is identically and independently distributed and then determine the optimal modulation scheme based
(i.i.d.) between frames. Without loss of generality, we consider on the justification results. In this section, we first introduce
that both TX and RX are equipped with a single antenna the RVP to justify the robustness of the inference results in
and the RX uses the least square (LS) method for channel the semantic communication system. Then, we formulate a
equalization, i.e., modulation scheme selection problem based on the robustness
justification.
z̃𝑛 = ℎ −1 (ℎz𝑛 + n), ∀𝑛. (2)
where n is the additive noise, modeled as a complex Gaussian A. Robustness Verification Problem in Wireless Systems
random vector. It has zero mean and its correlation matrix is
Given the transmitted features X at the TX, we cannot
𝐶N (0, 𝜎 2 I), where 𝜎 2 denotes the average noise power. Note
directly obtain the perturbed features X̃ and the predicted label
that the similar problem formulation, analysis, and conclusion
𝑐˜ at the RX. However, for each transmitted signal, the TX
can be straightforwardly extended to the multi-antenna system
already knows the SNR and the selected modulation scheme,
with minor modifications.
which can be used to estimate the possible errors of the
features. Therefore, we introduce 𝑞 as the error gap between
C. Adaptive Transmission Scheme x𝑛 and x̃𝑛 . The distance between x𝑛 and x̃𝑛 can be measured
1) Modulation: As mentioned before, we adopt adaptive by the 𝑝-norm. It means that the received features x̃𝑛 satisfy
modulation scheme for improving spectral efficiency while the following constraint
keeping robust inference for the semantic task. For a given ∥ x̃𝑛 − x𝑛 ∥ 𝑝 ≤ 𝑞, ∀𝑛. (5)
modulation scheme such as PSK and QAM in uncoded
systems, we choose different approximation equations for We can then define the set P 𝑝,𝑞 = {X̃ = [x̃1 ; x̃2 ; . . . ; x̃ 𝑁 ] :
different channel conditions, as expressed by the following ∥ x̃𝑛 − x𝑛 ∥ 𝑝 ≤ 𝑞, ∀𝑛} to represent the collection of X̃ that
equations [37] satisfy the constraint (5). Without loss of generality, we utilize
√︁ 1-norm to measure the distance in the following discussion.
2 Ímax( 𝑀/4,1) (2𝑖−1) 𝜋
𝜖MPSK = max ( log2 𝑀,2) 𝑖=1 𝑄 2𝛾¯ 𝑏 log2 𝑀 sin 𝑀 , (3) It is obvious that 𝑞 is highly related to the wireless channel
Í√ √︃ condition. Given certain transmitted samples, the number of
4 3 𝛾¯ 𝑏 log2 𝑀 quantization bits, and the feature dimension, the distribution
√1
𝑀/2
𝜖MQAM = log2 𝑀 1− 𝑖=1 𝑄 (2𝑖 − 1) ( 𝑀 −1) 𝑁0 , (4)
𝑀
of ∥ x̃𝑛 − x𝑛 ∥ 1 is only related to BER 𝜖, following the normal
|ℎ| 2 distribution whose mean is E 𝜖 and variance is Var 𝜖 , i.e., ∥ x̃𝑛 −
where 𝛾¯ 𝑏 is the average energy per bit, i.e., 𝛾¯ 𝑏 = 2 . x𝑛 ∥ 1 ∼ N (E 𝜖 , Var 𝜖 ), where
𝜎
In coded systems, MCS is utilized to map the bit sequence
𝐷 max{|𝑥 𝑛𝑑 |}
into transmitted signals in wireless channels. MCSs define 𝑛,𝑑
E𝜖 = E(| 𝑔˜ 𝑛𝑑 − 𝑔𝑛𝑑 |), (6)
various modulation and code rate combinations depending on 2𝐾 − 1
the wireless link quality. Given specific coding method, we
𝐷 (max{|𝑥 𝑛𝑑 |}) 2
can analyze the noise threshold of each MCS, under which 𝑛,𝑑
Var 𝜖 = Var(| 𝑔˜ 𝑛𝑑 − 𝑔𝑛𝑑 |). (7)
signals can be perfectly recovered. In addition, by testing the (2𝐾 − 1) 2
BER performance, we can obtain the correlation between MCS
The detailed proof can be found in Appendix A.
and BER for different SNRs [38].
If the predicted labels of all possible features are the same,
2) Retransmission: The retransmission scheme ensures re-
we refer to the predicted label 𝑐˜ as robust given 1-norm input
liable wireless transmission by resending data that failed to
space by (5), i.e.,
reach the destination. In some simple retransmission tech-
niques, the RX discards erroneously received data and requests 𝛿 = 𝑓𝑐 ( X̃; 𝜃) − max{ 𝑓1−𝑐 ( X̃; 𝜃)} > 0, ∀X̃ ∈ P1,𝑞 , (8)

where 𝑓𝑐 is the classification score for class 𝑐 while 𝑓1−𝑐 is 𝑠.𝑡. 𝑝 𝑟 (𝜖) ≥ 𝑝 𝑟𝑡 , ∀𝜖 ≤ 𝜖 𝑈 . (11b)
the one for any other class that is not 𝑐. If 𝛿 > 0, it indicates
Given the mapping relationship from modulation 𝑀 to BER
that the classification score for class 𝑐 is larger than the others
𝜖, we can further acquire the optimal modulation scheme. In
and the sample is classified into class 𝑐.
the next section, we will discuss how to solve this problem in
In semantic communication systems, the TX can obtain the
detail.
feedback about the uplink channel’s SNR and then determine
the modulation scheme or MCS for the transmission. Besides,
the error tolerance of the inference model is monotonic. It IV. P ROPOSED ADAPTIVE MODULATION AND
implies that the inference model is also robust to fewer errors RETRANSMISSION SCHEME
if it is robust to a certain number of errors. Based on the In this section, we propose an algorithm to solve the modu-
statistical property and monotonic error tolerance, we propose lation scheme selection problem. Then, based on the proposed
a new metric named robustness probability to measure the algorithm, we develop an adaptive modulation scheme in
robustness of inference results of semantic communication semantic communication systems. However, when the channel
systems. We denote 𝑝 𝑟 as the robustness probability. Specif- condition becomes too bad, even selecting the lowest mod-
ically, it measures the robustness of the inference result 𝑐˜ ulation order still cannot satisfy the robustness requirement.
corresponding to the received signal X̃. To further express the Therefore, we propose a retransmission mechanism to further
formula of 𝑝 𝑟 , we define a new variable 𝑞𝑈 and formulate the enhance the robustness of semantic inference results. Finally,
following optimization problem: we discuss the generalization capability of our proposed
𝑞𝑈 = max 𝑞, (9a) schemes, which can be extended to other semantic sources
𝑞 and tasks. They can be also extended to downlink scenarios
𝑠.𝑡. 𝛿 = 𝑓𝑐 ( X̃; 𝜃) − max{ 𝑓1−𝑐 ( X̃; 𝜃)} > 0, ∀X̃ ∈ P1,𝑞 . (9b) and integrated with other source coding methods.

Due to the monotonic error tolerance, the inference result

A. Adaptive Modulation Scheme
is robust for any 𝑞 ≤ 𝑞𝑈 . Thus, we define the robustness
probability 𝑝 𝑟 as Pr(𝑞 ≤ 𝑞𝑈 ). Depending on the statistical For specific samples X, since the inference model has
property, it can be formulated as monotonic error tolerance, we can further utilize the bisection
𝑈 method to obtain 𝑞𝑈 in Problem (9). Specifically, we start
𝑞 − E𝜖 with an interval that contains 𝑞𝑈 and narrow it down by
𝑝𝑟 = 1 − 𝑄 √ . (10)
Var 𝜖 evaluating whether 𝛿 > 0 at the midpoint of the interval.
To give more analysis on 𝑝 𝑟 , we assume the case that the We select the right subinterval when 𝛿 > 0 and the left
received sample can be accurately classified during error-free subinterval when 𝛿 < 0. Then we continue the process until
transmission. As 𝑝 𝑟 decreases towards 0, it indicates that BER a desired level of accuracy is achieved. As for each searched
𝜖 increases and gets closer to 0.5. In this context, the channel 𝑞, we can utilize the linear bound propagation [33] to test
condition is too bad to classify the received sample into the whether 𝛿 > 0. Specifically, we compare the classification
correct class. Conversely, when 𝑝 𝑟 approaches 1, it suggests scores between the ground truth label and other labels to get
that BER 𝜖 decreases and becomes almost equal to zero. Then, 𝛿 as shown in Eq. (8). In Problem (11), the maximum BER
the received sample can be classified into the correct class 𝜖 𝑈 is achieved when 𝑝 𝑟 = 𝑝 𝑟𝑡 , since the robustness probability
with a high level of confidence. Note that 𝑝 𝑟 describes the 𝑝 𝑟 decreases monotonically as BER 𝜖 increases. With solved
probability. Therefore, there is still the possibility of 𝑞 ≤ 𝑞𝑈 , 𝑞𝑈 and given 𝑝 𝑟𝑡 , we can then use the bisection method
which leads to correct classification despite how low 𝑝 𝑟 is. to obtain 𝜖 𝑈 . Furthermore, since 𝜖 decreases monotonically
A specific robustness probability is demanded at the TX, as the modulation order 𝑀 decreases, we can choose the
denoted as 𝑝 𝑟𝑡 . If 𝑝 𝑟 ≥ 𝑝 𝑟𝑡 , the robustness requirement is maximum modulation order that satisfies the requirement on
satisfied. To guarantee the performance of semantic tasks, we 𝜖 𝑈 . If the minimum modulation order still cannot satisfy the
need to adapt the modulation scheme to satisfy the robustness BER requirement, we output the minimum modulation scheme
requirement1 . so as not to waste the channel usage.
For more intuitive, we show the relation between the per-
B. Modulation Scheme Selection Problem turbed samples and the sent samples in Fig. 2. As shown
in Fig. 2, 𝑞𝑈 is the distance between the sent samples X
Since we aim to find the maximum modulation order while and the decision boundary. The distance between perturbed
keeping the inference robustness of semantic tasks, we choose samples X̃ and X is determined by the BER 𝜖. If we specify a
to maximize the tolerant BER under the robustness constraint. certain robustness probability 𝑝 𝑟 , then the probability that the
When the transmitted sample is specified, the robustness perturbed samples appear in the same classification space is
probability 𝑝 𝑟 is only related to BER 𝜖, i.e., 𝑝 𝑟 = 𝑝 𝑟 (𝜖). also determined. As the robustness probability 𝑝 𝑟 increases,
Then the optimization problem can be formulated as the perturbed samples are more and more likely to appear
𝜖 𝑈 ≜ max 𝜖, (11a) in the same classification space, and thus the classification
𝜖 accuracy also increases.
1 We do not utilize the high-order modulation to directly achieve 𝑞𝑈 since However, to obtain 𝑞𝑈 , one must first possess the inference
it is heavily dependent on the performance of robustness verification methods. NN and then run a robustness verification algorithm, which is

transmit. As soon as 𝜖 min is lower than 𝜖 𝑈 , the TX transmits the

sample using the optimal modulation scheme or MCS, 𝑀opt .
After receiving the sample, the RX reconstructs the sample
and inputs it to the inference network to obtain the predicted
˜ Then, the TX starts a new transmission round.
label 𝑐.

