Semantic-Aware Task Clustering for Constructive and Cooperative Multi-Tasking Ahmad Halimi Razlighi , Maximilian H. V. Tillmann , Edgar Beck , Bho Matthiesen† , and Armin Dekorsy Department of Communications Engineering, University of Bremen, Germany † Communication Systems Group, Paderborn University, 33098 Paderborn, Germany
arXiv:2607.21426v1 [cs.LG] 23 Jul 2026
E-mails:{halimi, tillmann, beck, dekorsy}@ant.uni-bremen.de, † E-mail: [email protected]
Abstract—Cooperative multi-task semantic communication (CMT-SemCom) improves task execution performance by leveraging shared representations. However, as we demonstrated in [1], cooperative multi-tasking can be either constructive or destructive, depending on the semantic relationships among tasks. To ensure constructive cooperation, we propose a semanticaware task clustering method for CMT-SemCom. We have formulated a sequential multi-stage optimization problem in which semantically aligned tasks are clustered once after a short initial training phase, and then end-to-end (E2E) joint training is conducted exclusively within the discovered groups. Specifically, the problem decomposes into two stages: (i) a semantic clustering problem leveraging hierarchical density-based spatial clustering, and (ii) an intra-cluster E2E CMT-SemCom learning problem. Simulation results demonstrate that the proposed framework effectively mitigates destructive cooperation and negative transfer, yielding accuracy gains compared to unclustered multi-tasking and individual training baselines. Index Terms—Semantic communication, task-oriented communication, multi-task learning, task grouping, HDBSCAN, UMAP.
I. I NTRODUCTION Task-oriented semantic communication (SemCom) is recognized as a key enabler for emerging intelligent wireless networks [2]. In these networks, artificial intelligence (AI)-driven communication extends beyond traditional data-oriented services (e.g., voice and text transmission) to support the execution of intelligent tasks [3]. Practical task-oriented communication systems often need to support multiple downstream tasks simultaneously. Therefore, multi-task learning (MTL) strategies have recently been investigated for communication systems to exploit common features across tasks and avoid redundant processing [4]. MTL has been studied as a mechanism for improving generalization performance through shared representation learning [5]. Consequently, recent works have incorporated MTL into SemCom exclusively based on machine learning (ML) approaches and treating the semantic link as a black-box neural network [6], [7]. On the other hand, our previous work [1] introduces an information maximization (InfoMax) perspective on cooperative multi-task SemCom (CMT-SemCom), moving beyond the black-box use of deep neural networks (DNNs). This work investigates dividing the semantic encoder into a common unit (CU) and multiple specific units (SUs) to This work was partially supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy – EXC-3036 The Martian Mindset, project number: 533607631, and by the DFG under project number: 518671822.
enable multi-tasking. The CU captures common information from the observation among multiple tasks, while SUs further process this information for individual tasks. The resulting CMT-SemCom framework facilitates multi-tasking based on a single observation. Subsequently, [8], [9] extended CMTSemCom to scenarios with distributed partial observations and to rate-limited wireless channels, respectively. Despite these advancements, existing multi-task SemCom works, including CMT-SemCom, assume that all tasks are fully related and implicitly learn a shared feature subspace for them. However, applying these methods to unrelated tasks leads to suboptimal task execution and potential performance degradation. As demonstrated in [1], such mismatched cooperation results in destructive cooperation, where unrelated tasks negatively impact each other’s performance. This motivates identifying groups of semantically related tasks that should share latent semantic processing via the CU. In MTL, task grouping has traditionally been addressed either through extensive search [10], subjective human judgment [11], or by exploiting task similarities based on datasets, single-task model parameters, or model gradients [12]–[15]. These approaches can face significant challenges when adapted to the CMT-SemCom. Specifically, gradientbased or parameter-alignment methods require complex backward passes through the differentiable wireless channel model and multiple encoder-decoder pairs, introducing prohibitive computational overhead during the clustering phase. Recently, we proposed a lightweight semantic-aware clustering approach based on the empirical probability mass functions of the semantic variables for distributed federated learning (FL)-based multi-user scenarios [16]. The proposed method leveraged an information-theoretic metric, i.e., JS-divergence, to group semantically related tasks while maintaining low communication overhead during the FL process. Such an approach is particularly suitable for distributed SemCom systems, in which encoders are not connected with each other through a CU on the transmitter side, preventing transmission of high-dimensional feature representations, gradients, or intermediate model statistics to a central processor, and guaranteeing communication efficiency. However, although [16] provides a communication-efficient proxy for task-relatedness, it fails to capture the nuanced, higher-order semantic relationships embedded in the shared representation space. In the centralized CMT-SemCom system, where multi-task training is performed directly on the encoder
This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible.
