Pith. sign in

REVIEW 5 major objections 6 minor 36 references

Multi-Modal Beamforming with Model Compression and Modality Generation for V2X Networks

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A hierarchical Transformer fusing camera, LiDAR, radar, and GPS predicts mmWave beam indices for V2X links, reaching an overall DBA-score of 0.9129 on real-world urban scenarios, and remains robust under pruning and missing sensors.

desk verdict BeamTransFuser is a coherent engineering system with a flawed baseline comparison: the reported overall DBA score is internally inconsistent with its own per-scenario numbers. read the letter →

arxiv 2506.22469 v2 pith:T3AIMGNZ submitted 2025-06-19 eess.SP

Chen Shang , Dinh Thai Hoang , Jiadong Yu This is my paper
classification eess.SP
keywords integratedsensingandcommunicationmulti-modalbeamformingV2XnetworkstransformerfusionmodelcompressionpruningconditionalvariationalautoencoderDeepSense6G
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

BeamTransFuser is a multi-modal deep learning framework that predicts which beam a roadside unit should use to reach a vehicle, based on camera, LiDAR, radar, and GPS measurements. The paper claims it consistently beats existing multi-modal beamforming methods on real-world V2X scenarios, reaching an overall distance-based accuracy score of 0.9129. To make the model deployable, the authors add a module-aware pruning scheme that keeps accuracy high while cutting inference latency, and a conditional variational autoencoder that reconstructs missing radar or LiDAR inputs from the surviving sensors. If these claims hold, beam alignment in 6G V2X networks can tolerate sensor failures and run in real time on roadside hardware without exhaustive beam sweeping.

What carries the argument

The load-bearing object is BeamTransFuser, an end-to-end fusion network built from modality-specific ResNet-based encoders that feed four interleaved multi-modal fusion blocks. Each fusion block takes tokens from camera, LiDAR, radar, and GPS, concatenates them into one sequence, and applies shared multi-head self-attention plus residual connections, so attention scores carry the cross-modal dependencies; a learnable softmax-weighted aggregation then combines the final features before a three-layer MLP emits one of 64 beam indices. Two auxiliary mechanisms support deployment: a splitting-and-pruning scheme that uses KL-divergence importance scores to prune embedding dimensions, attention projections, and feed-forward neurons module by module, and a CVAE whose decoder generates missing modality features from the available ones without retraining.

What would settle it

Run every baseline checkpoint from [8], [9], [10], [11], [12], [13], [14], [18] on the same 10,118 training and 1,125 validation samples with the same beam codebook and evaluation script; if any method's reconstructed score (e.g., ICMFE's Scenario 34 DBA of 0.9074 or its assumed 100% Scenario 31 value) rises above BeamTransFuser's 0.9129 overall, the claim of consistent superiority fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that a hierarchical Transformer architecture, with separate branches for each sensing modality and four interleaved fusion blocks, can learn cross-modal correlations that translate directly into better beam index prediction. On the DeepSense 6G urban scenarios 31–34, the model reports an overall DBA-score of 0.9129, the highest among the compared methods, with per-scenario scores from 0.8945 to 1.0000. The same framework, after splitting-based pruning, retains an 88.71% DBA-score at 80% parameter retention and drops below the latency threshold used for 5G-V2X at 60% retention. When radar or LiDAR is missing, a CVAE trained on the observed modalities restores accuracy to 89.84% ($\pm 0.5\%$), close to the full-modality setting; the authors do not apply generation to the camera modality, whose absence degrades accuracy to 19.88%.

Load-bearing premise

The main empirical claim depends on baseline numbers taken from prior papers, several with missing entries that were filled by averaging Top-1/Top-3 or assuming a perfect score, so if those published numbers came from different training splits or evaluation protocols the performance margin could shrink or reverse.

