Pith. sign in

REVIEW 4 major objections 5 minor 40 references

Multi-Task Semantic Communication With Graph Attention-Based Feature Correlation Extraction

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Pith's one-line read This paper claims that encoding-block correlations are a compressible resource that multi-task semantic communication systems currently waste.

desk verdict The GAI architecture is a plausible and well-ablated multi-task feature fusion idea, but the paper never explains how the per-task transmitted features are multiplexed, so the headline bandwidth-constrained gains are not yet supported. read the letter →

arxiv 2501.02006 v1 pith:KL3NYGTJ submitted 2025-01-02 cs.LG cs.AI

classification cs.LGcs.AI
keywords multi-tasklearningsemanticcommunicationgraphattentionjointsource-channelcodingfeaturecorrelationbandwidth-constrainedtransmissiontask-nodeweighting
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper is trying to establish that a multi-task semantic communication transmitter can send less and decode better if it exploits the correlations among its own intermediate encoder features. It adds a graph attention inter-block module that treats each encoder block's output as a node, refines the nodes through graph attention, and learns per-task weights that reassemble the block features into the transmitted representation. The paper reports that this module improves average task performance over the strongest available baselines by 11.4% on CityScapes 2-task and 3.97% on NYU v2 3-task at a bandwidth ratio of $1/12$ (the compression level for transmission). It also reports that the advantage grows as the channel becomes noisier.

What carries the argument

The central mechanism is the graph attention inter-block (GAI) module, a three-layer add-on to the encoder. A Feature Transformation Layer normalizes each residual block's output $F_i$ into a graph node representation $V_i^0$ plus an interpolated tensor $K_i$; a Graph Attention Layer updates all node representations for $M$ iterations using attention coefficients $a_{i,j}^m$; and a Relation Mapping Layer turns each refined node into task-node weights $e_{i,t}$ through a two-layer multi-layer perceptron (a small neural network). The transmitted feature for task $t$ is $z_t = \sum_{i=1}^{N} e_{i,t} K_i$, so the correlation structure among blocks directly sets what each task's decoder receives.

What would settle it

Train GAI and its strongest baseline directly under the deployment conditions, for instance a signal-to-noise ratio of $-2$ dB and a bandwidth ratio of $1/12$, instead of training clean and then testing noisy; if GAI's margin collapses, the claimed robustness comes from the evaluation protocol rather than from graph attention. As a second check, replace the learned attention coefficients with random fixed values and rerun the NYU v2 3-task experiment; a retained margin would show correlation learning is not the source.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that attention across the outputs of an encoder's residual blocks gives a task-oriented transmitter a way to specialize a shared representation without extra decoders. Instead of transmitting only the last block's features, the GAI module computes task-node weights $e_{i,t}$ for every block $i$ and task $t$, and transmits $z_t = \sum_i e_{i,t} K_i$. The paper argues that this weighted fusion preserves details from earlier blocks that later blocks compress away, and it supports the argument by beating single-task, multi-task, feature-sharing, and dynamic-network baselines on six datasets while keeping the encoder's cost roughly flat in the number of tasks.

Load-bearing premise

Section IV.B.3 trains with plenty of bandwidth and almost no channel noise but evaluates with heavy noise and tight bandwidth, so the reported robustness assumes that clean-first protocol is a fair comparison.

Editorial extensions

If this is right

  • Existing decoders can stay unchanged; the gain comes from the transmitter's representation, so the module can be added to current multi-task encoder-decoder links.
  • The value of graph attention grows with the number of tasks; the ablation shows removing node updates costs 0.60% at two tasks and 4.24% at five tasks, so richer task sets benefit more from correlation modeling.
  • Extreme bandwidth limitation is where the module pays off; the lead over the strongest baseline on CityScapes grows from 3.93% at $R = 1/4$ to 11.40% at $R = 1/12$.
  • Noisy channels do not erase the gain; average relative improvements are slightly larger at the noisier end of the tested range and hold over a Rayleigh fading channel.
  • The encoder's computational cost stays about the same as tasks are added, at 19.73 GFLOPs for the 2-, 3-, and 5-task settings.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same correlation-as-compression idea could apply to other shared-representation systems where bandwidth, not model capacity, is the bottleneck, such as edge inference or distributed sensing.
  • The learned task-node weights suggest later encoder blocks dominate for two-task settings while shallow blocks become more influential for five tasks; a future system could hard-select which blocks to transmit to cut latency further.
  • Because training is clean while testing is noisy, retraining the model under deployment noise is a direct way to separate the architecture's benefit from the protocol's benefit.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a Graph Attention Inter-block (GAI) module for the encoder of a multi-task semantic communication system. The module treats the outputs of the encoder's residual blocks as graph nodes, updates their representations with a graph attention network, and generates task-specific weights through a relation-mapping MLP; these weights are fused with interpolated block features to form transmitted features. The authors evaluate on CityScapes 2Task, NYU v2 2/3Task, TaskonomyTiny 5Task, Oxford-IIIT Pet, and MVSA, reporting gains over several multi-task learning baselines, with headline results at a bandwidth ratio of R=1/12.