side with a shared CU, the intermediate CU outputs features that are critical for precise task grouping. Consequently, clustering solely based on the statistics of semantic variables may fail to fully characterize task-relatedness in CMT-SemCom. To address this issue, we propose an approach to manage the shared subspace, ensuring that only fully or partially related tasks cooperate to guarantee constructive cooperation in CMTSemCom. Specifically, we formulate a multi-stage optimization problem where semantic-aware clustering is performed to identify related tasks, followed by an end-to-end (E2E) learning step that ensures only informative tasks are jointly processed. In particular, we introduce a hybrid clustering pipeline that leverages uniform manifold approximation and projection (UMAP) [17] and hierarchical density-based spatial clustering of applications with noise (HDBSCAN) [18] to extract task-relatedness from the CU output feature space. Our method operates on the high-dimensional semantic embeddings generated by the shared encoder, capturing complex nonlinear relationships between tasks. The clustering is executed once after a short initial training phase, fixing the task groups for the remainder of the training. In summary, building upon our prior work on the CMTSemCom, we contribute to a framework capable of mitigating destructive cooperation. Key contributions include: • Proposing a multi-stage problem for Clustered-CMTSemCom, integrating semantic-aware clustering to identify related tasks, and an E2E learning step to ensure only informative tasks are jointly processed, • Proposing a task clustering method by applying UMAP and HDBSCAN on extracted semantic features, moving beyond common clustering approaches that rely on data distribution or similarity of models parameters, • Demonstrating that the proposed framework effectively mitigates destructive cooperation and improves task execution performance compared to unclustered multitasking and individual task training. II. P ROBLEM F ORMULATION In this section, we establish a probabilistic model to characterize the semantic source and communication aspects. Based on this foundation, we define the overall Clustered-CMTSemCom optimization problem. A. System Model We consider a multi-task communication system, consisting of a single transmitter (Tx), multiple receivers (Rxs), and a wireless channel in between. The Tx makes an observation S and extracts multiple task-specific information via a split structure, each corresponding to one of N downstream tasks. At Rx i, this information is used to decode a reconstruction of semantic variable Zi to support execution of its task. We denote the tuple (Z, S) as semantic source [1], fully described by its probability distribution p(z, s), where Z = [ Z1 Z2 . . . ZN ]. Our goal is to design an efficient multi-task semantic encoder using data-driven training. In particular, we consider
x1 C y1
SU Enc. Semantic Source (Z, S)
CU1 Enc.
.. .
s
CUK ⋆ Enc.
h a x2 n n e xN l
SU Enc.
.. .
SU Enc.
y2
Dec.
ẑ1
Dec.
ẑ2
Dec.
ẑN
.. .
