Semantic-Aware Task Clustering for Constructive and Cooperative Multi-Tasking

Ahmad Halimi Razlighi; Armin Dekorsy; Bho Matthiesen; Edgar Beck; Maximilian H. V. Tillmann

REVIEW 5 major objections 4 minor 23 references

Grouping tasks by shared-encoder features prevents harmful interference in multi-task semantic communication.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 07:27 UTC pith:6SGE27L2

load-bearing objection A plausible, clearly written incremental extension that groups SemCom tasks by clustering CU features, but the empirical case needs error bars and a quantified merge rule before 'ensures constructive cooperation' can be trusted. the 5 major comments →

arxiv 2607.21426 v1 pith:6SGE27L2 submitted 2026-07-23 cs.LG cs.ITeess.SPmath.ITstat.ML

Semantic-Aware Task Clustering for Constructive and Cooperative Multi-Tasking

Ahmad Halimi Razlighi , Maximilian H. V. Tillmann , Edgar Beck , Bho Matthiesen , Armin Dekorsy This is my paper

classification cs.LG cs.ITeess.SPmath.ITstat.ML

keywords semantic communicationmulti-task learningtask clusteringnegative transferdestructive cooperationUMAPHDBSCANInfoMax

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper asks which tasks should share a semantic encoder in a multi-task wireless communication system. Its claim: after a short joint training phase, the intermediate outputs of a shared encoder already contain enough information to tell semantically related tasks apart, and clustering those outputs with UMAP and HDBSCAN yields a task partition that avoids destructive cooperation. Training separate sub-encoders for each discovered cluster then improves accuracy relative to both unclustered multi-task training and single-task training in the tested scenarios. The contribution matters because semantic communication systems must serve many downstream tasks, and the answer to 'who shares what' is currently either assumed or left to costly search.

Core claim

The paper proposes Clustered-CMT-SemCom: train one unified common unit (CU) briefly with all tasks, collect its output features for a sample of each task, embed those features with UMAP, cluster them with HDBSCAN, and assign each task to a cluster by majority vote. This partition is fixed; then the unified CU is replaced by one sub-CU per cluster, and end-to-end InfoMax training continues only within clusters. The authors report that this pipeline separates tasks that dataset-based clustering would group together, and that the resulting error rates are lower than unclustered multi-task learning and at least as good as individual training in each of the three scenarios.

What carries the argument

UMAP + HDBSCAN on CU output features after E_init = 10 epochs of unified training, followed by majority-vote task assignment and intra-cluster end-to-end training. UMAP projects high-dimensional encoded features to two dimensions for density-based clustering; HDBSCAN chooses the number of clusters automatically based on cluster stability. A manual merge step combines clusters that split samples of the same tasks. The InfoMax objective (maximizing mutual information between received signal and semantic variable) is what couples tasks in a cluster through a shared sub-CU.

Load-bearing premise

The method hinges on the assumption that after only ten epochs of joint training, the unified encoder's output features already contain the task structure that will persist for the rest of training, so a single, early clustering can fix the partitions.

What would settle it

Run the identical pipeline but vary the initial training length (e.g., E_init in {1, 5, 10, 50, 100}) and record whether the resulting clusters and final error rates change materially; if the partition is not stable across E_init, the one-shot early clustering is not trustworthy.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

If valid, task grouping in multi-task semantic communication can be decided by a lightweight, one-time clustering step instead of exhaustive search or gradient-based affinity.
Negative transfer disappears: no task in the clustered system performs worse than its individually trained counterpart, and most improve.
Raw data distributions are insufficient proxies for task relatedness; the shared encoder's output space is where semantic relatedness shows up.
The clustered system retains most of the benefit of multi-task learning while isolating unrelated tasks, so it should generalize to other task sets where dataset-level grouping is misleading.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The paper does not test whether the 10-epoch clustering remains valid later in training; a periodic re-clustering might change the partition and could improve or destabilize results.
The manual merging of clusters is a subjective post-hoc step; an automatic criterion (e.g., HDBSCAN's own stability or a mutual-information check) would make the method more reproducible at scale.
The claim that semantic-aware clustering captures higher-order semantic relationships is inferred from the feature space; a direct information-theoretic measure of cluster quality beyond downstream accuracy would test that interpretation.
A natural next experiment is to use tasks with partially overlapping semantics (e.g., digits and letters sharing visual strokes) and see whether clusters reflect graded similarity rather than binary grouping.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

A plausible, clearly written incremental extension that groups SemCom tasks by clustering CU features, but the empirical case needs error bars and a quantified merge rule before 'ensures constructive cooperation' can be trusted.

read the letter

Take it as a useful engineering paper, not a breakthrough. The new bit is applying UMAP+HDBSCAN to the shared CU's latent outputs after a short warm-up, then fixing clusters and training separate sub-CUs. That is a modest but sensible extension of the authors' own CMT-SemCom and their federated clustering paper, and it is clearly described. Code is available, which helps.