B. Retransmission Scheme
When the robustness probability requirement cannot be
Fig. 2. The relation between the perturbed samples and the sent samples. satisfied even with the minimum modulation scheme, retrans-
mission is needed to enhance the robustness of the semantic
inference results. As mentioned in Section II.B, the channel
unacceptable for the TX with limited capabilities on storage condition is i.i.d. between frames. We can use the soft-bit
and computation. To tackle this problem, we introduce a much maximal-ratio combining (SBMRC) technique to exploit the
smaller NN 𝐼 (·; W) by mapping X to approximate 𝑞𝑈 , i.e., channel diversity [40]. To be more specific, if sample X
𝑞𝑈 ≈ 𝐼 (X; W). To train 𝐼 (·; W), we first conduct robustness is transmitted for the 𝑇 rounds, all 𝑇 copies of received
verification on the historical data to build the dataset and X̃(𝑡 𝑖 ), 𝑖 = {1, . . . , 𝑇 } can be combined as X̃(𝑇) by the SBMRC
then divide the dataset into a training set and a validation technique at the RX. In uncoded systems, the effective SNR
set. As for sequence features extracted from text or audio, of X̃(𝑇) can be formulated as
we design 𝐼 (·; W) as the sequence-to-sequence structure like 𝑇
∑︁
transformers, recurrent neural networks (RNNs) and pointer 𝛾(𝑇) = 𝑑 2𝑀𝑖 𝛾(𝑡𝑖 ), (12)
networks. While for other non-sequence features extracted 𝑖=1
from images, we can utilize deep neural networks (DNNs) or
convolutional neural networks (CNNs). To control the size of √︄ round, 𝛾(𝑡𝑖 ) is
where 𝑀𝑖 is the modulation order of the 𝑖-th
3 log2 (𝑀𝑖 )
𝐼 (·; W), we can adjust the hidden dimensions to accommodate the SNR of the 𝑖-th round, and 𝑑 𝑀𝑖 = . The
the device requirement. Finally, we choose the mean squared 2(𝑀𝑖 − 1)
error loss (MSELoss), mean absolute error loss (MAELoss), corresponding BER can be bounded by
or Huber loss as the loss function to minimize the gap between
√︁ √︁
𝜏𝑄( 2𝛾(𝑇)) < 𝜖 (𝑇) < 𝑄( 2𝛾(𝑇)), (13)
the predicted 𝑞𝑈 and real 𝑞𝑈 . After successful training on the
server, the well-trained 𝐼 (·; W) model can be deployed to the where 𝜏 is a constant determined by the combinations of 𝑇-
TX for utilization. After that, the TX can obtain the optimal round modulation. While in coded scenarios, the effective SNR
modulation order 𝑀opt based on the sent samples X and the and corresponding BER can also be acquired by experiments.
SNR 𝛾, as shown in Algorithm 1. In our proposed retransmission scheme, the SBMRC tech-
nique is utilized to combine multi-round received signals of
Algorithm 1 Modulation Selection Algorithm different modulation schemes. The BER decreases with the re-
1: Input: Trained parameters W, SNR 𝛾, input samples X = transmission proceeds and therefore improves the performance
[x1 ; . . . ; x 𝑁 ]. of semantic tasks. Since semantic information varies between
2: Acquire 𝑞𝑈 through 𝐼 (·; W). transmitted samples, our proposed retransmission scheme sets
different BER thresholds for different samples. It demonstrates
method to find 𝜖 that satisfies 𝑝 𝑟 =
3: Use the bisection 𝑈 𝑡
that not only channel conditions but also the meaning of
𝑈
𝑞 − E𝜖
1−𝑄 √ . data is considered when deciding the retransmission stop and
Var 𝜖
𝑈
4: Substitute 𝜖 into mapping relation between 𝑀 and 𝜖. modulation scheme selection.
5: Use the bisection method to find 𝑀opt . In order to exploit the waiting channel resource in the
adaptive modulation scheme, we choose to send the signal
using the minimum modulation scheme even it cannot satisfy
According to the proposed modulation scheme selection the robustness requirement. In this case, the optimal modu-
algorithm, we can further propose an adaptive modulation lation scheme is the minimum one. Thus, we re-design the
scheme in semantic communication systems, as shown in Fig. retransmission scheme as shown in Fig. 4 and describe as
3. Specifically, the TX initializes the connection with the RX follows. The TX first initializes the connection with the RX
and then sends a transmission signaling for new samples. before transmission. Then, the TX picks up a sample and
Then the RX feeds the current SNR 𝛾 and desired robustness sends a signaling requesting to transmit it. The RX feeds
probability threshold 𝑝 𝑟𝑡 to the TX. The desired 𝑝 𝑟𝑡 can be back the desired 𝑝 𝑟𝑡 and the TX can obtain 𝑞𝑈 and 𝜖 𝑈
determined by the accuracy requirement. The connection be- based on Algorithm 1. After the RX feeds back the SNR
tween the accuracy and 𝑝 𝑟𝑡 can be obtained from the historical information of the 𝑖-th round, 𝛾(𝑡 𝑖 ), the TX can obtain the
data. With the help of Algorithm 1, the TX gets 𝑞𝑈 and 𝜖 𝑈 . optimal modulation, 𝑀opt . Then, the TX transmits the data
Also, the TX calculates the minimum BER 𝜖 min by using the using the chosen modulation scheme or MCS. The effective
given feedback SNR and assuming the minimum modulation SNR 𝛾(𝑇) and effective BER 𝜖 (𝑇) can be further obtained.
scheme. If the minimum BER still cannot satisfy the BER Since one-time transmission may not satisfy the robust-
requirement, the TX waits for better channel conditions to ness requirement, the TX acquires the current robustness

Fig. 3. Adaptive modulation scheme in semantic communication systems.

probability 𝑝 𝑟 based on 𝜖 (𝑇) and Eq. (10). Then, the TX example, we can utilize 0-norm to measure the difference be-
checks whether 𝑝 𝑟 exceeds the threshold 𝑝 𝑟𝑡 . If 𝑝 𝑟 < 𝑝 𝑟𝑡 , tween received and sent bits. Then, Eq. (5) can be substituted
the TX continues to ask for SNR feedback and repeats the by ∥ z̃ − z∥ 0 ≤ 𝑞. After that, we can obtain the corresponding
procedure mentioned before. After the calculated 𝑝 𝑟 exceeds 𝑞𝑈 and analyze the distribution of ∥ z̃ − z∥ 0 to measure 𝑝 𝑟 in
the threshold 𝑝 𝑟𝑡 , the TX sends a stop signaling and the RX a similar way.
utilizes the SBMRC technique to combine the signals from In summary, the proposed schemes can be generalized to
all previous rounds. If the retransmission round 𝑇 exceeds the other semantic sources like images and videos, and other tasks
maximum allowed retransmission round 𝑇max , the TX stops the like regression and reconstruction. They can be also extended
retransmission and initiates a new transmission. Otherwise, the to various wireless scenarios and other source coding methods,
RX utilizes the combined signals to do inference. such as Huffman coding.

V. S IMULATION RESULTS
C. Generalization Discussion
In this section, we verify that the performance of semantic
Although we focus on the text classification task for se-
tasks can be guaranteed by properly selecting the robustness
mantic communication systems in this paper, the proposed
probability threshold. Besides, we show that the proposed
schemes can be easily extended to other semantic sources
schemes can achieve a higher data rate compared with the
and tasks. To be specific, the proposed schemes are mainly
traditional schemes given the same performance of semantic
according to the methods in [33] to first obtain the maximum
tasks.
tolerant perturbation 𝑞𝑈 and further determine the appro-
priate modulation scheme based on 𝑞𝑈 . Since linear bound
propagation is a universal robustness verification framework A. Simulation Settings
that can tackle various NNs, it can be easily applied in We use Stanford Sentiment Treebank (SST) dataset [41]
NNs that process other semantic sources like images and to execute binary sentiment analysis. It consists of 67, 349
videos. The proposed schemes can also generalize to other training samples, 872 validation samples, and 1, 892 test
semantic tasks, such as regression and reconstruction tasks. samples. Each sample is a sentence labeled with two feelings:
In classification tasks, 𝑞𝑈 is measured with regard to the positive or negative. We train a one-layer transformer using the
unchanged label. We can also define the maximum tolerant cross-entropy loss function with a hidden size of 64 for both
output perturbation and then obtain 𝑞𝑈 for the regression tasks. the embedding layer and the self-attention layers. The self-
While for reconstruction tasks, we can first transfer the quality- attention layers have four heads, and the feed-forward layers
of-service (QoS) requirements to the actual maximum output have a hidden size of 128. We divide the network into two
perturbation of NNs. Then, we can obtain 𝑞𝑈 in a similar way. parts at the embedding layer, resulting in a feature dimension
The proposed schemes can be also extended to the downlink of 𝐷 = 64. Fig. 5 shows the detailed structure of NNs at
systems by exchanging the functions on the TX and the RX. the TX and RX. To avoid the impact caused by classification
Besides, though we utilize NNs to extract features from errors, we select samples to let the TX only transmit correctly-
semantic source, the proposed schemes can be generalized to classified ones. It means that one can achieve an accuracy
other source coding methods. Taking Huffman coding as an of 100% under perfect channel conditions. For small NN

Fig. 4. Retransmission scheme in semantic communication systems.

𝐼 (·; W), we simply utilize a single-head self-attention layer system model, the total information bits of the sent sample,
and a feed forward layer with a hidden size of 8 to estimate denoted as 𝐿, equals to 𝑁 𝐷𝐾. The transmission rate can be
𝑞𝑈 2 . We choose Huber loss with parameter 𝛽 = 0.5 as loss 𝐿
correspondingly obtained as 𝑅 = . Besides, optional modu-
function since it achieves fast convergence while decreasing 𝑡
lation schemes in uncoded and coded semantic communication
the impacts of outliers.
systems are shown in Table I. The MCS in coded system
As for the wireless communication settings, we utilize the is designed with reference to the IEEE 802.11ac protocol
wireless communication model mentioned in Section II.B. [42]. We utilize low-density parity-check (LDPC) codes with
Specifically, we utilize the Rayleigh fading and the channel a block size of 64, 800 as the coding method.
gain is modeled as a complex random Gaussian vector. We
consider the quantization number 𝐾 as 10. It is an appropriate
TABLE I
value since it leads to a negligible quantization error. At O PTIONAL M ODULATION S CHEMES IN U NCODED AND C ODED S EMANTIC
the same time, it causes low computational complexity of C OMMUNICATION S YSTEMS .
encoding and decoding, and low storage space consumption.
Uncoded Coded
Note that the selection of 𝐾 does not affect the usage of our Index Modulation Index Modulation Code Rate
methods. We set the bandwidth as 4 MHz and the duration 1 QPSK 1/2
1 QPSK
time of each frame is 10 ms. At the beginning of each frame, 2 QPSK 3/4
the TX obtains the channel feedback and chooses the sample 3 16-QAM 1/2
2 16-QAM
to transmit. The transmission continues until the transmission 4 16-QAM 3/4
5 64-QAM 2/3
success. Since the channel estimation time is much smaller 3 64-QAM 6 64-QAM 3/4
than the data transmission time, we ignore this part of time 7 64-QAM 5/6
consumption. 8 256-QAM 3/4
4 256-QAM
Fig. 6 shows the transmission procedure of the proposed 9 256-QAM 5/6
adaptive modulation scheme and retransmission scheme. For
the adaptive modulation scheme, the TX waits until the chan-
nel condition becomes good enough to achieve the robustness B. Impact of Robustness Probability Threshold on Accuracy
requirement through one-time transmission. Thus, the duration
time, denoted as 𝑡, consists of waiting frames and one-time 1) Uncoded System: First, we test the performance of
transmission. While the retransmission scheme retransmits the semantic tasks, measured by the average classification ac-
sample with the minimum modulation scheme even though the curacy, and also record the average BER, as the robustness
current channel condition is too bad to achieve the robustness probability threshold, 𝑝 𝑟𝑡 , increases from 0 to 1 in the uncoded
requirement. In this way, 𝑡 includes the last 𝑇 − 1 frames and semantic communication system. The results for the proposed
the 𝑇-th retransmission time. Note that a higher modulation adaptive modulation scheme and retransmission scheme are
scheme leads to a shorter transmission time. Based on the summarized in Table II(a) and Table II(b), respectively.
From the tables, the average accuracy increases as 𝑝 𝑟𝑡 in-
2 Although this is an estimated value, our simulation results show that it creases and finally converges to 100% under all SNR settings.
has an excellent performance. Also, the average BER decreases as 𝑝 𝑟𝑡 increases for all

Fig. 5. The structure of NNs at the TX and RX.