yN
Fig. 1: Illustration of the proposed Clustered-CMT-SemCom. the split encoder structure in Fig. 1, consisting of K CUs indexed by k and N SUs indexed by i. Let (z, s) be a semantic source sample, where only s is available to the Tx. Then, CU k preprocesses this observation s for a subset of semantically related tasks Tk ⊆ T = {1, . . . , N }. SU i ∈ Tk takes CU k’s output ck and computes a length-d codeword xi . This codeword is transmitted over an AWGN channel to Rx i. The received signal yi = xi + ni with ni ∼ N (0d , γn2 Id ) is used by decoder i to infer ẑi , with the goal that ẑi = zi . Thus, we have the Markov chain for the i-th semantic variable as: p(zi , s, ck , xi , yi , ẑi ) = (1) SU p(s, zi )pCU k (ck |s)pi (xi |ck )p(yi |xi )p(ẑi |yi ). B. Problem Statement Previous work [1] shows that cooperative multi-tasking is constructive only when semantically related tasks are grouped together. This motivates the split-CU structure we are considering. Following the InfoMax principle, we maximize the mutual information between channel output Yi and corresponding semantic variable Zi to design encoder and decoders. That is, for fixed K, the joint clustering and InfoMax problem is: K X X maximize I(Yi ; Zi ) (2a) CU SU SU pCU 1 ,··· ,pK ,p1 ,··· ,pN ,T1 ,··· ,TK
subject to
k=1 i∈Tk K [
Tk = T ,
(2b)
Tk ∩ Tk′ = ∅,
∀k ̸= k ′ , (2c)
k=1
where I(Yi ; Zi ) and I(Yj ; Zj ) for i, j ∈ Tk , are coupled through the k-th CU, pCU k (ck |s), according to the Markov chain CU in (1). For simplicity in notation, we show pCU k (ck |s) by pk and SU SU also pi (xi |ck ) by pi , representing the distributions for the sub-CUs and SUs, respectively. We observe that constraints (2b) and (2c) ensure that the task clustering is a disjoint partition of T . This combines a variational optimization with a clustering problem. A further complication is that the optimal number of clusters K is not known a priori. Jointly optimizing the clustering structure and the variational distributions in (2) leads to a challenging mixed discrete-continuous optimization. Therefore, in the following section, we design a heuristic procedure to obtain a feasible solution to (2) with a high objective value. III. C ONSTRUCTIVE AND C OOPERATIVE M ULTI -TASKING We approach (2) and the optimal choice of K through a sequential multi-stage solution approach. First, we jointly determine K and T1 , . . . , TK with a semantics-aware clustering procedure after initial training epochs Einit based on
the CMT-SemCom training procedure from [1]. Then, we continue training, replacing the initial unified CU with K ⋆ CUs, exclusively within the task clusters. A. Semantic Clustering In the first stage, we identify semantically related tasks by directly analyzing the latent feature space of an initially trained unified CU. This allows us to fix the encoder topology before the main cooperative training begins. 1) Initial Training: To capture task similarities, we first train a single, unified CU parameterized by θu jointly with all N SUs for a short period of Einit epochs as in [1]. During this bootstrapping phase, the encoder learns to map the source observation s to a high-dimensional latent representation c = fθu (s) ∈ RdCU . To investigate task-relatedness through the CU outputs, we randomly select a representa(b) tive subset DiB = {(s(b) , zi )}B b=1 from the complete dataset (m) (m) Mi Di = {(s , zi )}m=1 of each task i ∈ T . We then define Ci as the corresponding set of CU outputs for the i-th task: n oB (b) (b) (b) (b) (b) Ci = ci | ci = fθu (si ) , (si , zi ) ∈ Di . (3) b=1 SN The collection C = i=1 Ci contains raw CU outputs that inherently encode task-relevant semantic information shaped by the InfoMax objective in [1]. 