What the paper does well: it shows that raw-input clustering fails to split tasks in their scenarios while CU-feature clustering does, and the error curves in Figs. 6a and 6c show clustered multi-task training at or below individual baselines. So the idea has legs. The InfoMax reformulation in (7) is standard variational estimation, but it is applied cleanly.

Soft spots, in order of importance. First, there are no error bars, no multiple seeds, and no statistical tests; the curves may be single runs. Second, the SNR is never stated, which is odd for a wireless communication paper. Third, the manual cluster merge rule in Section IV-A is post hoc and outside the algorithm in Section III-A; the paper gives no data on how often it fires or how much it changes results. That weakens the attribution of the gains to semantic feature clustering. Fourth, there is no sensitivity analysis for E_init, B, or the UMAP/HDBSCAN hyperparameters; everything is fixed, and one-shot clustering at 10 epochs could be lucky. Fifth, only three small grayscale classification scenarios are considered, so the generality claim is thin.

The reader's stress-test concern about E_init is legitimate and unaddressed, and the merge rule can indeed mask clustering failures. Still, the clustering itself is unsupervised and the reported task accuracy is not used to fit constants, so the circularity burden is mild rather than fatal.

This paper is for people working on task grouping in semantic communication or on practical MTL for edge-AI systems; the general MTL audience will not learn much from it. It deserves a serious referee, not a desk reject, because the approach is plausible, clearly presented, and reproducible in principle. I would send it to peer review with the expectation of major revision: add statistical rigor, report the SNR, and quantify the merge rule. If those are addressed, the contribution is solid for its niche.

Referee Report

5 major / 4 minor

Summary. The paper proposes Clustered-CMT-SemCom, a two-stage pipeline for cooperative multi-task semantic communication. In stage one, a single unified CU is trained for a short initialization period (E_init=10 epochs); the CU output features are then embedded with UMAP and clustered with HDBSCAN to discover semantically compatible task groups. In stage two, training continues end-to-end within each discovered cluster using a dedicated sub-CU. The authors formulate the joint clustering/InfoMax problem (2), relax it into a heuristic sequential procedure, and evaluate the method on three scenarios built from MNIST, EMNIST, and Fashion-MNIST, claiming lower error rates than unclustered multi-tasking and individual training baselines.

Significance. If the empirical claims hold, the paper would provide a practical way to avoid destructive cooperation in multi-task semantic communication by automatically discovering which tasks should share a common encoder. The idea is well motivated by the authors' prior InfoMax formulation, and the use of UMAP+HDBSCAN on CU features is a reasonable lightweight alternative to gradient-based task grouping. The paper ships a code link, uses standard baseline comparisons, and states its hyperparameters, which aids reproducibility. However, the current evidence is not strong enough to support the central claim: the evaluation lacks statistical rigor, the clustering stage has an ad-hoc manual override, and the key assumption that a one-shot clustering after 10 epochs is reliable is unvalidated.

major comments (5)