Significance. If the reported results are reproducible, the paper would offer a computationally light way to exploit inter-block feature correlations in multi-task joint source-channel coding. The ablations (GAI-w, SimpAtt) support the view that both the graph-based updating and the task-specific weighting contribute to the gains, and the complexity analysis is clearly presented. The stability analysis (Tables XIII-XV) is a useful addition. However, the headline bandwidth-constrained claims are not currently substantiated because the channel-use accounting is ambiguous and the reported improvement percentages are not margins over the best baseline.

major comments (4)
  1. [Section III.C, Eq. (14); Section II.C, Eq. (4); Algorithm 1] The manuscript defines a separate transmitted feature z_t for each task t in Eq. (14) and Algorithm 1, but the channel model and the bandwidth ratio R are defined for a single encoder output z in Section II.C. The paper never specifies how the T tensors z_1,...,z_T are multiplexed into the single channel described by Eq. (2). If they are transmitted as T separate tensors, the actual number of channel symbols is T·k, so the effective bandwidth ratio is T·R, not R. For CityScapes 2Task at a claimed R=1/12, the true ratio would be 1/6; for NYU v2 3Task it would be 1/4. Since the baselines transmit a single tensor, the comparison at R=1/12 is not on equal footing. If instead the z_t are combined into one tensor, then the definition of k in Eq. (4) and the dimensionality of the received signal in Eq. (6) must be revised, and the paper does not explain how the combination is performed. This is load-bearing for the paper's central bandwidth-constrained claims.
  2. [Abstract; Section IV.E, Tables XI and XII; Eq. (21)] The abstract's statement that GAI 'surpasses the most competitive and publicly available models by 11.4% on the CityScapes 2Task dataset' is not supported by the tables. The quantity defined in Eq. (21) is a model's average relative improvement over the single-task model, not the margin over the best baseline. In Table XI, GAI's entry of 11.40% is this Δ value, while the raw metric differences against AdaShare at R=1/12 are far smaller (e.g., mIoU 45.1 vs. 38.1). Similarly, the '3.97%' on NYU v2 3Task in the abstract is GAI's own relative improvement over single-task (Table XII), not the advantage over AdaShare. The headline numbers therefore overstate the improvement over the state of the art and should be corrected.
  3. [Tables II-VII, XI, XII; Section IV.D] The main comparison tables do not report standard deviations or significance tests for the baseline models. Tables XIII-XV provide STDs for the proposed GAI model over different seeds, but no comparable variation is shown for the baselines. Several claimed advantages over the best baseline are small (e.g., Table IV: GAI Δ=+9.47% vs. AdaShare +8.98%; Table XII: a 0.71 percentage-point advantage in the segmentation Δ at R=1/12). Without error bars or multiple-seed results for all models, the reader cannot assess whether these differences are beyond run-to-run variability. The paper should either add error bars to the baseline columns or provide a statistical test for the key comparisons.
  4. [Section IV.B.3, Section IV.E, Section IV.F] All models are trained under sufficient bandwidth and negligible channel noise (SNR approaching infinity), but are evaluated at low bandwidth ratios and at SNR values from -2 dB to 14 dB. The paper presents this as demonstrating robustness and states that the improvement becomes more pronounced as SNR decreases. While the protocol is applied uniformly, the claim that the proposed model is robust under noisy, bandwidth-limited channels would be stronger if the models were also trained at the deployment SNR or if the authors discussed how the clean-channel training transfers to noisy test conditions. This is a limitation of the experimental methodology that should be acknowledged and addressed, since the main headline claims concern bandwidth-constrained and noisy channels.
minor comments (5)
  1. [Section III.C, Eq. (13)] Equation (13) has a missing closing parenthesis: the expression should read Lineari2,t(ReLU(Lineari1,t(V_i^M))) rather than Lineari2,t(ReLU(Lineari1,t(V_i^M))).
  2. [Section IV.B.1 and References [37], [38]] The Oxford-IIIT Pet dataset is cited as [38] in the text, which corresponds to the MVSA reference, while the MVSA dataset is cited as [37], which is the Pet dataset. These citations appear to be swapped.
  3. [Throughout] The phrase 'multi-layer perceptual network' should be 'multilayer perceptron' for consistency with standard terminology.
  4. [Table IX and Section IV.A.1] The notation 'δ < 1.252' and 'δ < 1.253' should be typeset as δ < 1.25^2 and δ < 1.25^3, to match the definition given in Section IV.A.1.
  5. [Section III.E] The complexity analysis states that the overall complexity is O(N C_out^2) and 'little affected by the number of tasks T', but the inference-time numbers in Table XVI show a clear increase with T. The sentence should be qualified to clarify that the asymptotic analysis ignores the T-dependent terms that are visible in practice.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the GAI module is trained end-to-end and evaluated against external baselines; no fitted parameter is recycled as a prediction.