Fig. 6. The transmission procedure of the adaptive modulation scheme and retransmission scheme.

TABLE II
T HE AVERAGE ACCURACY OF THE SEMANTIC TASK AND THE AVERAGE BER WITH DIFFERENT SNR S IN THE UNCODED SEMANTIC COMMUNICATION
SYSTEM .

(a) Adaptive modulation scheme. (b) Retransmission scheme.

SNR (dB) SNR (dB)
𝑝𝑟𝑡 𝑝𝑟𝑡
0 5 10 15 20 0 5 10 15 20
0 71.10% / 3.62E-01 76.02% / 2.40E-01 82.98% / 1.36E-01 91.34% / 6.43E-02 96.68% / 2.52E-02 0 71.10% / 3.62E-01 76.02% / 2.40E-01 82.98% / 1.36E-01 91.34% / 6.43E-02 96.68% / 2.52E-02
1E-08 81.76% / 1.30E-01 82.40% / 1.21E-01 84.62% / 9.11E-02 92.66% / 4.85E-02 96.96% / 2.00E-02 1E-09 82.38% / 1.47E-01 81.02% / 1.34E-01 85.10% / 9.62E-02 92.34% / 5.14E-02 96.70% / 2.10E-02
3E-08 82.54% / 1.21E-01 82.44% / 1.15E-01 85.48% / 8.64E-02 91.50% / 4.83E-02 96.76% / 1.99E-02 5E-09 83.66% / 1.34E-01 81.80% / 1.25E-01 84.90% / 9.22E-02 91.80% / 4.91E-02 97.04% / 2.06E-02
5E-08 83.52% / 1.17E-01 83.22% / 1.12E-01 86.60% / 8.54E-02 91.42% / 4.68E-02 97.06% / 1.91E-02 1E-08 84.12% / 1.30E-01 83.36% / 1.21E-01 85.40% / 9.15E-02 91.74% / 4.85E-02 97.04% / 1.96E-02
1E-07 84.66% / 1.12E-01 83.64% / 1.07E-01 87.10% / 8.37E-02 91.78% / 4.63E-02 96.88% / 1.90E-02 2.5E-08 85.70% / 1.23E-01 83.84% / 1.16E-01 85.76% / 8.70E-02 92.22% / 4.84E-02 96.52% / 1.91E-02
3E-07 86.24% / 1.03E-01 85.92% / 9.99E-02 87.04% / 7.92E-02 92.72% / 4.49E-02 96.78% / 1.89E-02 5E-08 86.16% / 1.18E-01 84.88% / 1.11E-01 86.36% / 8.59E-02 92.60% / 4.71E-02 97.04% / 1.90E-02
5E-07 87.58% / 9.84E-02 85.48% / 9.59E-02 87.42% / 7.70E-02 92.78% / 4.45E-02 97.32% / 1.88E-02 1E-07 86.54% / 1.13E-01 84.64% / 1.07E-01 86.68% / 8.37E-02 92.44% / 4.60E-02 97.08% / 1.87E-02
1E-06 86.74% / 9.25E-02 87.00% / 9.07E-02 88.20% / 7.45E-02 93.34% / 4.27E-02 96.98% / 1.82E-02 5E-07 88.44% / 9.93E-02 86.46% / 9.50E-02 88.36% / 7.69E-02 92.86% / 4.34E-02 97.42% / 1.72E-02
5E-06 89.64% / 7.90E-02 89.00% / 7.84E-02 89.10% / 6.76E-02 93.06% / 4.07E-02 97.10% / 1.65E-02 1E-06 88.96% / 9.36E-02 87.28% / 8.97E-02 88.38% / 7.41E-02 93.32% / 4.20E-02 97.12% / 1.71E-02
1E-05 90.06% / 7.35E-02 89.98% / 7.31E-02 91.04% / 6.34E-02 93.84% / 3.79E-02 97.44% / 1.57E-02 5E-06 90.96% / 8.09E-02 89.36% / 7.67E-02 90.00% / 6.55E-02 93.74% / 3.97E-02 97.60% / 1.66E-02
5E-05 92.40% / 6.09E-02 91.60% / 5.98E-02 91.78% / 5.41E-02 94.92% / 3.40E-02 97.94% / 1.50E-02 1E-05 90.76% / 7.58E-02 90.42% / 7.04E-02 90.22% / 6.29E-02 93.82% / 3.84E-02 97.38% / 1.63E-02
0.0001 93.64% / 5.57E-02 93.68% / 5.45E-02 92.30% / 5.02E-02 95.52% / 3.20E-02 98.10% / 1.44E-02 5E-05 92.48% / 6.06E-02 92.88% / 5.70E-02 92.28% / 5.33E-02 94.70% / 3.38E-02 97.96% / 1.51E-02
0.0003 95.10% / 4.72E-02 94.44% / 4.59E-02 94.40% / 4.33E-02 96.26% / 2.91E-02 97.76% / 1.28E-02 0.0001 93.56% / 5.52E-02 93.20% / 5.16E-02 93.48% / 4.86E-02 95.24% / 3.14E-02 97.90% / 1.38E-02
0.0005 95.06% / 4.40E-02 95.28% / 4.19E-02 94.56% / 3.98E-02 95.90% / 2.73E-02 98.48% / 1.23E-02 0.00025 93.80% / 4.80E-02 94.30% / 4.50E-02 94.60% / 4.27E-02 95.88% / 2.87E-02 98.12% / 1.31E-02
0.00075 95.42% / 4.10E-02 95.12% / 3.92E-02 95.20% / 3.71E-02 96.08% / 2.57E-02 98.32% / 1.19E-02 0.0005 94.00% / 4.33E-02 95.04% / 4.02E-02 94.98% / 3.84E-02 96.76% / 2.66E-02 98.38% / 1.24E-02
0.001 95.70% / 3.88E-02 95.30% / 3.71E-02 95.20% / 3.51E-02 96.84% / 2.45E-02 98.04% / 1.17E-02 0.00075 93.58% / 4.17E-02 95.12% / 3.77E-02 95.92% / 3.64E-02 96.84% / 2.51E-02 98.56% / 1.17E-02
0.00625 97.06% / 2.67E-02 97.78% / 2.44E-02 97.84% / 2.37E-02 98.36% / 1.78E-02 99.00% / 8.44E-03 0.001 95.36% / 3.98E-02 95.30% / 3.61E-02 96.12% / 3.41E-02 96.68% / 2.45E-02 98.64% / 1.13E-02
0.0125 97.90% / 2.25E-02 98.26% / 2.02E-02 98.08% / 1.94E-02 98.64% / 1.51E-02 99.22% / 7.53E-03 0.00625 95.94% / 3.01E-02 97.62% / 2.57E-02 97.82% / 2.34E-02 98.26% / 1.74E-02 98.98% / 8.71E-03
0.025 98.56% / 1.85E-02 98.86% / 1.63E-02 98.88% / 1.57E-02 99.22% / 1.25E-02 99.28% / 6.37E-03 0.025 96.94% / 2.23E-02 98.32% / 1.81E-02 98.64% / 1.59E-02 98.82% / 1.23E-02 99.48% / 6.36E-03
0.05 98.76% / 1.45E-02 99.04% / 1.26E-02 98.98% / 1.17E-02 99.20% / 9.80E-03 99.52% / 5.24E-03 0.05 98.00% / 1.82E-02 98.68% / 1.45E-02 99.18% / 1.22E-02 99.40% / 9.62E-03 99.66% / 5.25E-03
0.1 99.32% / 1.10E-02 99.54% / 9.05E-03 99.40% / 8.60E-03 99.64% / 7.16E-03 99.78% / 4.06E-03 0.1 98.50% / 1.39E-02 99.06% / 1.10E-02 99.58% / 8.99E-03 99.72% / 7.30E-03 99.80% / 4.00E-03
0.2 99.56% / 7.78E-03 99.74% / 6.40E-03 99.88% / 5.74E-03 99.72% / 4.98E-03 99.92% / 2.88E-03 0.2 99.36% / 1.02E-02 99.70% / 7.89E-03 99.82% / 6.23E-03 99.88% / 4.92E-03 99.78% / 2.88E-03
0.3 99.72% / 6.01E-03 99.78% / 4.57E-03 99.96% / 4.26E-03 99.94% / 3.68E-03 99.92% / 2.23E-03 0.3 99.58% / 7.88E-03 99.82% / 5.87E-03 99.94% / 4.67E-03 99.86% / 3.72E-03 99.98% / 2.24E-03
0.4 99.94% / 4.76E-03 99.98% / 3.71E-03 100.00% / 3.29E-03 99.98% / 2.88E-03 99.96% / 1.74E-03 0.4 99.82% / 6.38E-03 99.92% / 4.82E-03 99.92% / 3.74E-03 99.94% / 2.99E-03 99.98% / 1.70E-03
0.5 99.98% / 3.94E-03 100.00% / 3.02E-03 100.00% / 2.56E-03 100.00% / 2.28E-03 100.00% / 1.51E-03 0.5 99.94% / 5.18E-03 99.94% / 3.93E-03 100.00% / 2.96E-03 100.00% / 2.31E-03 100.00% / 1.42E-03
0.6 100.00% / 3.23E-03 100.00% / 2.28E-03 100.00% / 2.08E-03 100.00% / 1.84E-03 100.00% / 1.20E-03 0.6 99.98% / 4.28E-03 100.00% / 3.19E-03 100.00% / 2.37E-03 100.00% / 1.90E-03 100.00% / 1.13E-03
0.7 100.00% / 2.64E-03 100.00% / 1.93E-03 100.00% / 1.60E-03 100.00% / 1.49E-03 100.00% / 9.40E-04 0.7 100.00% / 3.61E-03 100.00% / 2.67E-03 100.00% / 1.87E-03 100.00% / 1.60E-03 100.00% / 9.60E-04
0.8 100.00% / 2.09E-03 100.00% / 1.48E-03 100.00% / 1.31E-03 100.00% / 1.09E-03 100.00% / 7.46E-04 0.8 100.00% / 2.90E-03 100.00% / 2.09E-03 100.00% / 1.50E-03 100.00% / 1.13E-03 100.00% / 7.63E-04
0.9 100.00% / 1.50E-03 100.00% / 1.06E-03 100.00% / 9.25E-04 100.00% / 7.66E-04 100.00% / 5.51E-04 0.9 100.00% / 2.10E-03 100.00% / 1.56E-03 100.00% / 1.05E-03 100.00% / 7.90E-04 100.00% / 5.53E-04