2) Clustering Procedure: We apply Hierarchical DensityBased Spatial Clustering of Applications with Noise (HDBSCAN) [18] to cluster tasks based on the extracted semantic features C. We note that, unlike common clustering algorithms such as K-Means, which require specifying the number and shape of the clusters (i.e., a Gaussian assumption in K-Means), HDBSCAN identifies the number of clusters K ⋆ inherently and assumes no shape for them. HDBSCAN builds a fully connected graph from all data points, whose edge weights are given by the mutual reachability distance (MRD). The MRD between two points is defined based on their Euclidean distance while accounting for the local density, such that it becomes larger than the Euclidean distance in low-density regions. Next, a minimum spanning tree is constructed from the graph, which is basically hierarchical clustering. Finally, HDBSCAN selects the clusters by maximizing a cluster stability measure, favoring dense clusters that are well separated from each other [18]. However, directly applying HDBSCAN to high-dimensional CU representations in RdCU is ineffective, as it requires more observed samples to produce enough density. To make the density more evident and make clustering easier for HDBSCAN, Uniform Manifold Approximation and Projection (UMAP) [17] has been used together with HDBSCAN as a dimensionality reduction technique [19]. UMAP preserves the global structure of the data and is designed specifically for clustering in addition to other applications. At a high level, UMAP first constructs a weighted kneighbour graph to represent the topology of the highdimensional data, where the weights are the distances between data points scaled by the size of a unit ball to the k-th nearest
neighbor. Next, an initial low-dimensional representation estimation is calculated, and an equivalent topological representation is constructed. UMAP then minimizes the cross-entropy between these two topological representations to minimize the error between the low-dimensional and the original highdimensional one [17]. We apply UMAP directly to the pooled CU outputs C. Let ΦUMAP : RdCU → Rdlow denote the learned non-linear projection, where dlow ≪ dCU . The projected latent features are given by: (b) (b) c̃i = ΦUMAP ci , ∀i ∈ T , b = 1, . . . , B. (4) SN ˜ ˜ We then apply HDBSCAN to the projected set C = i=1 Ci (b) (b) with C˜i = {c̃i }B ∈ {1, . . . , K ⋆ } denote the b=1 . Let ℓi cluster label assigned to sample b of task i. The final task to cluster mapping {T1⋆ , . . . , TK⋆ ⋆ } is obtained via majority voting over the ( sample-level assignments: ) B X (b) Tk⋆ = i ∈ T k = arg max I ℓi = k ′ , (5) k′ =1,...,K ⋆
b=1
(b)
(b)
for k = 1, . . . , K ⋆ , and I(ℓi = k ′ ) = 1 if ℓi = k ′ and ⋆ 0 otherwise. This yields a disjoint partition satisfying SK ⋆ ⋆ ⋆ ′ k=1 Tk = T (2b) and Tk ∩ Tk′ = ∅ for k ̸= k (2c). ⋆ Tasks assigned to the same cluster Tk are deemed semantically compatible for constructive cooperation. In the second stage, the training continues with dedicated sub-CU k assigned to each cluster, replacing the initial unified CU. B. Clustered-CMT-SemCom Learning Having the clusters (Tk⋆ ) and the number of sub-CUs (K ⋆ ) obtained in the first stage, we aim to solve the InfoMax problem by optimizing the SUs and sub-CUs in K ⋆ parallel learning steps. Therefore, the second-stage optimization problem is defined as: K⋆ X X maximize I(Yi ; Zi ) CU (6) pCU ⋆ 1 ,··· ,pK ⋆ , k=1 i∈Tk
SU pSU 1 ,··· ,pN
Expanding the mutual information in the optimization problem, as discussed in detail in [1], the approximated objective function is derived⋆ as: K X X L(θ, Φ, Ψ) = I(Yi ; Zi ) k=1 i∈Tk⋆
(7) }| z { K⋆ X X Ep(s,zi ) EpSU [ log qψi (zi |yi )] ≈ EpCU . ϕi θk ⋆
Task specific processing within each sub-CU
i∈Tk
k=1
|
{z
Intra-cluster cooperation through k-th Sub-CU
}
In (7), we employed the variational method using weights in neural networks (NNs) [20]. Thus, the posterior distributions SU are approximated by NNs, resulting in pCU θk and pϕi , where θk and ϕi represent the NN’s parameters approximating the k-th sub-CU and the i-th SU encoders, respectively. Although the i-th decoder, p(zi |yi ), can be fully determined using the known distributions, due to the high-dimensional integrals,
Initial training of Einit epochs
CU Enc.
SU Enc.
x 1 C y1
Dec.
ẑ1
SU Enc.
y2
Dec.
ẑ2
yN
Dec.
ẑN
.. .
SU Enc.
Clustering procedure
c
Clustered-CMTSemCom learning
h x2 na n e xN l
Clustering Procedure UMAP
{K ⋆ , T1⋆ , . . . , TK⋆ ⋆ } c1
.. .