full rationale

The paper's central claim is an empirical comparison: the proposed GAI module improves multi-task semantic communication accuracy on standard benchmarks (CityScapes 2Task, NYU v2, etc.). The derivation chain is architectural rather than analytical: Eq. (14) constructs task-specific transmitted features z_t as a weighted sum of interpolated block features K_i with learned task-node weights e_i,t, and these weights are optimized jointly with the encoder and decoders through the loss in Eq. (15). No equation in the paper defines a target result in terms of itself, and no fitted parameter is later presented as a prediction. The task-node weights are learned end-to-end and then interpreted post hoc as correlations between blocks and tasks; this is interpretation of learned representations, not circular derivation. The paper contains no self-citations of the authors' prior work, no imported uniqueness theorem, and no ansatz smuggled in via citation; the GAT mechanism is standard and cited to the external GAT literature. The bandwidth-ratio fairness concern raised by the reader is an evaluation-protocol question about how the T task-specific z_t tensors are multiplexed into the single channel modeled in Eq. (2), not a circularity in which an output is equivalent to an input by construction. Thus the paper is self-contained against external benchmarks and receives a score of 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The GAI module introduces no invented physical entities. It relies on standard neural components (GAT, MLP, convolution, interpolation) and on the listed hyperparameters. The most consequential assumptions are the noise-free training protocol and the use of pooled block features as graph nodes.

free parameters (6)
  • Graph attention update iterations M = 1
    Set to 1 in Section IV.B.3; no sensitivity analysis is provided, and it controls how much inter-node information is exchanged.
  • Relation mapping hidden size = 256
    The first linear layer in each task-specific MLP uses 256 channels (Section IV.B.1); an architectural choice.
  • Task loss weights w_t = not reported
    Eq. (15) uses weights configured to equalize loss magnitudes; specific values are not given, leaving the multi-task objective partly unspecified.
  • Output channel count C_out = 512
    All node representations are projected to 512 channels, matching the last ResNet block (Section IV.B.1).
  • Bandwidth downscale channels C_ds = not reported
    Section IV.E adds convolution modules to change channel count for each bandwidth ratio, but the actual C_ds values are not given, making the compression pipeline unrepeatable.
  • Learning rate = 1e-4
    Selected after sensitivity analysis in Tables VIII-X; a standard hyperparameter, but part of the reported configuration.
assumptions (5)
  • domain assumption The channel is modeled as AWGN (Eq. 2), with Rayleigh fading used as an additional setting.
    Section II.C assumes additive white Gaussian noise; the paper asserts the extension to Rayleigh fading is straightforward.
  • ad hoc to paper Training with negligible noise transfers to noisy test channels.
    Section IV.B.3 trains at SNR to infinity and tests at -2 to 14 dB without a stated justification for transfer.
  • domain assumption Pooled block-level features are sufficient to represent inter-feature correlations.
    Section III.B pools each block output to a single vector before graph attention, discarding spatial structure.
  • ad hoc to paper A fully connected graph over encoder blocks is an appropriate relational model.
    Section III states the blocks form a fully connected graph; no alternative topology is compared.
  • domain assumption Relative improvement over a single-task baseline (Eq. 20) is a valid summary metric.
    The paper uses Delta as the primary comparison metric, which can overstate gains when baselines are small or when different metrics are averaged.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Task Semantic Communication With Graph Attention-Based Feature Correlation Extraction." pith.science (2026). https://pith.science/paper/KL3NYGTJ