SNRs. It is because that the proposed schemes would select is 71.10% when SNR=0dB and 𝑝 𝑟𝑡 = 0. It seems relatively
lower modulation orders under the same channel conditions high since we normalize the accuracy and therefore the overall
as 𝑝 𝑟𝑡 increases. It is a clue to the fact that the perturbed accuracy is increased a lot. Also, the connection between
samples approximate to the sent ones, therefore increasing the performance of semantic tasks and wireless conditions
the classification accuracy. This result demonstrates that the highly relies on both the dataset and NN models. From
performance of semantic tasks can be guaranteed by setting our experiment results, some samples are far away from the
an appropriate 𝑝 𝑟𝑡 . We can also see that the average accuracy decision boundary of NNs and therefore can be still classified
increases with the increase of SNR when 𝑝 𝑟𝑡 equals to 0. The correctly even in harsh channel conditions. When the channel
reason is that there is no robustness probability requirement condition becomes better, the average BER becomes lower
when 𝑝 𝑟𝑡 = 0. Therefore, the TX would choose the maximum under the same modulation scheme and the gap between X̃
modulation scheme to send all samples. The average accuracy and X becomes smaller, leading to the improvement on the

average accuracy. system that the average BER decreases with 𝑝 𝑟𝑡 . Besides,
Additionally, comparing Table II(b) with Table II(a), we the average accuracy increases with 𝑝 𝑟𝑡 and increases more
can see that the average accuracy increases more flattened smoothly in the retransmission scheme. The results also show
as 𝑝 𝑟𝑡 increases with the proposed retransmission scheme that the average accuracy increases with SNR given 𝑝 𝑟𝑡 = 0.
under harsh channel conditions. The reason responsible for this The reasons for these are consistent with the discussion in the
phenomenon is that the average robustness probability under uncoded settings.
retransmission scheme is lower than that under the adaptive However, in contrast to the uncoded semantic commu-
modulation scheme. Specifically, the proposed retransmission nication system, the average accuracy is higher given the
scheme requires multiple retransmission rounds to exceed 𝑝 𝑟𝑡 same 𝑝 𝑟𝑡 and converges more rapidly to 100%. Take 𝑝 𝑟𝑡 =
while there is only one-time transmission under the adaptive 0 as an example. The average accuracy for 0dB, 5dB,
modulation scheme. Considering that the optional modulation 10dB, 15dB and 20dB are 71.10%, 76.02%, 82.98%, 91.34%
schemes are finite and the selected one must satisfy the and 96.68% for uncoded systems in Table II(a), while
robustness requirement, the average robustness probability 71.34%, 76.64%, 86.08%, 94.02% and 97.72% for coded sys-
under the retransmission scheme is closer to that under the tems in Table V(a). This is because additional error correction
actual setting 𝑝 𝑟𝑡 . coding can achieve much lower BER in the coded system than
Here, we show how a proper 𝑝 𝑟𝑡 can guarantee the accurate in the uncoded system. Even in the case of low robustness re-
classification result through imperfect transmission in a more quirements, the actual choice of MCS can achieve much lower
intuitive way. We set 𝑝 𝑟𝑡 to 0.9 under which all the samples can error probabilities than the corresponding BER requirements.
be correctly classified. We obtain the embedding representa-
tions of all tokens and then compare the received feature with C. Performance of Transmission Rate
these embeddings. Then, we select the token whose embedding
is the closest to the received embedding as the sent word. In this part, we test the performance of the proposed
We then compare the words sent and received, as well as schemes with different SNRs in both uncoded and coded se-
the maximum tolerant BER 𝜖 𝑈 under the proposed schemes, mantic communication systems. Based on the relation between
which are summarized in Tables III and IV. We bold the the average accuracy and the robustness probability threshold
inconsistencies between sent and italicize words that express shown in Table II and Table V, we then utilize the binary
strong emotions. search to find the minimum 𝑝 𝑟𝑡 that satisfies the accuracy
From Table III, we can find that there exists redundancy requirement for each SNR.
in the sent samples. That is, even if the samples cannot be
correctly recovered without errors, they can still be classified
correctly since the NN can capture semantic information
Pro osed AM (100%)
from different sentences. In addition, 𝜖 𝑈 acquired by setting 4 Proposed Retran (100%)
Traditional AM (100%)
appropriate 𝑝 𝑟𝑡 can indicate the semantics of transmitted sen- Traditional Retran (100%)
Average transmission rate (bit/s/Hz)

tences. To be specific, we find that 𝜖 𝑈 becomes smaller as Pro osed AM (95%)

Pro osed Retran (95%)
the number of words expressing positive or negative feelings 3 Traditional AM (95%)
Traditional Retran (95%)
increases. The first example only has one emotional word Pro osed AM (85%)
Pro osed Retran (85%)
‘engrossing’ and its 𝜖 𝑈 increases to 0.021 while the next Traditional AM (85%)
has ‘gorgeous’, ‘mind-blowing’, ‘breath-taking’, and ‘mess’, 2 Traditional Retran (85%)

a total of four emotional words and its tolerant BER is

0.012. Since emotional words affect the classification result
1
of sentiment analysis more significantly, our proposed scheme
inclines to allocate lower tolerant BER for those samples
full of emotional words to protect the inherent semantic 0
information. 0 5 10 15 20
From Table IV, it can be seen that the RX recovers more SNR (dB)
and more accurate words with the increase of retransmission
rounds. At the same time, more samples are classified from the Fig. 7. The average transmission rates of the proposed adaptive modulation
wrong category to the correct one. Though the final recovered and retransmission schemes compared to the baseline schemes for different
SNRs at various accuracy requirements in the uncoded semantic communica-
sentence is not the same as the sending one, it is still classified tion system.
correctly, which validates the effectiveness of 𝑝 𝑟𝑡 .
2) Coded System: We then test the connection between 1) Uncoded System: We choose the transmission schemes
the performance of semantic tasks and robustness probability with a fixed BER threshold as the traditional schemes. To be
threshold in the coded semantic communication systems. Also, more specific, the traditional adaptive modulation and retrans-
we record the average BER for each robustness probability mission schemes adopt the maximum modulation scheme to
threshold under all SNR settings. The results include the satisfy a certain BER requirement for all samples. We adjust
proposed adaptive MCS and retransmission scheme, which are the BER requirements, evaluate the average accuracy, and
summarized in Table V(a) and Table V(b), separately. From then select the corresponding BER requirement for specific
the tables, we can draw the same conclusion as in the uncoded accuracy requirements. Fig. 7 depicts the average transmission

TABLE III
C OMPARISONS B ETWEEN THE S ENDING AND R ECEIVING W ORDS UNDER A DAPTIVE M ODULATION S CHEME IN THE U NCODED S EMANTIC
C OMMUNICATION S YSTEM

Type Words Label 𝜖𝑈

‘kshi’, ‘examination’, ‘of’, ‘aquatic’, ‘life’, ‘off’, ‘the’, ‘shores’, ‘of’,
‘the’, ‘Baja’, ‘California’, ‘peninsula’, ‘of’, ‘Mexico’, ‘offers’,
Received Positive
‘an’, ‘engrossing’, ‘way’, ‘to’, ‘demonstrate’, ‘the’, ‘virtues’, 0.021
‘of’, ‘the’, ‘IMAX’, ‘format’, ‘.’
‘This’, ‘examination’, ‘of’, ‘aquatic’, ‘life’, ‘off’, ‘the’, ‘shores’, ‘of’,
‘the’, ‘Baja’, ‘California’, ‘peninsula’, ‘of’, ‘Mexico’, ‘offers’,
Sent Positive
‘an’, ‘engrossing’, ‘way’, ‘to’, ‘demonstrate’, ‘the’, ‘virtues’,
‘of’, ‘the’, ‘IMAX’, ‘format’, ‘.’
‘ abel ’, ‘film’, ‘abby’, ‘dodged’, ‘atv’, ‘local’, ·giorgio’,
Received ·gorgeous’, ‘strategy’, ‘hatch’, ‘fictional’, ·1775’, ‘##mament’, Positive
0.012
‘appreciate’, ‘[unused717]’, ‘##pw’, ‘hole’, ‘sprang’
‘The’, ‘film’, ‘is’, ‘just’, ‘a’, ‘big’, ‘,’, ‘gorgeous’,
Sent Positive
‘,’, ‘mind-blowing’, ‘,’, ‘breath-taking’, ‘mess’, ‘.’

TABLE IV
C OMPARISONS B ETWEEN THE S ENDING AND R ECEIVING W ORDS UNDER R ETRANSMISSION S CHEME IN THE U NCODED S EMANTIC C OMMUNICATION
S YSTEM

Type Words Label

‘meyer’, ‘distribution’, ‘s’, ‘like’, ‘wai’,
Round 1 Negative
‘poem’, ‘trapping’
Received
Round 2 ‘it’, ‘fate’, ‘s’, ‘like’, ‘wai’, ‘poem’, ‘trapping’ Negative
Round 3 ‘it’, “ ’ ”, ‘s’, ‘like’, ‘a’, ‘poem’, ‘trapping’ Positive
Sent ‘it’, “ ’ ”, ‘s’, ‘like’, ‘a’, ‘poem’, ‘.’ Positive

TABLE V
T HE AVERAGE ACCURACY OF THE SEMANTIC TASK AND THE AVERAGE BER WITH DIFFERENT SNR S IN THE CODED SEMANTIC COMMUNICATION
SYSTEM .

(a) Adaptive MCS. (b) Retransmission scheme.