Editorial extensions

If this is right

  • An RSU can select one of 64 beams in a single forward pass from camera, LiDAR, radar, and GPS data, avoiding exhaustive beam sweeping and its signaling overhead.
  • Pruning to 60% parameter retention keeps all tested GPU platforms below 40 ms inference latency, meeting the real-time threshold cited for 5G-V2X while retaining a DBA-score of 82.02%.
  • When radar or LiDAR disappears at inference time, the CVAE restores accuracy to about 89.84%, so the system keeps working without retraining under partial sensor failure.
  • A single trained model stays stable across daytime and nighttime scenarios (per-scenario DBA-scores from 0.8945 to 1.0000), suggesting it generalizes across LoS/NLoS and lighting changes.
  • Larger beam codebooks can be adopted with negligible extra latency because only the final linear projection layer changes, unlike exhaustive beam search which scales with codebook size.

Reading between the lines

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

  • A testable extension beyond the paper is camera reconstruction: because missing camera drops accuracy to 19.88% and the authors explicitly leave camera generation out, a higher-capacity conditional generator on LiDAR+radar+GPS would complete the robustness story.
  • An ablation not run in the paper — one or two fusion blocks instead of four — would test whether the hierarchical fusion design is necessary or simply sufficient, since the pruning curve suggests the fusion blocks carry redundancy.
  • Because radar loss (30.94%) degrades accuracy more than LiDAR loss (50.32%), sensor-failure planning should prioritize radar redundancy or a radar fallback path; this prioritization is an inference from the reported masking results.
  • The cleanest validation of the headline comparison would be a re-benchmark that fills every dash in Tables III and IV from the original models on a shared split; the paper's own reconstructed averages would then be checked directly.
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

5 major / 6 minor

Summary. The paper proposes BeamTransFuser, a hierarchical Transformer-based multi-modal fusion framework that combines camera, LiDAR, radar, and GPS data to predict mmWave beam indices on the DeepSense 6G urban scenarios 31-34. It also introduces a splitting-and-pruning compression scheme to reduce inference latency on roadside units and a conditional variational autoencoder to reconstruct missing radar or LiDAR features from the available modalities. The reported results claim a state-of-the-art overall DBA-score of 0.9129, competitive accuracy after pruning (e.g., 88.71% DBA at 0.8 parameter retention), and near-full-modality accuracy (~89.84%) when missing radar or LiDAR inputs are replaced by generated features.

Significance. If the empirical claims hold, the paper would make a useful contribution to multi-modal sensing-assisted beamforming: it uses a real-world public dataset, addresses practical deployment constraints (latency and missing sensors), and combines fusion, compression, and generation in one system. The work is not a theoretical derivation, and the machine-checked or reproducible-code strengths are absent (no code release or seed reporting); its value rests on experimental validation. The main risks are the internal inconsistency in the headline comparison and the reliance on reconstructed baseline numbers, which currently undercut the central claim of consistent outperformance. The compression and generative components are plausible but under-supported by ablations and statistical reporting.