@misc{pith2026250102006,
  author       = {Pith},
  title        = {Pith review of: Multi-Task Semantic Communication With Graph Attention-Based Feature Correlation Extraction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KL3NYGTJ}},
  note         = {Machine review of arXiv:2501.02006}
}
read the original abstract

Multi-task semantic communication can serve multiple learning tasks using a shared encoder model. Existing models have overlooked the intricate relationships between features extracted during an encoding process of tasks. This paper presents a new graph attention inter-block (GAI) module to the encoder/transmitter of a multi-task semantic communication system, which enriches the features for multiple tasks by embedding the intermediate outputs of encoding in the features, compared to the existing techniques. The key idea is that we interpret the outputs of the intermediate feature extraction blocks of the encoder as the nodes of a graph to capture the correlations of the intermediate features. Another important aspect is that we refine the node representation using a graph attention mechanism to extract the correlations and a multi-layer perceptron network to associate the node representations with different tasks. Consequently, the intermediate features are weighted and embedded into the features transmitted for executing multiple tasks at the receiver. Experiments demonstrate that the proposed model surpasses the most competitive and publicly available models by 11.4% on the CityScapes 2Task dataset and outperforms the established state-of-the-art by 3.97% on the NYU V2 3Task dataset, respectively, when the bandwidth ratio of the communication channel (i.e., compression level for transmission over the channel) is as constrained as 1 12 .

Figures

Figures reproduced from arXiv: 2501.02006 by the authors.

Figure 1
Figure 1. The proposed multi-task semantic communication system model, where the encoder extracts features, which are transmitted through the channel to the decoders for multiple simultaneous tasks. and by 1.2% and 5.19% on the Oxford-IIIT Pet and MVSA datasets, respectively. When the bandwidth ratio is reduced to as small as 1 12 , the GAI outperforms the baselines by 11.4% and 3.97% on the CityScapes 2Task and NYU V2 3Task … view at source ↗
Figure 2
Figure 2. The proposed GAI module, which involves generating encoded data zt for task t within the GAI module. This process begins with inputs from the output of each basic encoder block. The procedure comprises (i) a Feature Transformation Layer, (ii) a Graph Attention Layer, and (iii) a Relation Mapping Layer. The Feature Transformation Layer unifies each block’s feature size, and the Graph Attention Layer updates the node … view at source ↗
Figure 3
Figure 3. Graph Attention Layer: This diagram illustrates the iterative process of updating representation for the i-th node, specifically focusing on the m-th update iteration. Each node undergoes a total of M iterations for information updating. transmission in the channel process can be characterized by a conditional probability, as given by pch(zˆ|z). Without loss of generality, we use additive white Gaussian processes to… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Improvement of GAI over Ada under different SNR conditions in the CityScapes 2Task and NYU v2 3Task transmission. (a) CityScapes 2Task (b) NYU v2 3Task [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Improvement of GAI over Ada under different SNR conditions in the CityScapes 2Task and NYU v2 3Task transmission over fading channels. regime, particularly in the semantic segmentation task, as shown in Fig. 5a. On the NYU dataset, the GAI model shows notable performan…
Figure 6
Figure 6. Figure 6: The relationship among different tasks and each coding block on the three datasets [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: The input image. prediction, as shown in Fig. 5b. This consistent performance under low SNR conditions con￾firms the GAI model’s ability to leverage feature correlations while adapting task-specific weights effectively. Prioritizing critical features in noisy wireless …
Figure 8
Figure 8. Figure 8: (i) Shows the information encoded from the early blocks (1-4), which typically capture basic details such as edges and textures. (ii) Demonstrates the information encoded from the deeper blocks (13-16), where complex, high-level semantic information is encapsulated. (i…
Figure 9
Figure 9. Figure 9: Outage probability in the CityScapes 2Task transmission. For most metrics, curves closer to the bottom-right indicate better performance. Exceptions include “abs errs” and “rel errs”, where curves closer to the top-left are preferable [PITH_FULL_IMAGE:figures/full_fig…
Figure 10
Figure 10. Figure 10: Outage probability in the NYU v2 3-Task transmission. For most metrics, curves closer to the bottom-right indicate better performance. Exceptions include “abs errs”, “rel errs”, “angular means”, and “angular medians”, where curves closer to the top-left are preferable…
Figure 11
Figure 11. Figure 11: Training trails for different datasets and tasks. The x-axis represents the number of training steps, and the y-axis represents the training loss. counterpart enlarges with the SNR. While excelling in optimizing task-specific performance, the GAI model requires retrai…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 34 canonical work pages