SNR (dB) SNR (dB)
𝑝𝑟𝑡 𝑝𝑟𝑡
0 5 10 15 20 0 5 10 15 20
0 71.34% / 2.88E-01 76.64% / 2.01E-01 86.08% / 1.07E-01 94.02% / 4.28E-02 97.72% / 1.50E-02 0 71.34% / 2.88E-01 76.64% / 2.01E-01 86.08% / 1.07E-01 94.02% / 4.28E-02 97.72% / 1.50E-02
1E-08 81.88% / 1.33E-01 81.72% / 1.21E-01 88.24% / 7.46E-02 94.86% / 3.26E-02 98.06% / 1.11E-02 1E-08 83.76% / 1.34E-01 82.52% / 1.20E-01 88.60% / 7.56E-02 95.26% / 3.14E-02 98.12% / 1.09E-02
1E-07 85.98% / 1.06E-01 85.30% / 1.04E-01 89.26% / 6.80E-02 95.16% / 2.89E-02 98.50% / 9.64E-03 1E-07 85.96% / 1.10E-01 85.50% / 1.04E-01 90.24% / 6.70E-02 95.36% / 2.87E-02 98.34% / 9.80E-03
1E-06 91.24% / 7.41E-02 88.62% / 8.28E-02 91.56% / 5.75E-02 95.98% / 2.46E-02 98.62% / 8.54E-03 1E-06 90.26% / 8.25E-02 89.66% / 8.22E-02 91.32% / 5.67E-02 96.06% / 2.52E-02 98.52% / 9.34E-03
1E-05 96.32% / 4.19E-02 92.74% / 5.69E-02 93.80% / 4.43E-02 96.98% / 2.02E-02 98.88% / 7.14E-03 1E-05 94.10% / 5.67E-02 92.82% / 5.95E-02 93.86% / 4.47E-02 96.56% / 2.02E-02 99.02% / 7.17E-03
1E-04 98.12% / 2.00E-02 96.94% / 3.30E-02 96.68% / 2.82E-02 98.22% / 1.37E-02 99.40% / 5.17E-03 1E-04 96.30% / 3.51E-02 96.48% / 3.67E-02 96.78% / 2.96E-02 98.70% / 1.31E-02 99.24% / 5.13E-03
0.001 99.44% / 8.90E-03 98.90% / 1.54E-02 98.70% / 1.40E-02 99.46% / 6.85E-03 99.86% / 2.56E-03 0.001 98.22% / 1.90E-02 98.54% / 1.85E-02 98.92% / 1.46E-02 99.36% / 7.02E-03 99.84% / 2.42E-03
0.0125 99.86% / 5.46E-03 99.86% / 5.45E-03 99.80% / 5.15E-03 99.82% / 2.61E-03 99.96% / 9.86E-04 0.0125 99.66% / 7.26E-03 99.54% / 6.42E-03 99.66% / 5.47E-03 99.76% / 2.85E-03 99.96% / 1.05E-03
0.025 99.88% / 3.66E-03 99.78% / 3.67E-03 99.68% / 3.81E-03 99.92% / 2.03E-03 99.98% / 8.30E-04 0.025 99.78% / 5.48E-03 99.76% / 4.45E-03 99.88% / 3.86E-03 99.86% / 1.95E-03 99.98% / 6.98E-04
0.0375 99.96% / 2.54E-03 99.92% / 2.94E-03 99.90% / 2.90E-03 99.90% / 1.58E-03 100.00% / 6.11E-04 0.0375 99.90% / 4.35E-03 99.84% / 3.56E-03 99.82% / 3.27E-03 99.90% / 1.57E-03 99.98% / 5.98E-04
0.05 99.98% / 2.46E-03 99.96% / 2.41E-03 99.86% / 2.59E-03 99.92% / 1.34E-03 99.98% / 4.35E-04 0.05 99.92% / 3.38E-03 99.90% / 3.17E-03 99.92% / 2.61E-03 99.96% / 1.31E-03 99.98% / 5.69E-04
0.0625 99.96% / 2.23E-03 99.94% / 2.22E-03 99.88% / 2.29E-03 99.98% / 1.06E-03 100.00% / 4.21E-04 0.0625 99.96% / 2.73E-03 99.98% / 2.65E-03 99.94% / 2.41E-03 99.96% / 1.25E-03 99.98% / 4.37E-04
0.075 100.00% / 2.01E-03 99.98% / 2.00E-03 99.94% / 1.90E-03 99.94% / 1.06E-03 100.00% /4.20E-04 0.075 99.96% / 2.72E-03 99.98% / 2.59E-03 99.96% / 2.03E-03 99.98% / 1.17E-03 99.98% / 4.13E-04
0.0875 99.98% / 1.91E-03 100.00% / 1.86E-03 99.90% / 1.80E-03 99.98% / 1.01E-03 100.00% / 3.01E-04 0.0875 99.96% / 2.43E-03 99.98% / 2.13E-03 99.96% / 1.85E-03 99.98% / 9.52E-04 99.98% / 3.49E-04
0.1 99.98% / 1.78E-03 99.98% / 1.67E-03 100.00% / 1.67E-03 100.00% / 8.78E-04 100.00% / 2.82E-04 0.1 99.96% / 2.03E-03 99.98% / 1.89E-03 99.98% / 1.67E-03 99.98% / 9.08E-04 100.00% / 2.96E-04
0.2 100.00% / 1.28E-03 99.98% / 9.26E-04 100.00% / 8.33E-04 100.00% / 6.05E-04 100.00% / 2.21E-04 0.2 100.00% / 1.16E-03 100.00% / 1.08E-03 100.00% / 9.52E-04 100.00% / 4.95E-04 100.00% / 1.88E-04
0.3 100.00% / 6.83E-04 100.00% / 6.47E-04 100.00% / 6.43E-04 100.00% / 2.93E-04 100.00% / 1.53E-04 0.3 100.00% / 7.75E-04 100.00% / 6.69E-04 100.00% / 6.64E-04 100.00% / 2.90E-04 100.00% / 1.36E-04
0.4 100.00% / 5.77E-04 100.00% / 5.60E-04 100.00% / 5.16E-04 100.00% / 2.85E-04 100.00% / 7.87E-05 0.4 100.00% / 5.27E-04 100.00% / 4.57E-04 100.00% / 4.32E-04 100.00% / 2.65E-04 100.00% / 1.01E-04
0.5 100.00% / 3.97E-04 100.00% / 3.82E-04 100.00% / 2.98E-04 100.00% / 1.82E-04 100.00% / 6.75E-05 0.5 100.00% / 4.01E-04 100.00% / 3.73E-04 100.00% / 3.12E-04 100.00% / 1.83E-04 100.00% / 8.06E-05
0.6 100.00% / 3.01E-04 100.00% / 2.94E-04 100.00% / 2.46E-04 100.00% / 1.07E-04 100.00% / 5.42E-05 0.6 100.00% / 2.81E-04 100.00% / 2.61E-04 100.00% / 2.22E-04 100.00% / 1.25E-04 100.00% / 3.96E-05
0.7 100.00% / 2.55E-04 100.00% / 2.12E-04 100.00% / 1.94E-04 100.00% / 9.38E-05 100.00% / 4.51E-05 0.7 100.00% / 1.98E-04 100.00% / 1.66E-04 100.00% / 1.42E-04 100.00% / 7.98E-05 100.00% / 4.16E-05
0.8 100.00% / 1.46E-04 100.00% / 1.13E-04 100.00% / 9.33E-05 100.00% / 6.44E-05 100.00% / 2.41E-05 0.8 100.00% / 1.36E-04 100.00% / 1.35E-04 100.00% / 1.06E-04 100.00% / 6.52E-05 100.00% / 1.86E-05
0.9 100.00% / 8.01E-05 100.00% / 7.85E-05 100.00% / 7.64E-05 100.00% / 4.69E-05 100.00% / 1.88E-05 0.9 100.00% / 9.08E-05 100.00% / 6.98E-05 100.00% / 6.56E-05 100.00% / 3.50E-05 100.00% / 9.76E-06

rates of the proposed adaptive modulation and retransmission contain relatively less semantic information.
schemes in comparison with baseline algorithms for varying Besides, the proposed retransmission scheme can achieve a
SNRs and varying accuracy requirements. higher transmission rate than the adaptive modulation scheme.
From Fig. 7, we can find that the proposed adaptive mod- It is because that the retransmission scheme effectively uti-
ulation scheme achieves a higher average transmission rate lizes the channel resources that are wasted by the adaptive
than the traditional scheme. The same conclusion can be modulation scheme when the robustness requirement cannot
drawn for the retransmission scheme as well. The reason is be satisfied by the minimum modulation scheme. Since some
that the proposed schemes take into account the differences samples have strict BER requirements, there still exists the
between samples. We can use a low modulation order to case that we have to retransmit the samples for multiple
protect the samples with more semantic information, while rounds especially when the channel experiences deep fading
a high-order modulation can be used to transmit samples that even when SNR=20dB. Thus, the transmission rate of the

retransmission scheme is still slightly higher than that of the time transmission can already meet the robustness requirement
adaptive modulation scheme when SNR=20dB. However, it contrast to multiple transmission rounds in the uncoded sys-
also demonstrates that the retransmission scheme consumes tem. However, the error correction capability does not come
more energy. When SNR is low, the improvements of the for free. The performance enhancement at low SNRs is paid
proposed schemes are not significant. It is because that the by the decrease in data rate per frame. At high SNRs, the
additional waiting and retransmission time is much longer than uncoded system can already achieve low BER using existing
the saved time of the last round transmission. Some points are modulation schemes, while low data rate MCS in the coded
not plotted, since the average accuracy has already exceeded system hinders the increase of transmission rate.
the accuracy requirement even selecting the maximum modu-
lation scheme for these SNRs. D. Performance Gain Analysis

4.0 Pr p sed AMC 1.0

Pr p sed C ded Retran
3.5 Traditi nal AMC
Average transmissi n rate (bit/s/Hz)

Traditi nal C ded Retran 0.8

3.0 Pr p sed AM
Traditi nal AM
2.5 Pr p sed Unc ded Retran
Traditi nal Unc ded Retran 0.6

P(ε > E)
2.0
(2.82e-03,0.46)
1.5 0.4

1.0
0.2
0.5 prt=0.1
prt=0.5
0.0 0.0 prt=0.9
0 5 10 15 20 10−5 10−4 10−3 10−2
SNR (dB) E

Fig. 8. Average transmission rates of the proposed adaptive MCS and coded Fig. 9. The proportion of the maximum BER that can be allowed for the
retransmission scheme compared to baseline and uncoded schemes, when transmitted samples with different robustness probability thresholds.
average accuracy is 100%.
To further illustrate the performance gain, we plot the
2) Coded System: Since coded transmission has high relia- proportion of transmitted samples that can withstand the max-
bility, we choose accuracy requirement as 100% and the base- imum BER at various robustness probability thresholds. The
line algorithms adopt the maximum MCS under the premise of results are summarized in Fig. 9. From the figure, the marked
perfect bit transmission. We further put the result of uncoded point means that if the BER threshold is set as 0.00282, 46%
schemes given average accuracy equals to 100% to compare of the total samples can satisfy the BER requirement given
their rates variation. The results are summarized in Fig. 8. 𝑝 𝑟𝑡 = 0.9. From the figure, the BER requirement for the same
The results show that the proposed coded schemes outperform sample also increases as 𝑝 𝑟𝑡 grows. It is because increasing 𝑝 𝑟𝑡
their baseline algorithms, respectively, indicating the ability means that the perturbed samples are more likely to appear in
to generalize on the coded schemes. The same explanations the same classification space as the initial samples. As a result,
can be found in the uncoded system as mentioned above. we have to reduce the BER to achieve more robustness.
We also find that the proposed coded retransmission scheme Besides, given the same 𝑝 𝑟𝑡 , different samples have different
achieves a little bit higher transmission rate than the proposed BER requirements due to various semantic information. How-
MCS. The performance gain comes from the utilization of ever, the traditional method assigns the same BER requirement
waiting channel resources. However, since the average number to all samples. If the BER requirement is not set to the lowest
of transmission rounds approaches one, the diversity gain from tolerable BER among the samples, then some samples are
combining multiple rounds is limited. subject to misclassification, since samples rich in semantic in-
However, the transmission rate increases more smoothly formation need more stringent protection during transmission.
with SNR in the coded system than in the uncoded system, In contrast, the transmission rate of some samples with little
since the coded system increases the number of optional semantic information is reduced if the lowest tolerant BER is
modulation schemes. In addition, the proposed MCS and set. Thus, the traditional method cannot make a fair balance
coded retransmission scheme show higher transmission rates between the accuracy and transmission rate, which explains
at low SNRs but lower transmission rates at high SNRs. The the performance gain.
result verifies the impact of modulation scheme combinations
on the transmission rate. Since error correction coding reduces VI. C ONCLUSIONS
the BER, the coded system does not need to take advantages In this paper, we mainly studied the adaptive modulation
of diversity gain in fading. Given enough bandwidth, one- and retransmission scheme for semantic communication sys-