major comments (5)
  1. [Table III] The headline overall DBA comparison is internally inconsistent. For baseline [11], the per-scenario DBA differences relative to BeamTransFuser are 0, +0.0018, +0.0114, and -0.0129 for scenarios 31-34, while the reported overall difference is 0.9129 - 0.8969 = 0.0160. Since any sample-weighted or pooled overall DBA is a convex combination of per-scenario values, the maximum possible weighted difference is 0.0114 (achieved by giving scenario 33 weight 1 and scenario 34 weight 0). No nonnegative weights summing to 1 can reproduce both rows. The authors must either correct the aggregation, or state explicitly that the per-scenario and overall numbers come from different validation splits or protocols. This inconsistency is load-bearing because the paper's main empirical claim is the highest overall DBA-score.
  2. [Section V-C-2, Table IV] The baseline comparison relies on reconstructed numbers from prior papers rather than re-running the baselines under the same protocol. The authors fill missing Top-2 entries by averaging Top-1 and Top-3 (e.g., for Position-based [8] and Multi-modal [8]) and assume ICMFE [11] achieves 100% accuracy in Scenario 31 based on a claim in its paper. These approximations are not guaranteed to be unbiased, and differences in train/validation splits, preprocessing, and evaluation protocols across papers can change the performance margins. To support the 'consistently outperforms all baselines' claim, the authors should rerun the public baselines on the same split or clearly characterize the comparability limits and quantify the sensitivity of the margin to the reconstruction assumptions.
  3. [Section IV-A, Fig. 6] The claim that the splitting-based pruning scheme is superior to conventional one-size-fits-all pruning is not tested. Figure 6 shows only the accuracy-latency tradeoff of BeamTransFuser under the proposed pruning, with no comparison against standard global magnitude pruning, per-layer uniform pruning, or other compression baselines at matched parameter budgets. Since the motivation in Section IV-A explicitly argues that global pruning is inappropriate for this architecture, the evaluation should include at least one standard pruning baseline to substantiate that claim.
  4. [Section V-C-4, Table V] The generative evaluation is too thin to support the broad missing-modality claim. Table V reports a single generated-accuracy number per condition (with a +/-0.5% range but no stated number of seeds), covers only radar and LiDAR, and explicitly omits the camera case, which is dropped to 19.88% missing accuracy. In addition, the text says the generated accuracy 'matches the full-modality setting,' but the reported 89.84% differs from the full-model overall DBA of 91.29% in Table III; the metric used in Table V (Top-1 accuracy vs. DBA-score) is also not defined. The authors should specify the metric, report multiple seeds, and scope the generative claim to the modalities actually supported.
  5. [Section V-A and V-C] The paper does not report how the 'overall' DBA-score and the 'Overall Avg.' Top-k accuracy are computed. For Table III, the overall value is not the simple average of the per-scenario DBA values (e.g., for BeamTransFuser the simple average would be 0.9243, not 0.9129). For Table IV, the reported overall average of 77.27% does not match the simple average of the per-scenario Top-1/Top-2/Top-3 entries (approximately 80.01%). The authors should state the weighting formula or correct the reported values.
minor comments (6)
  1. [Section V-C-2] The statement that 'Top-2 typically lies between Top-1 and Top-3' does not justify using the mean of Top-1 and Top-3 as an estimate of Top-2; monotonicity alone does not imply that the average is unbiased. Please report exact values or rerun the baselines.
  2. [Fig. 6] The axis label 'remaining parameter ration' contains a typo ('ration' should be 'ratio'), and the y-axis scaling for DBA accuracy should be clarified (it appears to start near 100% rather than at 0).
  3. [Section IV-B] There is a typo 'FNN' where 'FFN' (feed-forward network) is intended, and the KL-divergence name is split across a line break in the text; please fix these presentation issues.
  4. [Eq. (3)] The notation is inconsistent: Eq. (2) defines gamma(f_m), but Eq. (3) uses gamma_n(f_m) without defining the subscript n; please align the notation.
  5. [Section V-C-3] The latency claims would be easier to verify if the values in Fig. 6(b) were reported in a table with exact numbers per device and parameter-retention ratio, rather than only in the plot.
  6. [Section V-C-1] The text says BeamTransFuser is 'slightly outperformed by QTNs (0.9124) and the method in [11] (0.9074) in Scenario 34' but then claims it 'consistently outperforms' all methods; please rephrase to 'achieves the highest overall DBA-score' to avoid an apparent contradiction.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the framework, pruning, and generative model are evaluated on held-out data with standard losses; baseline reconstruction issues are correctness risks, not circularity.

full rationale

This is an empirical deep-learning paper rather than a derivation chain. The claimed prediction model M_Θ maps multi-modal inputs to beam indices and is trained with cross-entropy/focal loss on a 90/10 train/validation split; the DBA-Score and Top-k metrics are external evaluation criteria computed from the holdout validation set, not quantities defined in terms of the model's own outputs. The CVAE in Section IV-B uses the standard conditional ELBO (Eq. 16) with the usual relaxation p(z|x)=N(0,I), and the generated features are tested on the same frozen BeamTransFuser model, so the 89.84% generated accuracy is not forced by construction. The pruning scheme is evaluated after fine-tuning on the same training protocol; accuracy and latency are measured, not derived from the pruning ratios themselves. There are no load-bearing self-citations: the authors cite prior work for baselines and standard architectures, but the central claims are supported by their own experiments. The main caveat is that some baseline numbers in Tables III and IV are reconstructed (e.g., averaging Top-1 and Top-3 to estimate Top-2, and assuming ICMFE's missing Scenario 31 score is 100%), and the Table III overall DBA difference is not arithmetically consistent with the per-scenario differences under any convex weighting. Those are evaluation-comparability and internal-consistency concerns affecting the strength of the 'consistently outperforms' claim, but they are not examples of a fitted parameter being renamed as a prediction or an equation reducing to its own input. The limitation that camera modality is not generated is explicitly acknowledged in Section V-C.4. No circular step meeting the required evidence standard was found.