⋆ CUK ⋆ Enc. cK
TABLE II: NN architectures used for the compared methods HDBSCAN
c̃
Extracted features
CU1 Enc.
.. .
(b)
ℓi
Task to cluster mapping
SU Enc.
x 1 C y1
Dec.
ẑ1
SU Enc.
x2 na y2
Dec.
ẑ2
SU Enc.
n e xN l
Dec.
ẑN
.. .
h
yN
TABLE I: List of simulation scenarios. Scenario Tasks Scenario A T1 (MNIST), T2 (EMNIST-I) , T4 (Fashion-MNIST-I) Scenario B T1 (MNIST), T2 (EMNIST-I), T3 (EMNIST-II) Scenario C T1 (MNIST), T4 (Fashion-MNIST-I), T5 (Fashion-MNIST-II)
.. .
Fig. 2: Clustered-CMT-SemCom system design illustration. we approximate it with qψi (zi |yi ), where ψi represents NNs approximating the distribution of the i-th decoder. Finally, we obtain the empirical estimate of the Clustered-CMT-SemCom objective function for training by approximating the expectations using Monte Carlo sampling, considering the training (m) (m) data set D = {(s(m) , z1 , . . . , zN )}M m=1 as done in [1]. The overall design of the proposed Clustered-CMT-SemCom framework is summarized and illustrated in Fig. 2. IV. S IMULATION R ESULTS To evaluate the performance of Clustered-CMT-SemCom, we compare its task execution error rate against two baselines: Individual training, corresponding to single-task training, and Unclustered-CMT-SemCom, MTL unaware of task-relatedness as in [1]. We note that the semantic-aware task clustering method in [16] reduces to Individual training in our considered scenarios since tasks label domains do not overlap. Therefore, it is implicitly included through the Individual training baseline. The simulation code of this work is also available online1 . A. Simulation Setup To demonstrate SN the effectiveness of our approach, we use the dataset D = i=1 Di and consider multinomial classification tasks, i.e., each semantic variable has a multinomial distribution Zi ∼ Multinomial. We define the first task T1 on the MNIST dataset of handwritten digits [21] as the classification of 0 to 9. For tasks T2 and T3 we consider the EMNIST dataset of handwritten letters [22] such that T2 (EMNIST-I) is the classification of A to J and T3 (EMNIST-II) is the classification of K to T . For T4 and T5 we use the FashionMNIST dataset of fashion items [23], such that T4 (FashionMNIST-I) is the classification of: pullover, dress, coat, sandals, bag, and T5 (Fashion-MNIST-II) is the classification of: tshirt, trousers, shirt, sneakers, boots. All datasets are grayscale images of size 28 × 28. 1 github.com/ant-uni-bremen/Semantic Aware Task Clustering
Layer Conv2D (# filters: F1 ), BatchNorm, ReLU, MaxPool2D, Dropout (0.1) CU Conv2D (# filters: F2 ), BatchNorm, ReLU, MaxPool2D, Dropout (0.1) Conv2D (# filters: F3 ), BatchNorm, ReLU, MaxPool2D Conv2D (# filters: F4 ), BatchNorm, ReLU Conv2D (# filters: F5 ), BatchNorm, ReLU, Flatten SU Dense (output size: 20), ReLU, Dense (output size: 16), ReLU Dense (output size: d), ReLU, Power normalization Dense (output size: 16), ReLU, Dense (output size: 16), ReLU Dec Dense (output size: number of classes for classification)
To assess the generality of the proposed framework, we evaluate it across different combinations of our defined tasks in three scenarios, with N = 3 in each, as detailed in Table I. For the semantic clustering stage, we set the initial training period Einit = 10 and B = 15 000 samples are used. The UMAP hyperparameters are specified as: the output dimension dlow = 2, min_dist=0, and n_neighbors=60. The HDBSCAN hyperparameters are set as: allow_single_cluster=True, min_samples=1, and min_cluster_size=1000. As the distance metric for both UMAP and HDBSCAN, we use the Euclidean distance. Furthermore, we merged those clusters containing large portions of samples of the same tasks if such a case exists. This