Í𝐷
tems, which adapted the transmission rate over time-varying the distribution of 𝑑=1 | 𝑔˜ 𝑛𝑑 − 𝑔𝑛𝑑 | approximates the normal
channels. We focused on the inference stage of NNs, rather distribution, i.e.,
than how to train feature selection NNs. It empowered our 𝐷
∑︁
scheme to generalize across diverse semantic sources, dy- | 𝑔˜ 𝑛𝑑 − 𝑔𝑛𝑑 | ∼ N (𝐷 E(| 𝑔˜ 𝑛𝑑 − 𝑔𝑛𝑑 |), 𝐷 Var(| 𝑔˜ 𝑛𝑑 − 𝑔𝑛𝑑 |)). (18)
namic channel statistics, and multiple semantic tasks without 𝑑=1
joint-retraining. To be more specific, we first used robust- Thus, we have
ness verification to bridge between wireless transmission and
𝐷 max{|𝑥 𝑛𝑑 |} 𝐷 (max{|𝑥 𝑛𝑑 |}) 2
task performance. The robustness probability threshold was ∥ x̃𝑛 − x𝑛 ∥ 1 ∼ N (
𝑛,𝑑
E(| 𝑔˜ 𝑛𝑑 − 𝑔𝑛𝑑 |),
𝑛,𝑑
Var(| 𝑔˜ 𝑛𝑑 − 𝑔𝑛𝑑 |)) .
proposed to guarantee the performance of semantic tasks. 2𝐾 −1 (2𝐾 − 1) 2
(19)
Next, we formulated and resolved the challenge of modulation
scheme selection with the help of the proposed robustness
probability threshold. Based on this, we further designed the R EFERENCES
adaptive modulation and retransmission schemes to improve [1] W. Weaver, “Recent contributions to the mathematical theory of commu-
the transmission rate compared to the traditional schemes. nication,” ETC: A Review of General Semanics, pp. 261–281, 1953.
Finally, simulation results demonstrated that the proposed [2] D. Gündüz, Z. Qin, I. E. Aguerri, H. S. Dhillon, Z. Yang, A. Yener, K. K.
Wong, and C. -B. Chae, “Beyond transmitting bits: Context, semantics,
schemes outperformed the baseline methods, which adapted and task-oriented communications,” IEEE J. Sel. Areas Commun., vol.
transmission rates depending on semantic information behind 41, no. 1, pp. 5–41, Jan. 2023.
the data. The proposed schemes shed lights on the adaptive rate [3] Q. Lan, D. Wen, Z. Zhang, Q. Zeng, X. Chen, P. Popovski, and K. Huang,
“What is semantic communication? A view on conveying meaning in the
control for practical machine learning enabled communication era of machine intelligence,” J. Commun. Netw., vol. 6, no. 4, pp. 336–
systems. 371, Dec. 2021.
[4] X. Luo, H. -H. Chen, and Q. Guo, “Semantic Communications: Overview,
open Issues, and future research directions,” IEEE Wireless Commun., vol.
A PPENDIX A 29, no. 1, pp. 210–219, Feb. 2022.
P ROOF OF THE STATISTICAL PROPERTY OF THE [5] H. Xie, Z. Qin, G. Y. Li, and B. -H. Juang, “Deep learning enabled
DISTORTION BETWEEN THE RECEIVED AND SENT semantic communication systems,” IEEE Trans. Signal Process., vol. 69,
pp. 2663–2675, Apr. 2021.
FEATURES [6] L. Yan, Z. Qin, R. Zhang, Y. Li, and G. Y. Li, “Resource allocation for
Given the quantization in Eq. (1), ∥ x̃𝑛 − x𝑛 ∥ 1 ≤ 𝑞 can be text semantic communications,” IEEE Wireless Commun. Lett., vol. 11,
no. 7, pp. 1394–1398, Jul. 2022.
reformulated as [7] D. B. Kurka and D. Gündüz, “DeepJSCC-f: Deep joint source-channel
max{|𝑥 𝑛𝑑 |} ∑︁ coding of images with feedback,” IEEE Trans. Inf. Theory, vol. 1, no. 1,
𝐷
𝑛,𝑑 pp. 178–193, May 2020.
| 𝑔˜ 𝑛𝑑 − 𝑔𝑛𝑑 | ≤ 𝑞, ∀𝑛, 𝑑, (14) [8] M. Yang, C. Bian, and H.-S. Kim, “Deep joint source channel coding for
2𝐾 − 1 𝑑=1 wireless image transmission with OFDM,” in IEEE ICC 2021, online,
Jun. 2021, pp. 1–6.
where 𝑔˜ 𝑛𝑑 = 𝐾
Í 𝑘 𝑘,𝑑 and 𝑔 𝑑 = Í𝐾 2 𝑘 𝑜 𝑘,𝑑 .
𝑘=1 2 𝑜˜ 𝑛 𝑛 𝑘=1 𝑛 [9] M. Jankowski, D. Gündüz, and K. Mikolajczyk, “Deep joint source-
To analyze the distribution of the left part of inequality (14), channel coding for wireless image retrieval,” in IEEE ICASSP 2020,
Í 𝑘,𝑑 Barcelona, Spain, May 2020, pp. 5070–5074.
we introduce the number of error bits as 𝑒 𝑛𝑑 = 𝐾 𝑘=1 𝑛 −
| ˜
𝑜 [10] Z. Weng and Z. Qin, “Semantic communication systems for speech
𝑜 𝑛𝑘,𝑑 | ∈ {0, . . . , 𝐾 }. We denote the 𝑗-th error situation that transmission,” IEEE J. Sel. Areas Commun., vol. 39, no. 8, pp. 2434–
𝑑,𝑒 𝐾 2444, Aug. 2021.
𝑔𝑛𝑑 has 𝑒 𝑛𝑑 errors as 𝑔˜ 𝑛, 𝑗 , 𝑗 ∈ {1, . . . , 𝑒𝑛𝑑 }. For example, if [11] F. Zhai, Y. Eisenberg, and A. K. Katsaggelos, “Joint source-channel
𝐾 = 3, 𝑒 𝑛𝑑 = 1 and 𝑔𝑛𝑑 = 0(𝑜 1,𝑑 2,𝑑 3,𝑑
𝑛 𝑜 𝑛 𝑜 𝑛 = 000), there are three coding for video communications,” Handbook Image Video Process.,
error situations, i.e., 2( 𝑜˜ 𝑛 𝑜˜ 𝑛 𝑜˜ 𝑛 = 001), 4( 𝑜˜1,𝑑
1,𝑑 2,𝑑 3,𝑑 2,𝑑 3,𝑑
𝑛 𝑜˜ 𝑛 𝑜˜ 𝑛 =
2005, pp. 1065–1082.
1,𝑑 2,𝑑 3,𝑑 [12] Q. Hu, G. Zhang, Z. Qin, Y. Cai, G. Yu, and G. Y. Li “Ro-
010), and 8( 𝑜˜ 𝑛 𝑜˜ 𝑛 𝑜˜ 𝑛 = 100). Then, due to i.i.d. channel bust semantic communications with masked VQ-VAE enabled code-
impairments, the expectation of | 𝑔˜ 𝑛𝑑 − 𝑔𝑛𝑑 | can be expressed as book.” IEEE Trans. Wireless Commun., early access, Apr. 12, 2023,
doi:10.1109/TWC.2023.3265201.
𝐾 −1
2∑︁ 𝐾 ( 𝑒𝐾𝑛𝑑 ) [13] G. Zhang, Q. Hu, Z. Qin, Y. Cai, and G. Yu, “A unified multi-
∑︁ ∑︁ 1 task semantic communication system with domain adaptation,” in IEEE
𝑑,𝑒
E(| 𝑔˜ 𝑛𝑑 − 𝑔𝑛𝑑 |) = 𝑝(𝑔𝑛𝑑 ) 𝑝(𝑒 𝑛𝑑 ) 𝐾
| 𝑔˜ 𝑛, 𝑑
𝑗 − 𝑔 𝑛 |, GLOBECOM, Rio de Janeiro, Brazil, Dec. 2022, pp. 3971–3976.
𝑔𝑛𝑑 =0 𝑒𝑛𝑑 =0 𝑗=1 𝑒𝑛𝑑 [14] E. Bourtsoulatze, D. Burth Kurka, and D. Gündüz, “Deep joint source
(15) channel coding for wireless image transmission,” IEEE Trans. Cogn.
Commun. Netw., vol. 5, no. 3, pp. 567–579, Sep. 2019.
where 𝑝(𝑒 𝑛𝑑 ) = 𝑒𝐾𝑛𝑑 𝜖 𝑒𝑛 (1 − 𝜖) 𝐾 −𝑒𝑛 is the probability of
𝑑 𝑑

[15] D. Tse and P. Viswanath, Fundamentals of Wireless Communication.

appearing 𝑒 𝑛𝑑 error bits and 𝑝(𝑔𝑛𝑑 ) is the prior distribution of Cambridge, U.K.: Cambridge Univ. Press, 2005.
𝑔𝑛𝑑 . Similarly, we can obtain the variance of | 𝑔˜ 𝑛𝑑 − 𝑔𝑛𝑑 | as [16] A. Goldsmith, Wireless Communications. Cambridge, U.K.: Cambridge
Univ. Press, 2005.
Var(| 𝑔˜ 𝑛𝑑 − 𝑔𝑛𝑑 |) = E(| 𝑔˜ 𝑛𝑑 − 𝑔𝑛𝑑 | 2 ) − E2 (| 𝑔˜ 𝑛𝑑 − 𝑔𝑛𝑑 |), (16) [17] J. Shao, Y. Mao, and J. Zhang, “Learning task-oriented communication
for edge inference: An information bottleneck approach,” IEEE J. Sel.
where Areas Commun., vol. 40, no. 1, pp. 197–211, Jan. 2022.
[18] F., Liu, W. Tong, Z. Sun, and C. Guo, “Task-oriented semantic commu-
𝐾 −1
2∑︁ 𝐾 ( 𝑒𝐾𝑛𝑑 ) nication systems based on extended rate-distortion theory.” arXiv preprint
∑︁ ∑︁ 1 arXiv:2201.10929, Jan. 2022.
E(| 𝑔˜ 𝑛𝑑 − 𝑔𝑛𝑑 | 2 ) = 𝑝(𝑔𝑛𝑑 ) 𝑝(𝑒 𝑛𝑑 ) 𝐾
𝑑,𝑒
| 𝑔˜ 𝑛, 𝑑 2
𝑗 − 𝑔𝑛 | . [19] M. Yang and H.-S. Kim, “Deep joint source-channel coding for wireless
𝑔𝑛𝑑 =0 𝑒𝑛𝑑 =0 𝑗=1 𝑒𝑛𝑑 image transmission with adaptive rate control,” in IEEE ICASSP 2022,
(17) Singapore, May 2022, pp. 5193–5197.
[20] S. Guo, Y. Wang, and P. Zhang, “Signal shaping for semantic commu-
Note that the feature dimension 𝐷 is always very large and nication systems with a few message candidates,” in IEEE VTC 2022,
| 𝑔˜ 𝑛𝑑 − 𝑔𝑛𝑑 | is i.i.d.. According to the central limit theorem, London, United Kingdom, Sep. 2022, pp. 1–5.