Assumptions & free parameters 5 free parameters · 7 assumptions · 0 invented entities

The central empirical claims rely on standard deep learning background (attention, ELBO), on the DeepSense 6G dataset labels, and on several paper-specific modeling choices: relaxing the CVAE prior to N(0,I), simulating missing sensors with Gaussian noise, and assuming GPS availability. These are not hidden, but they are assumptions a reader must accept to trust the reported performance.

free parameters (5)
  • Number of attention heads in fusion blocks = 4
    Chosen to match the four sensor modalities; a hand-set design choice that affects cross-modal capacity and pruning constraints (Section V-B).
  • Embedding dimensions of fusion blocks D = [64, 128, 256, 512]
    Progressive capacity schedule selected by the authors; trade-off between representational power and computational cost (Section V-B).
  • CVAE latent dimension = 128
    Capacity of the generative model; hand-set in Section V-B.
  • Global pruning ratio r = 0.1, 0.2, 0.3, 0.4, 0.5, 0.6
    Compression target chosen for evaluation; results in Fig. 6 depend on it.
  • Training/fine-tuning epoch counts = 30 epochs pretrain, 10 epochs fine-tune
    Hyperparameters reported in Section V-B; not derived from theory.
assumptions (7)
  • standard math Standard Transformer attention formulation (Eqs. 6-9) is correct and provides the cross-modal fusion mechanism.
    Equations (6)-(9) follow Vaswani et al. [16]; no proof is required.
  • standard math The ELBO lower bound for conditional variational inference (Eqs. 14-15) is valid.
    The derivation using Jensen's inequality is standard variational inference.
  • domain assumption The 64-beam codebook and the power-vector labels from DeepSense 6G accurately identify the optimal beam.
    The dataset provides the 64x1 power vector per time step, used as ground truth for beam index; Section V-A.
  • domain assumption The optimal beam is a deterministic function of the multi-modal sensing inputs, so a mapping M_Theta can learn it.
    The problem formulation in Eq. (5) assumes such a mapping exists; Section II.
  • ad hoc to paper The conditional prior p(z|x) can be relaxed to N(0,I), making z independent of x.
    Section IV-B states this relaxation 'to simplify training and ensure tractability'; it is a modeling choice specific to this paper.
  • domain assumption Simulating a missing modality by replacing it with standard Gaussian noise is representative of real sensor failure.
    Section V-C-4 uses this simulation to measure missing-modality accuracy; if real failures differ, the results may not transfer.
  • domain assumption GPS is always available and camera reconstruction is not needed.
    Section V-C-4 explains GPS is not generated because it comes from the vehicle, and camera is excluded due to complexity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Modal Beamforming with Model Compression and Modality Generation for V2X Networks." pith.science (2026). https://pith.science/paper/T3AIMGNZ

@misc{pith2026250622469,
  author       = {Pith},
  title        = {Pith review of: Multi-Modal Beamforming with Model Compression and Modality Generation for V2X Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T3AIMGNZ}},
  note         = {Machine review of arXiv:2506.22469}
}
read the original abstract

Integrating sensing and communication (ISAC) is a promising technology for predictive beamforming in 6G vehicle-to-everything (V2X) networks. However, current ISAC paradigms rely solely on radio-frequency (RF)-based sensing, which limits sensing resolution and beamforming robustness in complex wireless environments. Fortunately, the widespread deployment of diverse non-RF sensors such as cameras and LiDAR, along with the integration of artificial intelligence (AI) and communication systems, offers new opportunities to improve the synergy between sensing and communication. Motivated by this, this work develops a multi-modal sensing-assisted beamforming framework for realistic V2X scenarios. Specifically, we propose BeamTransFuser, a hierarchical Transformer-based multi-modal learning framework that exploits cross-modal correlations among camera, LiDAR, radar, and GPS observations to improve beam prediction accuracy and robustness. To facilitate practical deployment on roadside units, we further develop a module-aware pruning scheme to reduce inference latency while preserving competitive performance. Furthermore, to address potential missing-modality conditions in real-world scenarios, we introduce a generative model that is able to reconstruct missing inputs from available observations, allowing the framework to operate reliably even under incomplete sensing conditions. Extensive experimental results conducted on real-world datasets demonstrate that the proposed scheme consistently outperforms existing baselines across various metrics.