  1. [1]

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

    T. Standley et al. , “Which tasks should be learned together in multi- task learning?” in Proc. Int. Conf. Mach. Learn. (ICML) , Jul. 2020, pp. 9120–9132

  2. [2]

    Driving scene perception network: Real-time joint detection, depth estimation and semantic segmentation,

    L. Chen, Z. Yang, J. Ma, and Z. Luo, “Driving scene perception network: Real-time joint detection, depth estimation and semantic segmentation,” in IEEE Winter Conf. Appl. Comput. Vis. (WACV) , Lake Tahoe, NV , USA, Mar. 2018, pp. 1283–1291

  3. [3]

    Task-Oriented Communications for 6G: Vision, Principles, and Technologies

    Y . Shi et al., “Task-oriented communications for 6g: Vision, principles, and technologies,” arXiv preprint arXiv:2303.10920 , 2023

  4. [4]

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

    D. G ¨und¨uz et al. , “Beyond transmitting bits: Context, semantics, and task-oriented communications,” IEEE J. Sel. Areas Commun. , vol. 41, no. 1, pp. 5–41, 2022

  5. [5]

    Toward wisdom-evolutionary and primitive-concise 6G: A new paradigm of semantic communication networks,

    P. Zhang et al., “Toward wisdom-evolutionary and primitive-concise 6G: A new paradigm of semantic communication networks,” Engineering, vol. 8, pp. 60–73, 2022

  6. [6]

    Communication beyond transmitting bits: Semantics-guided source and channel coding,

    J. Dai, P. Zhang, K. Niu, S. Wang, Z. Si, and X. Qin, “Communication beyond transmitting bits: Semantics-guided source and channel coding,” IEEE Wirel. Commun., vol. 30, no. 4, pp. 170–177, 2023

  7. [7]

    Edge computing for metaverse: Incentive mech- anism versus semantic communication,

    N. C. Luong et al. , “Edge computing for metaverse: Incentive mech- anism versus semantic communication,” IEEE Trans. Mob. Comput. , vol. 23, no. 5, pp. 6196–6211, 2024

  8. [8]

    Deep joint source- channel coding for wireless image transmission,

    E. Bourtsoulatze, D. B. Kurka, and D. G ¨und¨uz, “Deep joint source- channel coding for wireless image transmission,” IEEE Trans. Cogn. Commun. Netw., vol. 5, no. 3, pp. 567–579, 2019

Show all 40 references
  1. [9]

    Deep learning enabled semantic communication systems,

    H. Xie, Z. Qin, G. Y . Li, and B.-H. Juang, “Deep learning enabled semantic communication systems,” IEEE Trans. Signal Process., vol. 69, pp. 2663–2675, 2021

  2. [10]

    A new communication paradigm: from bit accuracy to semantic fidelity,

    G. Shi et al. , “A new communication paradigm: from bit accuracy to semantic fidelity,” arXiv preprint arXiv:2101.12649 , 2021

  3. [11]

    Task-oriented multi-user semantic communications for VQA,

    H. Xie, Z. Qin, and G. Y . Li, “Task-oriented multi-user semantic communications for VQA,” IEEE Wirel. Commun. Lett. , vol. 11, no. 3, pp. 553–557, 2021

  4. [12]

    Learning task-oriented communication for edge inference: An information bottleneck approach,