means, if T1 and T2 are placed into two different clusters, both containing samples of the same tasks relatively equal, then we merged these clusters, concluding that T1 and T2 are semantically related and clustered together. The NN architectures in our simulations include Conv2dbased Convolutional Neural Network (CNN) layers with 3 × 3 convolution filters, a stride of 1, and zero-padding for the encoders, and fully connected layers for the decoders. Furthermore, 2D MaxPooling of size 2 × 2 for dimensionality reduction, BatchNorm2d, ReLU activations, max-pooling, and dropout layers are used, as summarized in Table II. We set the batch size to 100, and use a decaying learning rate strategy with an initial value of 10−3 . The CU and SU structures provided in Table II represent the overall system architecture with variables F = [F1 F2 F3 F5 F6 ], representing the number of convolution filters of the CNN. The adjustment of variable F is done so that the total number of trainable parameters remains approximately equal across the Clustered-CMT-SemCom, Unclustered-CMT-SemCom, and Individual training cases. As we have N = 3 tasks, there exist three possible task groupings. In case all tasks are clustered together, which is also the Unclustered-CMT-SemCom, we use F = [9, 9, 10, 6, 3], i.e., the CU for all three tasks followed by three SUs with filters F4 = 6 and F5 = 3, resulting in 8, 745 trainable parameters. When two tasks are clustered together and the third one is alone, we use F = [8, 8, 8, 6, 3] for the sub-CU with both clustered tasks and F = [6, 6, 6, 6, 3] for the separate task,
integrating the other sub-CU capacity into its SU. This results in a total of 8, 661 trainable parameters for all tasks. Finally, for Individual training, we use F = [6, 6, 6, 6, 3], integrating the unified CU capacity into the three separate SUs, resulting in 8, 661 trainable parameters for all tasks. All simulations consider an AWGN channel with d = 8 channel uses.
(a) Semantic-aware clustering Cluster 1
(b) Dataset-based clustering Cluster 1
Cluster 2
10
10 8
5 6 4
0
B. Evaluations In this section, we present first the performance of the proposed task clustering procedure, then the comparison of Clustered-CMT-SemCom with the aforementioned benchmarks with respect to task execution error rate. 1) Semantic-aware vs. Dataset-based Clustering: For each scenario, we are comparing the proposed clustering’s performance on two cases. It is executed on either (i) the output of the unified CU c, as proposed, named semantic-aware clustering, or (ii) the raw input image samples s, named Dataset-based clustering. As shown in Figs. 3 to 5, clustering performed on raw input data results in Unclustered-CMT-SemCom, where all tasks are jointly trained, for all scenarios. On the other hand, the semantic-aware clustering groups tasks differently in Scenarios A and C. We also observe that in Scenario B, both approaches yield the same clustering result. The impact of these clusters on task execution performance is evaluated in the next section. 2) Task Execution Performance: Here, we compare the performance of Clustered-CMT-SemCom, representing the semantic-aware clustering, against Unclustered-CMTSemCom, which coincides here with the dataset-based clustering result, where all tasks are jointly trained, and Individual training, which is simply the single-task training. As shown in Figs. 6a and 6c, the proposed Clustered-CMTSemCom achieves a significantly lower task execution error rate for all tasks compared to Unclustered-CMT-SemCom. Furthermore, the negative transfer is obvious in these scenarios, where Unclustered-CMT-SemCom performs worse than Individual training. This means MTL for all