Figures

Figures reproduced from arXiv: 2506.22469 by the authors.

Figure 1
Figure 1. Multi-modal sensing-assisted communication system. The [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The overview of BeamTransFuser. BeamTransFuser consists of four modality-specific branches (i.e., camera, LiDAR, radar, and [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. The splitting scheme of multi-modal fusion backbone. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: The overview of the multi-modal fusion block and the model compression process. The compression pipeline consists of: (1) KL [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: The overview of generative model. The trained decoder [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: (a) DBA accuracy vs. remaining parameter ratio and (b) [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 24 canonical work pages

  1. [11]

    Advancing multi-modal beam prediction with cross-modal feature enhancement and dynamic fusion mechanism,

    Q. Zhu, Y . Wang, W. Li, H. Huang, and G. Gui, “Advancing multi-modal beam prediction with cross-modal feature enhancement and dynamic fusion mechanism,” IEEE Trans. Commun. , pp. 1–1, 2025

    work page 2025

  2. [8]

    Multimodal deep learning empowered millimeter-wave beam prediction,

    B. Shi, M. Li, M.-M. Zhao, M. Lei, and L. Li, “Multimodal deep learning empowered millimeter-wave beam prediction,” in 2024 IEEE 99th Vehicular Technology Conference, 2024, pp. 1–6

    work page 2024

  3. [1]

    Draft New Recommendation ITU-R M. [IMT. Frame- work for 2030 and Beyond],

    ITU-R WP5D, “Draft New Recommendation ITU-R M. [IMT. Frame- work for 2030 and Beyond],” 2023

    2023

  4. [2]

    Integrated sensing and communications: Toward dual-functional wire- less networks for 6G and beyond,

    F. Liu, Y . Cui, C. Masouros, J. Xu, T. X. Han, Y . C. Eldar, and S. Buzzi, “Integrated sensing and communications: Toward dual-functional wire- less networks for 6G and beyond,” IEEE J. Sel. Areas Commun., vol. 40, no. 6, pp. 1728–1767, 2022

    2022

  5. [3]

    V2X functional and performance test report: Test procedures and results,

    “V2X functional and performance test report: Test procedures and results,” 5G Automotive Association, Tech. Rep. 5GAA P-190033, 2019. [Online]. Available: https://5gaa.org/content/uploads/2018/11/5GAA P-190033 V2X-Functional-and-Performance-Test-Report final-1.pdf

    work page 2019

  6. [4]

    Toward ISAC-empowered vehicular networks: Framework, Advances, and Opportunities,

    Z. Du, F. Liu, Y . Li, W. Yuan, Y . Cui, Z. Zhang, C. Masouros, and B. Ai, “Toward ISAC-empowered vehicular networks: Framework, Advances, and Opportunities,” IEEE Wireless Commun. , vol. 32, no. 2, pp. 222– 229, 2025

    work page 2025

  7. [5]

    On the fundamental tradeoff of integrated sensing and communications under gaussian channels,

    Y . Xiong, F. Liu, Y . Cui, W. Yuan, T. X. Han, and G. Caire, “On the fundamental tradeoff of integrated sensing and communications under gaussian channels,” IEEE Trans. Inf. Theory , vol. 69, no. 9, pp. 5723– 5751, 2023

    2023

  8. [6]

    Real time object detection using lidar and camera fusion for autonomous driving,

    H. Liu, C. Wu, and H. Wang, “Real time object detection using lidar and camera fusion for autonomous driving,” Scientific Reports, vol. 13, no. 1, p. 8056, 2023

    work page 2023

Show all 36 references
  1. [7]