[§IV-A, Fig. 6] The paper reports that Clustered-CMT-SemCom achieves 'significantly lower' error rates than the baselines, but no SNR is specified (only 'd=8 channel uses' is given), no number of random seeds is reported, no error bars are shown, and no statistical tests are performed. The reported curves are single-run trajectories, so the central empirical claim is not reproducible or quantitatively supported. Please specify the channel SNR/noise variance, repeat experiments over multiple seeds, and report means with confidence intervals or significance tests. Also state whether the clustering stage is evaluated at the same SNR as training.
[§IV-A] The manual cluster-merging rule introduced in the simulation setup is not part of the algorithm in §III-A. It is described only as merging clusters that 'contain large portions of samples of the same tasks relatively equal', with no threshold, no algorithmic definition, and no report of how often it fired in the three scenarios. This post-hoc rule can override the HDBSCAN output and makes it impossible to determine whether the reported gains are due to semantic-aware clustering or to manual correction. The rule must be formalized, its activation reported, and the experiments run without it to isolate the method's contribution.
[§III-A-1, §IV-A] The entire pipeline rests on the assumption that after only E_init=10 epochs of unified-CU training the UMAP+HDBSCAN clustering identifies the correct task partition for all subsequent training. No sensitivity analysis is provided for E_init, nor for the UMAP/HDBSCAN hyperparameters (n_neighbors, min_dist, min_cluster_size), and no re-clustering or multiple clustering initializations are considered. Since the first stage fixes the encoder topology, an unstable or incorrect clustering would invalidate the claimed guarantee of constructive cooperation. Please provide an ablation across E_init and the main clustering hyperparameters, and report cluster stability over training epochs or bootstrap resamples.
[§IV-B-1, Figs. 3-5] Clustering quality is assessed only qualitatively via 2D scatter plots. The paper does not report quantitative cluster-quality metrics such as adjusted Rand index against task labels (the ground truth is known), silhouette score, or bootstrap stability. Without such metrics, the claim that semantic-aware clustering discovers meaningful semantic groups rather than artifacts of UMAP/HDBSCAN parameter choices is not established. Please add quantitative clustering evaluation for each scenario and, ideally, compare to a simpler baseline such as clustering raw input features or model parameters.
[§IV, Table I] All evaluations use N=3 tasks per scenario, and the three scenarios are manually selected and are also the ones used to design the heuristic. With N=3, exhaustive search over all possible task partitions is trivial, so the paper does not demonstrate that the proposed clustering is needed or that it scales. The authors should add at least one scenario with N>3 and/or a held-out set of tasks, and compare against exhaustive grouping or one of the existing task-grouping methods cited in [12]-[15]. Otherwise the general claim that semantic-aware clustering 'effectively mitigates destructive cooperation' is only weakly supported.

minor comments (4)

[Eq. (5)] HDBSCAN can assign a noise label (typically -1) to points in low-density regions. Equation (5) defines ℓ_i^{(b)} in {1,...,K*}, so the handling of noise labels is not specified. Please clarify how noise points are treated in the majority-vote step.
[§IV-A, Table II] The text defines F=[F1 F2 F3 F5 F6], but the table and subsequent discussion mention F4; either the vector or the table index is inconsistent. Please correct.
[§IV-B-2, Fig. 6b] The 'Intuitive clustering' baseline is introduced without a precise definition of how T2 and T3 are grouped and how the architecture is allocated in that case. Please state the exact partition and parameter allocation used.
[§IV-B-2] The sentence 'the semantic-aware task clustering method in [16] reduces to Individual training in our considered scenarios' is not explained. Since [16] is not implemented, this claim should either be justified or removed.

Circularity Check

0 steps flagged

No significant circularity: the claimed clustering gains are not forced by construction or self-citation.

full rationale

The paper's derivation chain is empirical and non-circular. The clustering stage (UMAP+HDBSCAN on CU output features after E_init=10 epochs) is unsupervised: it takes the features f_theta_u(s) from a short initial training phase and produces a disjoint task partition via majority voting. The final task accuracy is not used to fit the clustering constants or to select clusters; the reported gains come from a subsequent end-to-end training stage within the discovered groups. The manual merge rule in Section IV-A is a post-hoc heuristic that adjusts HDBSCAN output based on sample-to-task overlap, which weakens the claim that the grouping is fully automatic, but it does not make the accuracy comparison circular: the merge rule does not optimize task error and the subsequent performance is still measured on the resulting partition. The paper's reliance on prior self-citations, especially [1] for the InfoMax formulation and the constructive/destructive cooperation premise, is ordinary incremental work; those cited results are external publications and the current paper also independently demonstrates negative transfer (Unclustered-CMT-SemCom worse than Individual training). No equation is identical to another by construction, no fitted parameter is renamed as a prediction, and no uniqueness theorem is imported from the authors to force the choice. The main weaknesses (one-shot clustering at a fixed E_init, no sensitivity analysis, manual cluster merging) are robustness/validation concerns, not circularity.