[21] H. Xie, Z. Qin, and G. Y. Li, “Semantic communication with memory.” Huiguo Gao Huiguo Gao received the B.S. degree in
arXiv preprint arXiv:2303.12335, Mar. 2023. information engineering from Zhejiang University,
[22] M. Ding, J. Li, M. Ma, and X. Fan, “SNR-adaptive deep joint source- Hangzhou, China, in 2020. He is currently pursuing
channel coding for wireless image transmission,” in IEEE ICASSP 2021, the Ph.D. degree with the College of Information
Toronto, Ontario, Canada, Jun. 2021, pp. 1555–1559. Science and Electronic Engineering, Zhejiang Uni-
[23] Q. Zhou, R. Li, Z. Zhao, Y. Xiao, and H. Zhang, “Adaptive bit rate versity. His current research interests mainly include
control in semantic communication with incremental knowledge-based machine learning and wireless communications.
HARQ,” IEEE Open J. Commun. Soc., vol. 3, pp. 1076–1089, Jul. 2022.
[24] P. Jiang, C. -K. Wen, S. Jin, and G. Y. Li, “Deep source-channel coding
for sentence semantic transmission with HARQ,” IEEE Trans. Commun.,
vol. 70, no. 8, pp. 5225–5240, Aug. 2022.
[25] G. I. Parisi, R. Kemker, J. L. Part, C. Kanan, and S. Wermter, “Continual
lifelong learning with neural networks: A review,” Neural Networks, vol.
113, pp. 54–71, Feb. 2019.
[26] M. McCloskey and N. J. Cohen, “Catastrophic interference in connec-
tionist networks: The sequential learning problem,” Psychol. Learn Motiv., Guanding Yu (S’05-M’07-SM’13) received the B.E.
Elsevier, 1989, vol. 24, pp. 109–165. and Ph.D. degrees in communication engineering
[27] R. R. Bunel, I. Turkaslan, P. Torr, P. Kohli, and P. K. Mudigonda, “A from Zhejiang University, Hangzhou, China, in 2001
unified view of piecewise linear neural network verification,” in Proc. and 2006, respectively. He joined Zhejiang Uni-
Adv. Neural Inf. Process. Syst. (NeurIPS), Montreal, Canada, Dec. 2018, versity in 2006, and is now a Professor with the
pp. 4795–4804. College of Information and Electronic Engineering.
[28] V. Tjeng, K. Xiao, and R. Tedrake, ”Evaluating robustness of From 2013 to 2015, he was also a Visiting Professor
neural networks with mixed integer programming,” arXiv preprint at the School of Electrical and Computer Engi-
arXiv:1711.07356, Nov. 2017. neering, Georgia Institute of Technology, Atlanta,
[29] G. Katz, C. Barrett, D. L. Dill, K. Julian, and M. J. Kochenderfer, “Re- GA, USA. His research interests include integrated
luplex: An efficient SMT solver for verifying deep neural networks,” in sensing and communications (ISAC), mobile edge
Proc. Int. Conf. Comput.-Aided Verification (CAV), Heidelberg, Germany, computing/learning, and machine learning for wireless networks.
Jul. 2017, pp. 97–117. Dr. Yu has served as a guest editor of IEEE Communications Magazine
[30] M. Lee, G. Yu, and H. Dai, “Decentralized inference with graph special issue on Full-Duplex Communications, an Editor of IEEE Journal
neural networks in wireless communication systems,” IEEE Trans. Mob. on Selected Areas in Communications Series on Green Communications
Comput., vol. 22, no. 5, May 2023, pp. 2582–2598. and Networking, and Series on Machine Learning in Communications and
[31] T. Gehr, M. Mirman, D. Drachsler-Cohen, P. Tsankov, S. Chaudhuri, and Networks, an Editor of IEEE Wireless Communications Letters, a lead
M. Vechev, “AI2: Safety and robustness certification of neural networks Guest Editor of IEEE Wireless Communications Magazine special issue on
with abstract interpretation,” in Proc. IEEE Symp. Secur. Privacy (SP), LTE in Unlicensed Spectrum, an Editor of IEEE Transactions on Green
San Francisco, California, USA, May 2018, pp. 3–18. Communications and Networking, and an Editor of IEEE Access. He is
[32] Matthias Hein and Maksym Andriushchenko, “Formal guarantees on the now serving as an editor of IEEE Transactions on Machine Learning in
robustness of a classifier against adversarial manipulation.” in Proc. Adv. Communications and Networking. He received the 2016 IEEE ComSoc
Neural Inf. Process. Syst. (NeurIPS), Long Beach, California, USA, Dec. Asia-Pacific Outstanding Young Researcher Award. He regularly sits on the
2017, pp. 2266–2276. technical program committee (TPC) boards of prominent IEEE conferences
[33] Z. Shi, H. Zhang, K.-W. Chang, M. Huang, and C.-J. Hsieh, “Robustness such as ICC, GLOBECOM, and VTC. He also serves as a Symposium Co-
verification for transformers,” in Int’l. Conf. Learn. Represent. (ICLR), Chair for IEEE Globecom 2019 and a Track Chair for IEEE VTC 2019’Fall.
online, Apr. 2020, pp. 1–22.
[34] Y. Kang, J. Hauswald, C. Gao, A. Rovinski, T. Mudge, J. Mars, and L.
Tang, “Neurosurgeon: Collaborative intelligence between the cloud and
mobile edge,” ACM SIGARCH Comput. Archit. News, vol. 45, no. 1, pp.
615–629, Apr. 2017.
[35] D. R. Pauluzzi and N. C. Beaulieu, “A comparison of SNR estimation
techniques for the AWGN channel,” IEEE Trans. Commun., vol. 48, no.
10, pp. 1681–1691, Oct. 2000. Yunlong Cai (S’07–M’10–SM’16) received the
[36] H. Ye, G. Y. Li, and B. -H. Juang, “Power of deep learning for channel Ph.D. degree in electronic engineering from the Uni-
estimation and signal detection in OFDM systems,” IEEE Wireless versity of York, U.K., in 2010. From 2010 to 2011,
Commun. Lett., vol. 7, no. 1, pp. 114–117, Feb. 2018. he was a Postdoctoral fellow with the Electronics
[37] J. Lu, K. B. Letaief, J. C.-I. Chuang, and M. L. Liou, “M-PSK and and Communications Laboratory, CNAM, France.
M-QAM BER computation using signal-space concepts,” IEEE Trans. Since February 2011, he has been with the College
Commun., vol. 47, no. 2, pp. 181–184, Feb. 1999. of Information Science and Electronic Engineering,
[38] J. Hou, P. H. Siegel, and L. B. Milstein, “Performance analysis and Zhejiang University, Hangzhou, China, where he is
code optimization of low density parity-check codes on Rayleigh fading currently a Professor. He has also held research vis-
channels,” IEEE J. Sel. Areas Commun., vol. 19, no. 5, pp. 924–934, May iting appointments at Georgia Institute of Technol-
2001. ogy, McGill University, and University of California
[39] P. Frenger, S. Parkvall, and E. Dahlman, “Performance comparison of Irvine. His research interests include transceiver design for multiple-antenna
HARQ with chase combining and incremental redundancy for HSDPA,” systems, cooperative and relay communications, UAV communications, and
in IEEE VTC Fall 2001, Atlantic City, New Jersey, USA, Oct. 2001, vol. machine learning for communications. He has published over 200 journal and
3, pp. 1829–1833. conference papers in these areas.
[40] A. Bin Sediq and H. Yanikomeroglu, “Performance analysis of soft-bit Dr. Cai currently serves as an Associate Editor for IEEE Transactions
maximal ratio combining in cooperative relay networks,” IEEE Trans. on Communications and a Senior Area Editor for IEEE Signal Processing
Wireless Commun., vol. 8, no. 10, pp. 4934–4939, Oct. 2009. Letters. He is the Lead Guest Editor of the special issue on “Next Generation
[41] R. Socher, A. Perelygin, J. Y. Wu, J. Chuang, C. D. Manning, A. Y. Advanced Transceiver Technologies” of the IEEE Journal on Selected Areas
Ng, and C. Potts, “Recursive deep models for semantic compositionality in Communications. He was an Associate Editor for IEEE Signal Processing
over a sentiment treebank,” in Proc. Conf. Empirical Methods Natural Letters from 2018 to 2023. He has served as the General Chair for the eigh-
Language Process., Seattle, Washington, USA, Oct. 2013, pp. 1631–1642. teenth IEEE International Symposium on Wireless Communications Systems,
[42] Eng Hwee Ong, J. Kneckt, O. Alanen, Z. Chang, T. Huovinen, and which was held in Hangzhou, in October 2022. He regularly sits on the
T. Nihtilä, “IEEE 802.11ac: Enhancements for very high throughput technical program committee boards of prominent IEEE conferences, such
WLANs,” in IEEE PIMRC, Toronto, Canada, Sep. 2011, pp. 849–853. as ICC, GLOBECOM, and VTC.