    Intelligent multi-modal sensing-communication integration: Synesthesia of machines,

    X. Cheng, H. Zhang, J. Zhang, S. Gao, S. Li, Z. Huang, L. Bai, Z. Yang, X. Zheng, and L. Yang, “Intelligent multi-modal sensing-communication integration: Synesthesia of machines,” IEEE Commun. Surv. Tutorials , vol. 26, no. 1, pp. 258–301, 2024

    2024

  2. [9]

    Multimodal transformers for wireless communications: A case study in beam pre- diction,

    Y . Tian, Q. Zhao, F. Boukhalfa, K. Wu, F. Bader et al. , “Multimodal transformers for wireless communications: A case study in beam pre- diction,” arXiv preprint arXiv:2309.11811 , 2023

    2023 arXiv

  3. [10]

    Advancing multi-modal beam prediction with multipath-like data augmentation and efficient fusion mechanism

    Q. Zhu, Y . Wang, W. Li, H. Huang, J. Yin, L. Guo, Y . Lin, and G. Gui, “Advancing multi-modal beam prediction with multipath-like data augmentation and efficient fusion mechanism.” New York, NY , USA: Association for Computing Machinery, 2024. [Online]. Available: https://doi...

    2024

  4. [12]

    Deep quantum-transformer networks for multimodal beam prediction in isac systems,

    S. Tariq, B. E. Arfeto, U. Khalid, S. Kim, T. Q. Duong, and H. Shin, “Deep quantum-transformer networks for multimodal beam prediction in isac systems,” IEEE Internet Things J. , vol. 11, no. 18, pp. 29 387– 29 401, 2024

    2024

  5. [13]

    Multi-modality sensing in mmwave beamforming for connected vehicles using deep learning,

    M. B. Mollah, H. Wang, M. A. Karim, and H. Fang, “Multi-modality sensing in mmwave beamforming for connected vehicles using deep learning,” IEEE Trans. Cognit. Commun. Networking. , pp. 1–1, 2025

    2025

  6. [14]

    Multi-modal transformer and reinforcement learning-based beam man- agement,

    M. Ghassemi, H. Zhang, A. Afana, A. B. Sediq, and M. Erol-Kantarci, “Multi-modal transformer and reinforcement learning-based beam man- agement,” IEEE Networking Letters , vol. 6, no. 4, pp. 222–226, 2024

    2024

  7. [15]

    Trans- fuser: Imitation with transformer-based sensor fusion for autonomous driving,

    K. Chitta, A. Prakash, B. Jaeger, Z. Yu, K. Renz, and A. Geiger, “Trans- fuser: Imitation with transformer-based sensor fusion for autonomous driving,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 45, no. 11, pp. 12 878–12 895, 2022

    2022

  8. [16]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems , vol. 30, 2017

    2017

  9. [17]

    Aligning beam with imbalanced multi-modality: A generative federated learning approach,

    J. Liang, M. Wen, S. Wang, Y . Liang, and S. Gao, “Aligning beam with imbalanced multi-modality: A generative federated learning approach,” arXiv preprint arXiv:2504.14835 , 2025

    2025 arXiv

  10. [18]

    Deepsense itu multi modal beam prediction challenge 2022 – deepsense,

    “Deepsense itu multi modal beam prediction challenge 2022 – deepsense,” Deepsense6g.net, 2022. [Online]. Available: https://www. deepsense6g.net/ml-task-multi-modal-beam-prediction/

    2022

  11. [19]

    Cross-entropy loss functions: Theoretical analysis and applications,

    A. Mao, M. Mohri, and Y . Zhong, “Cross-entropy loss functions: Theoretical analysis and applications,” in International Conference on Machine Learning. PMLR, 2023, pp. 23 803–23 828

    2023

  12. [20]

    Focal loss for dense object detection,

    T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Doll ´ar, “Focal loss for dense object detection,” in Proceedings of the IEEE international conference on computer vision , 2017, pp. 2980–2988

    2017

  13. [21]