tasks together is not efficient. In addition, the Clustered-CMT-SemCom outperforms the Individual training on average, either performing better than or matching single-task training. We note that the match is because the corresponding task is clustered alone. For Scenario B, we consider an additional case where we group the tasks intuitively named as Intuitive clustering, where T2 and T3 are clustered together since they come from the same dataset, and T1 is trained individually. Fig. 6b shows that the proposed framework outperforms the Individual training and also the Intuitive clustering on average. As the semantic-aware clustering groups all tasks together, we observe that Unclustered-CMT-SemCom has the same performance as Clustered-CMT-SemCom. V. C ONCLUSION We proposed the Clustered-CMT-SemCom framework that uses semantic-aware clustering to mitigate destructive cooperation between unrelated tasks by formulating a sequential multistaged optimization problem. The proposed system adaptively
2 −5
0
5
T1 (MNIST)
10
2
4
6
T2 (EMNIST-I)
8
10
12
14
T4 (Fashion-MNIST-I)
Fig. 3: Semantic-aware vs. dataset-based: Scenario A. (a) Semantic-aware clustering
(b) Dataset-based clustering Cluster 1
10
Cluster 1
12 10
5
8 6
0
4
0
5
10
T1 (MNIST)
15
0
2
T2 (EMNIST-I)
4
6
8
10
T3 (EMNIST-II)
Fig. 4: Semantic-aware vs. dataset-based: Scenario B. (b) Dataset-based clustering
(a) Semantic-aware clustering Cluster 2 15
Cluster 1
8
Cluster 1
6 10
4 2
5
0 0 −5
−2 −5
0
T1 (MNIST)
5
10
15
−4
−2
T4 (Fashion-MNIST-I)
0
2
4
6
8
10
T5 (Fashion-MNIST-II)
Fig. 5: Semantic-aware vs. dataset-based: Scenario C. assigns semantically related tasks to shared subspaces, ensuring constructive cooperation. We demonstrated that our approach improves task performance in various scenarios over different datasets, where benchmarks like Individual training and Unclustered-CMT-SemCom suffer due to neglect of statistical relationships among semantic variables. R EFERENCES [1] A. Halimi Razlighi, C. Bockelmann, and A. Dekorsy, “Semantic communication for cooperative multi-task processing over wireless networks,” IEEE Wireless Communications Letters, vol. 13, no. 10, pp. 2867–2871, 2024. [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, and task-oriented communications,” IEEE Journal on Selected Areas in Communications, vol. 41, no. 1, pp. 5–41, 2023. [3] C. You, Y. Cai, Y. Liu, M. Di Renzo, T. M. Duman, A. Yener, and A. L. Swindlehurst, “Next generation advanced transceiver technologies for 6G and beyond,” IEEE Journal on Selected Areas in Communications, vol. 43, no. 3, pp. 582–627, 2025.
0.5
(a) Scenario A
(b) Scenario B
0.2
Error rate
T1 (MNIST)
0.2 Individual training Unclustered-CMT-SemCom Clustered-CMT-SemCom
0.1
0.05
T1 (MNIST) Individual training Unclustered-CMT-SemCom Intuitive clustering Clustered-CMT-SemCom
0.1 0.05 0.02
0.02 0.5 T2 (EMNIST-I)
Error rate
(c) Scenario C T1 (MNIST)
0.2
Individual training Unclustered-CMT-SemCom Clustered-CMT-SemCom
0.1 0.05 0.02 0.2
T2 (EMNIST-I)
0.2
T4 (Fashion-MNIST-I)
0.2 0.1
0.1
0.1 0.05
0.05 Error rate
0.2
0.2 0.2
T4 (Fashion-MNIST-I)
T5 (Fashion-MNIST-II)
T3 (EMNIST-II)
0.1
0.1
0.1 0.05
0.05 010 50
100 150 Epochs
200
250
010 50
100 150 Epochs
200
250
010 50
100 150 Epochs
200
250