Axiom & Free-Parameter Ledger

9 free parameters · 6 axioms · 0 invented entities

The central claim rests on a heuristic clustering pipeline with several hand-set hyperparameters and a post-hoc merge rule; the InfoMax objective (2) is not shown to be optimized by the clustering stage. No new physical entities are introduced.

free parameters (9)

E_init = 10 epochs
Duration of initial unified-CU training before clustering; set by hand, no sensitivity analysis.
B = 15,000 samples per task
Size of representative subset used for clustering; chosen by hand.
UMAP output dimension d_low = 2
Embedding dimension for HDBSCAN; chosen by hand.
UMAP min_dist = 0
UMAP hyperparameter; chosen by hand, affects local density.
UMAP n_neighbors = 60
UMAP hyperparameter; chosen by hand, controls local/global balance.
HDBSCAN min_cluster_size = 1000
HDBSCAN hyperparameter; directly limits minimum cluster size and affects discovered K*.
HDBSCAN min_samples = 1
HDBSCAN hyperparameter; chosen by hand, shapes density threshold.
Cluster merge rule = unquantified ('relatively equal')
Manual post-hoc merging of clusters with similar task composition; no formal criterion stated.
AWGN noise variance / SNR = not specified
The paper says 'AWGN channel with d=8 channel uses' but never states the SNR or noise power; error rates cannot be interpreted or reproduced from text alone.

axioms (6)

domain assumption Cooperative multi-tasking is constructive only when semantically related tasks share a CU (from [1]).
This premise motivates the whole clustering stage; it is cited from the authors' prior work, not re-derived here.
domain assumption The Markov chain p(zi,s,ck,xi,yi,ẑi)=p(s,zi)p^CU_k(ck|s)p^SU_i(xi|ck)p(yi|xi)p(ẑi|yi) in Eq. (1) holds.
Standard conditional independence structure assumed for the split encoder and AWGN channel.
ad hoc to paper After E_init=10 epochs, the unified CU output f_θu(s) encodes enough task-related structure for clustering.
No sensitivity analysis or theoretical argument is given; this is load-bearing for the clustering stage.
ad hoc to paper UMAP projection preserves the task-relevant cluster structure so that HDBSCAN can recover semantic groups.
Invoked from [17], [19] but not validated for CU outputs.
ad hoc to paper HDBSCAN's stability-maximizing clusters, after majority voting, identify semantically compatible task partitions.
No formal link to the InfoMax objective; the paper itself calls the procedure heuristic.
domain assumption Variational NN approximations and Monte Carlo sampling accurately estimate the mutual information objective (7).
Standard practice from [20] and [1]; accepted without error analysis.

pith-pipeline@v1.3.0-alltime-deepseek · 9342 in / 12401 out tokens · 112913 ms · 2026-08-01T07:27:24.378311+00:00 · methodology

0 comments

read the original 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 semantic-aware 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.

Figures

Figures reproduced from arXiv: 2607.21426 by Ahmad Halimi Razlighi, Armin Dekorsy, Bho Matthiesen, Edgar Beck, Maximilian H. V. Tillmann.

**Figure 2.** Figure 2: Clustered-CMT-SemCom system design illustration. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: Semantic-aware vs. dataset-based: Scenario A. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 4.** Figure 4: Semantic-aware vs. dataset-based: Scenario B. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

**Figure 5.** Figure 5: Semantic-aware vs. dataset-based: Scenario C. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗

**Figure 6.** Figure 6: Task execution error rate over the training epochs for the three compared scenarios. [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

23 extracted references · 4 linked inside Pith

[1]

Semantic commu- nication for cooperative multi-task processing over wireless networks,

A. Halimi Razlighi, C. Bockelmann, and A. Dekorsy, “Semantic commu- nication for cooperative multi-task processing over wireless networks,” IEEE Wireless Communications Letters, vol. 13, no. 10, pp. 2867–2871, 2024

2024
[2]

Beyond transmitting bits: Context, semantics, and task-oriented communications,

D. G ¨und¨uz, 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

2023
[3]

Next generation advanced transceiver technologies for 6G and beyond,

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. (a) Scenario A MNIST0.02 0.05 0.1 0.2 0.5 T1 (MNIST) Error rate Individual training Unclustered-CMT-SemCom Clustered...

2025
[4]

Mutual information-empowered task-oriented communication: Principles, applications and challenges,

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 Maga- zine, vol. 64, no. 4, pp. 164–171, 2026

2026
[5]

Multitask learning,

R. Caruana, “Multitask learning,”Machine learning, vol. 28, pp. 41–75, 1997

1997
[6]

Multi - receiver task-oriented communications via multi - task deep learning,

Y . E. Sagduyu, T. Erpek, A. Yener, and S. Ulukus, “Multi - receiver task-oriented communications via multi - task deep learning,” in2023 IEEE Future Networks World F orum (FNWF), 2023, pp. 1–6

2023
[7]