Semantic Communications Overview Open Issues and Future Research Directions
No ratings yet
Semantic Communications Overview Open Issues and Future Research Directions
10 pages
Galois Field Computations With Matlab
No ratings yet
Galois Field Computations With Matlab
31 pages
A Generalized Semantic Communication System From Sources To Channels
No ratings yet
A Generalized Semantic Communication System From Sources To Channels
9 pages
Deep Learning Based Superposition Coded Modulation For Hierarchical Semantic Communications Over Broadcast Channels
No ratings yet
Deep Learning Based Superposition Coded Modulation For Hierarchical Semantic Communications Over Broadcast Channels
15 pages
Deep Learning Enabled Semantic Communication Systems
No ratings yet
Deep Learning Enabled Semantic Communication Systems
13 pages
Hybrid Digital-Analog Semantic Communications
No ratings yet
Hybrid Digital-Analog Semantic Communications
16 pages
Contrastive Learning-Based Semantic Communications
No ratings yet
Contrastive Learning-Based Semantic Communications
16 pages
(Survey) Semantic Communications - Principles and Challenges
No ratings yet
(Survey) Semantic Communications - Principles and Challenges
32 pages
Semantic Communication On Digital Wireless Communi
No ratings yet
Semantic Communication On Digital Wireless Communi
21 pages
Outage Probability Ref1
No ratings yet
Outage Probability Ref1
14 pages
Gunduz 等 - 2023 - Beyond Transmitting Bits Context, Semantics, And
No ratings yet
Gunduz 等 - 2023 - Beyond Transmitting Bits Context, Semantics, And
37 pages
深度学习通过语音识别和合成实现语义通信
No ratings yet
深度学习通过语音识别和合成实现语义通信
14 pages
A Theory of Semantic Communication
No ratings yet
A Theory of Semantic Communication
18 pages
A Unified Multi-Task Semantic Communication System For Multimodal Data
No ratings yet
A Unified Multi-Task Semantic Communication System For Multimodal Data
16 pages
A Lite Distributed Semantic Communication System For Internet of Things
No ratings yet
A Lite Distributed Semantic Communication System For Internet of Things
12 pages
Contrastive Learning Based Semantic
No ratings yet
Contrastive Learning Based Semantic
5 pages
Research Topics For The Year 2024
No ratings yet
Research Topics For The Year 2024
2 pages
Semantic Communications With Explicit Semantic Bases Model, Architecture, and Open Problems
No ratings yet
Semantic Communications With Explicit Semantic Bases Model, Architecture, and Open Problems
8 pages
Energy Efficient Semantic Communication Over Wireless Networks With Rate Splitting
No ratings yet
Energy Efficient Semantic Communication Over Wireless Networks With Rate Splitting
12 pages
Less Data More Knowledge Semantic Walid
No ratings yet
Less Data More Knowledge Semantic Walid
38 pages
Semantic Communications in Networked Systems A Data Significance Perspective
No ratings yet
Semantic Communications in Networked Systems A Data Significance Perspective
8 pages
What Is Semantic Communication A View On Conveying Meaning in The Era of Machine Intelligence
No ratings yet
What Is Semantic Communication A View On Conveying Meaning in The Era of Machine Intelligence
36 pages
Semantic
No ratings yet
Semantic
12 pages
Classification-Driven Discrete Neural Representation Learning For Semantic Communications
No ratings yet
Classification-Driven Discrete Neural Representation Learning For Semantic Communications
13 pages
Tech Seminar 5301
No ratings yet
Tech Seminar 5301
16 pages
An Observational Theory For Mobile Ad Hoc Networks
No ratings yet
An Observational Theory For Mobile Ad Hoc Networks
29 pages
Effective Communications: A Joint Learning and Communication Framework For Multi-Agent Reinforcement Learning Over Noisy Channels
No ratings yet
Effective Communications: A Joint Learning and Communication Framework For Multi-Agent Reinforcement Learning Over Noisy Channels
31 pages
Arda Journal 15946
No ratings yet
Arda Journal 15946
13 pages
Simultaneously Transmitting and Reflecting STAR RIS Aided Wireless Communications
No ratings yet
Simultaneously Transmitting and Reflecting STAR RIS Aided Wireless Communications
23 pages
Cognitive Radio Networks: Highlights of Information Theoretic Limits, Models, and Design
No ratings yet
Cognitive Radio Networks: Highlights of Information Theoretic Limits, Models, and Design
12 pages
Large AI Model Empowered Multimodal Semantic Communications
No ratings yet
Large AI Model Empowered Multimodal Semantic Communications
7 pages
A Survey On Goal-Oriented Semantic Communication Techniques Challenges and Future Directions
No ratings yet
A Survey On Goal-Oriented Semantic Communication Techniques Challenges and Future Directions
52 pages
On Distributed Communication Networks
No ratings yet
On Distributed Communication Networks
41 pages
Learning Task-Oriented Communication For Edge Inference: An Information Bottleneck Approach
No ratings yet
Learning Task-Oriented Communication For Edge Inference: An Information Bottleneck Approach
14 pages
Wireless End-to-End Image Transmission System Using Semantic Communications
No ratings yet
Wireless End-to-End Image Transmission System Using Semantic Communications
15 pages
Digital Signal Processing Using Deep Neural Networks: Brian Shevitski, Yijing Watkins, Nicole Man and Michael Girard
No ratings yet
Digital Signal Processing Using Deep Neural Networks: Brian Shevitski, Yijing Watkins, Nicole Man and Michael Girard
21 pages
Deep Learning-Driven Wireless Communication For Edge-Cloud Computing: Opportunities and Challenges
No ratings yet
Deep Learning-Driven Wireless Communication For Edge-Cloud Computing: Opportunities and Challenges
14 pages
On The Role of ViT and CNN in Semantic Communications Analysis and Prototype Validation
No ratings yet
On The Role of ViT and CNN in Semantic Communications Analysis and Prototype Validation
14 pages
MWC Soft
No ratings yet
MWC Soft
85 pages
Screenshot 2025-01-24 at 03.47.29
No ratings yet
Screenshot 2025-01-24 at 03.47.29
31 pages
Physical Layer Communications System Design Over-the-Air Using Adversarial Networks
No ratings yet
Physical Layer Communications System Design Over-the-Air Using Adversarial Networks
4 pages
Compressive Sensing-Based Adaptive Active User Detection and Channel Estimation Massive Access Meets Massive MIMO
No ratings yet
Compressive Sensing-Based Adaptive Active User Detection and Channel Estimation Massive Access Meets Massive MIMO
16 pages
Deep Learning in Physical Layer Communications: Zhijin Qin, Hao Ye, Geoffrey Ye Li, and Biing-Hwang Fred Juang
No ratings yet
Deep Learning in Physical Layer Communications: Zhijin Qin, Hao Ye, Geoffrey Ye Li, and Biing-Hwang Fred Juang
14 pages
Modern Digital Radio Communication Signals and Systems 2nbsped 9783030577056 3030577058 Compress
No ratings yet
Modern Digital Radio Communication Signals and Systems 2nbsped 9783030577056 3030577058 Compress
760 pages
0912 0599 PDF
No ratings yet
0912 0599 PDF
13 pages
Task-Oriented Communications For 6G: Vision, Principles, and Technologies
No ratings yet
Task-Oriented Communications For 6G: Vision, Principles, and Technologies
8 pages
E C A U C R S: Rgodic Apacity Nalysis FOR Nderlay Ognitive Adio Ystem
No ratings yet
E C A U C R S: Rgodic Apacity Nalysis FOR Nderlay Ognitive Adio Ystem
7 pages
A Conceptual Foundation For The Shannon-Weaver Model of Communication
No ratings yet
A Conceptual Foundation For The Shannon-Weaver Model of Communication
8 pages
Electronics Letters - 2019 - Liu - Dynamic Power Splitting Scheme For DF Relaying Based SWIPT Networks With Direct Link
No ratings yet
Electronics Letters - 2019 - Liu - Dynamic Power Splitting Scheme For DF Relaying Based SWIPT Networks With Direct Link
4 pages
Innovative Variational AutoEncoder For An End-To-End Communication System
No ratings yet
Innovative Variational AutoEncoder For An End-To-End Communication System
14 pages
2020handout4 Supervisory Control Circuits
No ratings yet
2020handout4 Supervisory Control Circuits
21 pages
Advanced Topics and Smart Systems For Wireless Com
No ratings yet
Advanced Topics and Smart Systems For Wireless Com
4 pages
A Survey of Recent Advances in Optimization
No ratings yet
A Survey of Recent Advances in Optimization
39 pages
Project Themes
No ratings yet
Project Themes
23 pages
Challenges of Intervehicle Ad Hoc Networks
No ratings yet
Challenges of Intervehicle Ad Hoc Networks
5 pages
Interference Assisted Energy Harvesting Short Pa - 2024 - International Journal
No ratings yet
Interference Assisted Energy Harvesting Short Pa - 2024 - International Journal
10 pages
Communication Systems Using ARM
No ratings yet
Communication Systems Using ARM
62 pages
Introduction To Communication Systems: Saeid Karamzadeh
No ratings yet
Introduction To Communication Systems: Saeid Karamzadeh
27 pages
Tcomm Zhang 2973263 Proof PDF
No ratings yet
Tcomm Zhang 2973263 Proof PDF
34 pages
Integrating Convex Optimization and Deep Learning For Downlink Resource Allocation in LEO Satellites Networks
No ratings yet
Integrating Convex Optimization and Deep Learning For Downlink Resource Allocation in LEO Satellites Networks
15 pages
Joint NTN Slicing and Admission Control For Infrastructure-as-a-Service A Deep Learning Aided Multi-Objective Optimization
No ratings yet
Joint NTN Slicing and Admission Control For Infrastructure-as-a-Service A Deep Learning Aided Multi-Objective Optimization
16 pages
Building Energy Efficient Semantic Segmentation in Intelligent Edge Computing
No ratings yet
Building Energy Efficient Semantic Segmentation in Intelligent Edge Computing
11 pages
5G Embraces Satellites For 6G Ubiquitous IoT Basic Models For Integrated Satellite Terrestrial Networks
No ratings yet
5G Embraces Satellites For 6G Ubiquitous IoT Basic Models For Integrated Satellite Terrestrial Networks
19 pages
Test Score Descriptions
No ratings yet
Test Score Descriptions
1 page
Mock Analysis
No ratings yet
Mock Analysis
1 page
RRL
No ratings yet
RRL
20 pages
Amazing Adventures Book of Powers
67% (3)
Amazing Adventures Book of Powers
50 pages
13 Fault Tree Analysis
100% (1)
13 Fault Tree Analysis
13 pages
Uts Finals Reviewer Complete
No ratings yet
Uts Finals Reviewer Complete
6 pages
Chapter 4 and Appendxeix
No ratings yet
Chapter 4 and Appendxeix
11 pages
Lyrics
No ratings yet
Lyrics
42 pages
Strategic Marketing Plan For Black Thunder
No ratings yet
Strategic Marketing Plan For Black Thunder
24 pages
Ministry of Resin Exposure Times - Durable Grey
No ratings yet
Ministry of Resin Exposure Times - Durable Grey
1 page
TDS - en - Tecnotop S-3000 T
No ratings yet
TDS - en - Tecnotop S-3000 T
5 pages
Conventions CoralReefs
No ratings yet
Conventions CoralReefs
22 pages
Review-Midterm-1-Eng 6
No ratings yet
Review-Midterm-1-Eng 6
6 pages
Lect 07
No ratings yet
Lect 07
44 pages
.Trashed 1724238737 BPSC Senior Secondary Teacher Booklets Sanskrit
No ratings yet
.Trashed 1724238737 BPSC Senior Secondary Teacher Booklets Sanskrit
32 pages
GE8 - What Is The Primary Reason of The Author in Writing The Documents How Was It Produced
No ratings yet
GE8 - What Is The Primary Reason of The Author in Writing The Documents How Was It Produced
2 pages
EAI-05 Warehouse
No ratings yet
EAI-05 Warehouse
2 pages
Text To Image Survey
No ratings yet
Text To Image Survey
40 pages
7673 Final Report - ST-2016-7673-1
No ratings yet
7673 Final Report - ST-2016-7673-1
58 pages
Resume Math
No ratings yet
Resume Math
4 pages
Unit - 2 TC
No ratings yet
Unit - 2 TC
23 pages
Kinerja Ruas Dan Simpang Jalan
No ratings yet
Kinerja Ruas Dan Simpang Jalan
43 pages
Nursing Management and Leadership Approaches From The Perspective of Registered Nurses in Portugal
No ratings yet
Nursing Management and Leadership Approaches From The Perspective of Registered Nurses in Portugal
8 pages
Agri Sba
96% (25)
Agri Sba
15 pages
Free Online AI Face Swap 2
No ratings yet
Free Online AI Face Swap 2
1 page
Global Supply Chains
No ratings yet
Global Supply Chains
25 pages
Eastman Turbo Oil 2180 MSDS
No ratings yet
Eastman Turbo Oil 2180 MSDS
14 pages
Exp - S5 - Vapour Liquid Equilibrium - Corrected
No ratings yet
Exp - S5 - Vapour Liquid Equilibrium - Corrected
6 pages
Island of Ignorance 31 Aug 23 Digital Draft
No ratings yet
Island of Ignorance 31 Aug 23 Digital Draft
41 pages
CSET106 DMS Course File
No ratings yet
CSET106 DMS Course File
4 pages

Adaptive Modulation and Retransmission Scheme For Semantic Communication Systems

Uploaded by

Adaptive Modulation and Retransmission Scheme For Semantic Communication Systems

Uploaded by

This article has been accepted for publication in IEEE Transactions on Cognitive Communications and Networking.

Adaptive Modulation and Retransmission Scheme

a universal framework that can tackle different NNs. A. Classification Problem

Considering the trade-off between communication and com-

B. Wireless Communication Model a retransmission. However, in practice, the RX usually keeps

Due to the monotonic error tolerance, the inference result

transmit. As soon as 𝜖 min is lower than 𝜖 𝑈 , the TX transmits the

Fig. 3. Adaptive modulation scheme in semantic communication systems.

Fig. 4. Retransmission scheme in semantic communication systems.

Fig. 5. The structure of NNs at the TX and RX.

(a) Adaptive modulation scheme. (b) Retransmission scheme.

tences. To be specific, we find that 𝜖 𝑈 becomes smaller as Pro osed AM (95%)

a total of four emotional words and its tolerant BER is

Type Words Label 𝜖𝑈

Type Words Label

(a) Adaptive MCS. (b) Retransmission scheme.

4.0 Pr p sed AMC 1.0

Traditi nal C ded Retran 0.8

[15] D. Tse and P. Viswanath, Fundamentals of Wireless Communication.

You might also like