    Multi-modal beam prediction challenge 2022: Towards generalization,

    G. Charan, U. Demirhan, J. Morais, A. Behboodi, H. Pezeshki, and A. Alkhateeb, “Multi-modal beam prediction challenge 2022: Towards generalization,” arXiv preprint arXiv:2209.07519 , 2022

    2022 arXiv

  14. [22]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 770–778

    2016

  15. [23]

    GPT-4 Technical Report,

    OpenAI, “GPT-4 Technical Report,” https://openai.com/research/gpt-4, 2023

    2023

  16. [24]

    A survey on vision transformer,

    K. Han, Y . Wang, H. Chen, X. Chen, J. Guo, Z. Liu, Y . Tang, A. Xiao, C. Xu, Y . Xu, Z. Yang, Y . Zhang, and D. Tao, “A survey on vision transformer,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 45, no. 1, pp. 87–110, 2023

    2023

  17. [25]

    A survey on deep neural network pruning: Taxonomy, comparison, analysis, and recommendations,

    H. Cheng, M. Zhang, and J. Q. Shi, “A survey on deep neural network pruning: Taxonomy, comparison, analysis, and recommendations,” IEEE Trans. Pattern Anal. Mach. Intell. , 2024

    2024

  18. [26]

    Distilling the knowledge in a neural network,

    G. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a neural network,” arXiv preprint arXiv:1503.02531 , 2015

    2015 arXiv

  19. [27]

    J. M. Joyce, Kullback-Leibler Divergence . Berlin, Heidelberg: Springer Berlin Heidelberg, 2011, pp. 720–722. [Online]. Available: https://doi.org/10.1007/978-3-642-04898-2 327

    2011 doi

  20. [28]

    Adaptive head pruning for attention mechanism in the maritime domain,

    W. Messaoud, R. Trabelsi, A. Cabani, and F. Abdelkefi, “Adaptive head pruning for attention mechanism in the maritime domain,” IEEE Transactions on Artificial Intelligence , pp. 1–12, 2025

    2025

  21. [29]

    Deep generative modelling: A comparative review of vaes, gans, normalizing flows, energy-based and autoregressive models,

    S. Bond-Taylor, A. Leach, Y . Long, and C. G. Willcocks, “Deep generative modelling: A comparative review of vaes, gans, normalizing flows, energy-based and autoregressive models,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 44, no. 11, pp. 7327–7347, 2022

    2022

  22. [30]

    Auto-encoding variational bayes,

    D. P. Kingma and M. Welling, “Auto-encoding variational bayes,” in International Conference on Learning Representations , Apr. 2014, pp. 1–14

    2014

  23. [31]

    Generative adversarial nets,

    I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial nets,” Advances in neural information processing systems , vol. 27, 2014

    2014

  24. [32]

    Diffusion models in vision: A survey,

    F.-A. Croitoru, V . Hondru, R. T. Ionescu, and M. Shah, “Diffusion models in vision: A survey,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 45, no. 9, pp. 10 850–10 869, 2023

    2023

  25. [33]

    An introduction to variational autoen- coders,

    D. P. Kingma and M. Welling, “An introduction to variational autoen- coders,” Foundations and Trends in Machine Learning , vol. 12, no. 4, pp. 307–392, Nov. 2019

    2019

  26. [34]

    Learning structured output representation using deep conditional generative models,

    K. Sohn, H. Lee, and X. Yan, “Learning structured output representation using deep conditional generative models,” in Advances in Neural Information Processing Systems , C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett, Eds., vol. 28. Curran Associates, Inc., 2015. ...

    2015

  27. [35]

    Semi-supervised learning with deep generative models,

    D. P. Kingma, S. Mohamed, D. Jimenez Rezende, and M. Welling, “Semi-supervised learning with deep generative models,” Advances in neural information processing systems , vol. 27, 2014

    2014

  28. [36]

    Deepsense 6G: A large-scale real-world multi-modal sensing and communication dataset,

    A. Alkhateeb, G. Charan, T. Osman, A. Hredzak, J. Morais, U. Demirhan, and N. Srinivas, “Deepsense 6G: A large-scale real-world multi-modal sensing and communication dataset,” IEEE Commun. Mag., vol. 61, no. 9, pp. 122–128, 2023

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.