A scalable multi-device semantic communication system for multi-task execution,

M. Gong, S. Wang, and S. Bi, “A scalable multi-device semantic communication system for multi-task execution,” inGLOBECOM 2023- 2023 IEEE Global Communications Conference. IEEE, 2023, pp. 2227– 2232

2023
[8]

Cooperative and collaborative multi-task semantic com- munication for distributed sources,

A. Halimi Razlighi, M. H. V . Tillmann, E. Beck, C. Bockelmann, and A. Dekorsy, “Cooperative and collaborative multi-task semantic com- munication for distributed sources,” inICC 2025 - IEEE International Conference on Communications, 2025, pp. 3966–3971

2025
[9]

Semantic commu- nication for cooperative multi-tasking over rate-limited wireless channels with implicit optimal prior,

A. Halimi Razlighi, C. Bockelmann, and A. Dekorsy, “Semantic commu- nication 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

2025
[10]

Which tasks should be learned together in multi-task learning?

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

2020
[11]

A survey on multi-task learning,

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

2022
[12]

Task groupings regularization: data-free meta-learning with heterogeneous pre-trained models,

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,” inProceedings of the 41st International Conference on Machine Learning, ser. ICML’24. JMLR.org, 2024

2024
[13]

Task grouping for automated multi-task machine learning via task affinity prediction,

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

Pith/arXiv arXiv 2023
[14]

Learning with whom to share in multi-task feature learning,

Z. Kang, K. Grauman, and F. Sha, “Learning with whom to share in multi-task feature learning,” inProceedings of the 28th International Conference on International Conference on Machine Learning, ser. ICML’11. Madison, WI, USA: Omnipress, 2011, p. 521–528

2011
[15]

Efficiently identifying task groupings for multi-task learning,

C. Fifty, E. Amid, Z. Zhao, T. Yu, R. Anil, and C. Finn, “Efficiently identifying task groupings for multi-task learning,” inProceedings of the 35th International Conference on Neural Information Processing Systems. Red Hook, NY , USA: Curran Associates Inc., 2021

2021
[16]

Semantic-aware task clustering for federated cooperative multi- task semantic communication,

A. H. Razlighi, P. Dhingra, E. Beck, B. Matthiesen, and A. Dekorsy, “Semantic-aware task clustering for federated cooperative multi- task semantic communication,” 2026. [Online]. Available: https: //arxiv.org/abs/2601.17419

arXiv 2026
[17]

UMAP: Uniform manifold approximation and projection for dimension reduction,

L. McInnes, J. Healy, and J. Melville, “UMAP: Uniform manifold approximation and projection for dimension reduction,”arXiv preprint arXiv:1802.03426, 2018

Pith/arXiv arXiv 2018
[18]

Density-based clustering based on hierarchical density estimates,

R. J. G. B. Campello, D. Moulavi, and J. Sander, “Density-based clustering based on hierarchical density estimates,” inAdvances 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

2013
[19]

Timecluster: dimension reduction applied to temporal data for visual analytics,

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

2019
[20]

Auto-encoding variational bayes,

D. P. Kingma and M. Welling, “Auto-encoding variational bayes,”arXiv preprint arXiv:1312.6114, 2013

Pith/arXiv arXiv 2013
[21]

The mnist database of handwritten digit images for machine learning research,

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

2012
[22]

Emnist: Extending mnist to handwritten letters,

G. Cohen, S. Afshar, J. Tapson, and A. Van Schaik, “Emnist: Extending mnist to handwritten letters,” in2017 international joint conference on neural networks (IJCNN). IEEE, 2017, pp. 2921–2926

2017
[23]

Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,

H. Xiao, K. Rasul, and R. V ollgraf, “Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,”arXiv preprint arXiv:1708.07747, 2017

Pith/arXiv arXiv 2017

[1] [1]

Semantic commu- nication for cooperative multi-task processing over wireless networks,

A. Halimi Razlighi, C. Bockelmann, and A. Dekorsy, “Semantic commu- nication for cooperative multi-task processing over wireless networks,” IEEE Wireless Communications Letters, vol. 13, no. 10, pp. 2867–2871, 2024

2024

[2] [2]

Beyond transmitting bits: Context, semantics, and task-oriented communications,

D. G ¨und¨uz, 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

2023

[3] [3]

Next generation advanced transceiver technologies for 6G and beyond,

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. (a) Scenario A MNIST0.02 0.05 0.1 0.2 0.5 T1 (MNIST) Error rate Individual training Unclustered-CMT-SemCom Clustered...