    J. Shao, Y . Mao, and J. Zhang, “Learning task-oriented communication for edge inference: An information bottleneck approach,” IEEE J. Sel. Areas Commun., vol. 40, no. 1, pp. 197–211, 2021

  5. [13]

    Semantic transfer between different tasks in the semantic communication system,

    Q. Wu et al., “Semantic transfer between different tasks in the semantic communication system,” in Proc. IEEE Wireless Commun. and Netw. Conf. (WCNC), Austin, TX, United States, Apr. 2022, pp. 566–571

  6. [14]

    A unified multi-task semantic communication system for multimodal data,

    G. Zhang et al., “A unified multi-task semantic communication system for multimodal data,” IEEE Trans. Commun. , pp. 1–1, 2024

  7. [15]

    A unified multi-task semantic communication system with domain adaptation,

    G. Zhang, Q. Hu et al., “A unified multi-task semantic communication system with domain adaptation,” in Proc. IEEE Global Commun. Conf. (GLOBECOM), Virtual, Online, Brazil, Dec. 2022, pp. 3971–3976

  8. [16]

    Image semantic communications: An extended rate-distortion theory based scheme,

    W. Tong, F. Liu, Z. Sun, Y . Yang, and C. Guo, “Image semantic communications: An extended rate-distortion theory based scheme,” in Proc. IEEE Globecom Workshops. (GC Wkshps), Virtual, Online, Brazil, Dec. 2022, pp. 1723–1728

  9. [17]

    Semantic communication approach for multi-task image transmission,

    Z. Zhang et al. , “Semantic communication approach for multi-task image transmission,” in Proc. IEEE 96th Veh. Technol. Conf. (VTC-Fall), London, United Kingdom, Sept. 2022, pp. 1–2

  10. [18]

    A multi-task semantic communication system for natural language processing,

    Y . Sheng et al. , “A multi-task semantic communication system for natural language processing,” in Proc. IEEE 96th Veh. Technol. Conf. (VTC-Fall), London, United Kingdom, Sept. 2022, pp. 1–5

  11. [19]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. D. M.-W. C. Kenton and L. K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” in Proc. Conf. North Amer.Chapter Assoc. Comput. Linguistics-Hum.Lang. Technol. (NAACL HLT), Minneapolis, Minnesota, Jun. 2019, pp. 4171–4186

  12. [20]

    A mathematical theory of communication,

    C. E. Shannon, “A mathematical theory of communication,” The Bell system technical journal , vol. 27, no. 3, pp. 379–423, 1948

  13. [21]

    Scalable multi-task semantic communication system with feature importance ranking,

    J. Hu, F. Wang, W. Xu, H. Gao, and P. Zhang, “Scalable multi-task semantic communication system with feature importance ranking,” in Proc. IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP), Rhodes Island, Greece, Jun. 2023, pp. 1–5

  14. [22]

    Cross-stitch networks for multi-task learning,

    I. Misra et al., “Cross-stitch networks for multi-task learning,” in Proc. IEEE Comput. Soc. Conf. Comput. Vision Pattern Recognit. (CVPR) , Las Vegas, NV , United States, Jun. 2016, pp. 3994–4003

  15. [23]

    Latent multi-task architecture learning,

    S. Ruder, J. Bingel, I. Augenstein, and A. Søgaard, “Latent multi-task architecture learning,” in Proc. AAAI Conf. on Artificial Intelligence (AAAI), vol. 33, no. 01, Honolulu, HI, United States, Jan. 2019, pp. 4822–4829

  16. [24]

    End-to-end multi-task learning with attention,

    S. Liu et al. , “End-to-end multi-task learning with attention,” in Proc. IEEE Comput. Soc. Conf. Comput. Vision Pattern Recognit. (CVPR) , Long Beach, CA, United States, Jun. 2019, pp. 1871–1880

  17. [25]

    Nddr-cnn: Layerwise feature fusing in multi-task cnns by neural discriminative dimensionality reduction,

    Y . Gao et al., “Nddr-cnn: Layerwise feature fusing in multi-task cnns by neural discriminative dimensionality reduction,” in Proc. IEEE Comput. 18 Soc. Conf. Comput. Vision Pattern Recognit. (CVPR) , Long Beach, CA, United States, Jun. 2019, pp. 3205–3214

  18. [26]