Fig. 6: Task execution error rate over the training epochs for the three compared scenarios. [4] H. Li, S. Xie, J. Shao, Z. Wang, H. He, S. Song, J. Zhang, and K. B. Letaief, “Mutual information-empowered task-oriented communication: Principles, applications and challenges,” IEEE Communications Magazine, vol. 64, no. 4, pp. 164–171, 2026. [5] R. Caruana, “Multitask learning,” Machine learning, vol. 28, pp. 41–75, 1997. [6] Y. E. Sagduyu, T. Erpek, A. Yener, and S. Ulukus, “Multi - receiver task-oriented communications via multi - task deep learning,” in 2023 IEEE Future Networks World Forum (FNWF), 2023, pp. 1–6. [7] M. Gong, S. Wang, and S. Bi, “A scalable multi-device semantic communication system for multi-task execution,” in GLOBECOM 20232023 IEEE Global Communications Conference. IEEE, 2023, pp. 2227– 2232. [8] A. Halimi Razlighi, M. H. V. Tillmann, E. Beck, C. Bockelmann, and A. Dekorsy, “Cooperative and collaborative multi-task semantic communication for distributed sources,” in ICC 2025 - IEEE International Conference on Communications, 2025, pp. 3966–3971. [9] A. Halimi Razlighi, C. Bockelmann, and A. Dekorsy, “Semantic communication for cooperative multi-tasking over rate-limited wireless channels with implicit optimal prior,” IEEE Open Journal of the Communications Society, vol. 6, pp. 8523–8538, 2025. [10] T. Standley, A. Zamir, D. Chen, L. Guibas, J. Malik, and S. Savarese, “Which tasks should be learned together in multi-task learning?” in Proceedings of the 37th International Conference on Machine Learning, ser. ICML’20. JMLR.org, 2020. [11] Y. Zhang and Q. Yang, “A survey on multi-task learning,” IEEE Transactions on Knowledge and Data Engineering, vol. 34, no. 12, pp. 5586–5609, 2022. [12] Y. Wei, Z. Hu, L. Shen, Z. Wang, Y. Li, C. Yuan, and D. Tao, “Task groupings regularization: data-free meta-learning with heterogeneous pre-trained models,” in Proceedings of the 41st International Conference on Machine Learning, ser. ICML’24. JMLR.org, 2024. [13] A. Ayman, A. Mukhopadhyay, and A. Laszka, “Task grouping for automated multi-task machine learning via task affinity prediction,” arXiv preprint arXiv:2310.16241, 2023.
[14] Z. Kang, K. Grauman, and F. Sha, “Learning with whom to share in multi-task feature learning,” in Proceedings of the 28th International Conference on International Conference on Machine Learning, ser. ICML’11. Madison, WI, USA: Omnipress, 2011, p. 521–528. [15] C. Fifty, E. Amid, Z. Zhao, T. Yu, R. Anil, and C. Finn, “Efficiently identifying task groupings for multi-task learning,” in Proceedings of the 35th International Conference on Neural Information Processing Systems. Red Hook, NY, USA: Curran Associates Inc., 2021. [16] A. H. Razlighi, P. Dhingra, E. Beck, B. Matthiesen, and A. Dekorsy, “Semantic-aware task clustering for federated cooperative multitask semantic communication,” 2026. [Online]. Available: https: //arxiv.org/abs/2601.17419 [17] L. McInnes, J. Healy, and J. Melville, “UMAP: Uniform manifold approximation and projection for dimension reduction,” arXiv preprint arXiv:1802.03426, 2018. [18] R. J. G. B. Campello, D. Moulavi, and J. Sander, “Density-based clustering based on hierarchical density estimates,” in Advances in Knowledge Discovery and Data Mining, J. Pei, V. S. Tseng, L. Cao, H. Motoda, and G. Xu, Eds. Berlin, Heidelberg: Springer Berlin Heidelberg, 2013, pp. 160–172. [19] M. Ali, M. W. Jones, X. Xie, and M. Williams, “Timecluster: dimension reduction applied to temporal data for visual analytics,” The Visual Computer, vol. 35, no. 6, pp. 1013–1026, 2019. [20] D. P. Kingma and M. Welling, “Auto-encoding variational bayes,” arXiv preprint arXiv:1312.6114, 2013. [21] L. Deng, “The mnist database of handwritten digit images for machine learning research,” IEEE Signal Processing Magazine, vol. 29, no. 6, pp. 141–142, 2012. [22] G. Cohen, S. Afshar, J. Tapson, and A. Van Schaik, “Emnist: Extending mnist to handwritten letters,” in 2017 international joint conference on neural networks (IJCNN). IEEE, 2017, pp. 2921–2926. [23] H. Xiao, K. Rasul, and R. Vollgraf, “Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,” arXiv preprint arXiv:1708.07747, 2017.