2025

[4] [4]

Mutual information-empowered task-oriented communication: Principles, applications and challenges,

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 Maga- zine, vol. 64, no. 4, pp. 164–171, 2026

2026

[5] [5]

Multitask learning,

R. Caruana, “Multitask learning,”Machine learning, vol. 28, pp. 41–75, 1997

1997

[6] [6]

Multi - receiver task-oriented communications via multi - task deep learning,

Y . E. Sagduyu, T. Erpek, A. Yener, and S. Ulukus, “Multi - receiver task-oriented communications via multi - task deep learning,” in2023 IEEE Future Networks World F orum (FNWF), 2023, pp. 1–6

2023

[7] [7]

A scalable multi-device semantic communication system for multi-task execution,

M. Gong, S. Wang, and S. Bi, “A scalable multi-device semantic communication system for multi-task execution,” inGLOBECOM 2023- 2023 IEEE Global Communications Conference. IEEE, 2023, pp. 2227– 2232

2023

[8] [8]

Cooperative and collaborative multi-task semantic com- munication for distributed sources,

A. Halimi Razlighi, M. H. V . Tillmann, E. Beck, C. Bockelmann, and A. Dekorsy, “Cooperative and collaborative multi-task semantic com- munication for distributed sources,” inICC 2025 - IEEE International Conference on Communications, 2025, pp. 3966–3971

2025

[9] [9]

Semantic commu- nication for cooperative multi-tasking over rate-limited wireless channels with implicit optimal prior,

A. Halimi Razlighi, C. Bockelmann, and A. Dekorsy, “Semantic commu- nication 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

2025

[10] [10]

Which tasks should be learned together in multi-task learning?

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

2020

[11] [11]

A survey on multi-task learning,

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

2022

[12] [12]

Task groupings regularization: data-free meta-learning with heterogeneous pre-trained models,

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,” inProceedings of the 41st International Conference on Machine Learning, ser. ICML’24. JMLR.org, 2024

2024

[13] [13]

Task grouping for automated multi-task machine learning via task affinity prediction,

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

Pith/arXiv arXiv 2023

[14] [14]

Learning with whom to share in multi-task feature learning,

Z. Kang, K. Grauman, and F. Sha, “Learning with whom to share in multi-task feature learning,” inProceedings of the 28th International Conference on International Conference on Machine Learning, ser. ICML’11. Madison, WI, USA: Omnipress, 2011, p. 521–528

2011

[15] [15]

Efficiently identifying task groupings for multi-task learning,

C. Fifty, E. Amid, Z. Zhao, T. Yu, R. Anil, and C. Finn, “Efficiently identifying task groupings for multi-task learning,” inProceedings of the 35th International Conference on Neural Information Processing Systems. Red Hook, NY , USA: Curran Associates Inc., 2021

2021

[16] [16]

Semantic-aware task clustering for federated cooperative multi- task semantic communication,

A. H. Razlighi, P. Dhingra, E. Beck, B. Matthiesen, and A. Dekorsy, “Semantic-aware task clustering for federated cooperative multi- task semantic communication,” 2026. [Online]. Available: https: //arxiv.org/abs/2601.17419

arXiv 2026

[17] [17]

UMAP: Uniform manifold approximation and projection for dimension reduction,

L. McInnes, J. Healy, and J. Melville, “UMAP: Uniform manifold approximation and projection for dimension reduction,”arXiv preprint arXiv:1802.03426, 2018

Pith/arXiv arXiv 2018

[18] [18]

Density-based clustering based on hierarchical density estimates,

R. J. G. B. Campello, D. Moulavi, and J. Sander, “Density-based clustering based on hierarchical density estimates,” inAdvances 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

2013

[19] [19]

Timecluster: dimension reduction applied to temporal data for visual analytics,

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

2019

[20] [20]

Auto-encoding variational bayes,

D. P. Kingma and M. Welling, “Auto-encoding variational bayes,”arXiv preprint arXiv:1312.6114, 2013

Pith/arXiv arXiv 2013

[21] [21]

The mnist database of handwritten digit images for machine learning research,

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

2012

[22] [22]

Emnist: Extending mnist to handwritten letters,

G. Cohen, S. Afshar, J. Tapson, and A. Van Schaik, “Emnist: Extending mnist to handwritten letters,” in2017 international joint conference on neural networks (IJCNN). IEEE, 2017, pp. 2921–2926

2017

[23] [23]

Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,

H. Xiao, K. Rasul, and R. V ollgraf, “Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,”arXiv preprint arXiv:1708.07747, 2017

Pith/arXiv arXiv 2017