    Deep elastic networks with model selection for multi-task learning,

    C. Ahn, E. Kim, and S. Oh, “Deep elastic networks with model selection for multi-task learning,” inProc. IEEE Int. Conf. Comput. Vision (ICCV), Gangnam-gu, Seoul, Korea., Oct. 2019, pp. 6529–6538

  19. [27]

    Adashare: Learning what to share for efficient deep multi-task learning,

    X. Sun, R. Panda, R. Feris, and K. Saenko, “Adashare: Learning what to share for efficient deep multi-task learning,” Adv. Neural Inf. Process. Syst., vol. 33, pp. 8728–8740, 2020

  20. [28]

    Dynashare: Task and instance conditioned parameter sharing for multi-task learning,

    E. Rahimian et al., “Dynashare: Task and instance conditioned parameter sharing for multi-task learning,” in Proc. IEEE Comput. Soc. Conf. Com- put. Vis. Pattern Recognit. Workshops (CVPR Workshops) , Vancouver, BC, Canada, Jun. 2023, pp. 4534–4542

  21. [29]

    DeepJSCC-f: Deep joint source-channel coding of images with feedback,

    D. B. Kurka and D. G ¨und¨uz, “DeepJSCC-f: Deep joint source-channel coding of images with feedback,” IEEE J. Sel. Areas Inf. Theory , vol. 1, no. 1, pp. 178–193, 2020

  22. [30]

    Wireless image transmission using deep source channel coding with attention modules,

    J. Xu, B. Ai, W. Chen, A. Yang, P. Sun, and M. Rodrigues, “Wireless image transmission using deep source channel coding with attention modules,” IEEE Trans. Circuits Syst. Video Technol., vol. 32, no. 4, pp. 2315–2328, 2021

  23. [31]

    A comprehensive survey on graph neural networks,

    Z. Wu et al., “A comprehensive survey on graph neural networks,” IEEE Trans. Neural Netw. Learn. Syst. , vol. 32, no. 1, pp. 4–24, 2021

  24. [32]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS) , vol. 30, 2017, pp. 5998–6008

  25. [33]

    Imagenet classification with deep convolutional neural networks,

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS) , vol. 25, 2012, pp. 1097–1105

  26. [34]

    The cityscapes dataset for semantic urban scene understanding,

    M. Cordts et al. , “The cityscapes dataset for semantic urban scene understanding,” in Proc. IEEE Comput. Soc. Conf. Comput. Vision Pattern Recognit. (CVPR) , Las Vegas, NV , United States, Jun. 2016, pp. 3213–3223

  27. [35]

    Indoor segmentation and support inference from rgbd images,

    N. Silberman, D. Hoiem, P. Kohli, and R. Fergus, “Indoor segmentation and support inference from rgbd images,” in Lect. Notes Comput. Sci. , Florence, Italy, Oct. 2012, pp. 746–760

  28. [36]

    Taskonomy: Disentangling task transfer learning,

    A. R. Zamir, A. Sax, W. Shen, L. J. Guibas, J. Malik, and S. Savarese, “Taskonomy: Disentangling task transfer learning,” in Proc. Int. Joint Conf. Artif. Intell. (IJCAI) , Macao, China, Aug. 2018, pp. 3712–3722

  29. [37]

    Cats and dogs,

    O. M. Parkhi et al., “Cats and dogs,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2012, pp. 3498–3505

  30. [38]

    Sentiment analysis on multi-view social data,

    L. Niu, W. Zhang, and C. Song, “Sentiment analysis on multi-view social data,” in Proc. Int. Joint Conf. Artif. Intell. (IJCAI). AAAI Press, 2016, pp. 3939–3945

  31. [39]

    On first-order meta-learning algorithms,

    A. Nichol, J. Achiam, and J. Schulman, “On first-order meta-learning algorithms,” arXiv preprint arXiv:1803.02999 , 2018

  32. [40]

    Model-agnostic meta-learning for fast adaptation of deep networks,

    C. Finn, P. Abbeel, and S. Levine, “Model-agnostic meta-learning for fast adaptation of deep networks,” in Proc. Int. Conf. Mach. Learn. (ICML). PMLR, 2017, pp. 1126–1135. Xi Yu (Graduate Student Member, IEEE) re- ceived the B.E. degree in communication engineer- ing from Beij...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.