Pith. sign in

REVIEW 4 major objections 6 minor 102 references

Learning Coarse-to-Fine Pruning of Graph Convolutional Networks for Skeleton-based Recognition

T0 review · 4 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read This paper claims that composing a coarse, group-level pruning mask with a fine, entry-level mask gives skeleton-recognition graph convolutional networks the speed of structured pruning and the accuracy of unstructured pruning, reporting…

desk verdict Plausible coarse-to-fine pruning idea, but Eq. 2's undefined P_r/P_c make the central mechanism uncheckable. read the letter →

arxiv 2412.12887 v1 pith:2NO3ZZJV submitted 2024-12-17 cs.CV

classification cs.CV
keywords coarse-to-finepruningfine-grainedgraphconvolutionalnetworksskeleton-basedrecognitionmagnitudestructuredunstructuredlightweightnetworkdesign
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

The paper tries to resolve a trade-off in neural-network pruning: removing whole groups of weights (structured pruning) makes networks faster but hurts accuracy, while removing individual weights (unstructured pruning) preserves accuracy but leaves the network just as slow. Its proposed coarse-to-fine (CTF) parametrization writes each connection mask as the Hadamard product of a coarse mask, built from channel-, column-, and row-wise pruning, and a fine entry-wise magnitude mask, so individual weights are cut only where no coarser cut applies. Trained with a cross-entropy loss plus a budget term that targets a pruning cost exactly, CTF is evaluated on skeleton-based action and hand-gesture recognition. On SBU at 98% pruning it reports 76.92% accuracy with a 43x speedup, compared with 49.23% accuracy at 235x for structured-only pruning and 90.76% accuracy with no speedup for unstructured-only pruning; the FPHA results show the same pattern. The claim a sympathetic reader would take away is that the mask composition itself, not any particular architecture, is what buys the accuracy-speed tradeoff.

What carries the argument

The carrying object is the CTF mask parametrization of Eq. 3, $\psi(\hat{W}) = \psi_c(\psi_f(\hat{W})) \odot \psi_f(\hat{W})$. The coarse map $\psi_c$ applies row-, column-, and block/channel-wise aggregations through binary adjacency matrices $P_r$ and $P_c$ (with a vectorization map $\varphi$ and its inverse), and the fine map $\psi_f$ is the symmetrized shifted sigmoid $2\,\mathrm{sigmoid}(\sigma \hat{W}^2) - 1$ acting entry-wise on magnitudes. The Hadamard product makes the granularities interact: a weight survives only if it passes both gates, so the network can delete large contiguous blocks for speed while carving out individual survivors where block deletion would be too destructive. The training objective adds a budget loss $\lambda(\sum_\ell \psi(\hat{W}_\ell) - c)^2$ with $\lambda$ large, driving the learned masks to the desired pruning rate; the ablation removes either factor and the trade-off collapses to one side or the other.

What would settle it

Take a small weight matrix $\hat{W}$, define $P_r$ and $P_c$ as Eq. 2 requires, and compute $\psi_c(\psi_f(\hat{W}))$; check whether every entry in each row, column, and block receives identical mask values and whether the masks are actually binary after training. If the coarse mask is not group-wise constant, or if $P_r$ and $P_c$ cannot be trained to crisp group patterns, the structured speedup claimed by CTF does not follow from the parametrization; equivalently, randomizing $P_r$ and $P_c$ on the SBU experiment should change the accuracy-speedup tradeoff if the grouping mechanism is doing the work.

Watch

Extended reading notes

Core claim

The central discovery is that pruning masks need not be chosen as either structured or unstructured; they can be parameterized as $\psi(\hat{W}) = \psi_c(\psi_f(\hat{W})) \odot \psi_f(\hat{W})$, where $\psi_f(\hat{W}) = 2\,\mathrm{sigmoid}(\sigma \hat{W}^2) - 1$ is a differentiable, nearly binary entry-wise magnitude gate, and $\psi_c(\hat{W})$ is built from three group-wise terms: $\varphi^{-1}(P_r \varphi(\hat{W}))$ for row-wise pruning, $\varphi^{-1}(\varphi(\hat{W})^\top P_c)$ for column-wise, and $\varphi^{-1}(P_r P_c^\top \varphi(\hat{W}))$ for block/channel-wise. For crisp masks, the composition gives block pruning highest priority, then column, then row, then entry pruning, so fine-grained cuts are enabled only where coarse-grained ones are disabled. The paper claims this yields both efficient computation and high accuracy, and supports the claim with ablations on SBU and FPHA where CTF at 98% pruning reaches 76.92% and 70.95% accuracy with 43x and 41x speedups, respectively, outperforming structured-only and regularization-based alternatives and preserving a real speedup that unstructured-only pruning cannot provide.

Load-bearing premise

The construction rests on the assumption that the binary matrices $P_r$ and $P_c$ in Eq. 2 genuinely implement row-wise, column-wise, and block-wise grouping of the weight tensor, so that the coarse mask $\psi_c$ removes whole groups and can be optimized into crisp values; the paper asserts this in Section 4.1 but gives no construction, proof, or learning rule for them.

Editorial extensions

If this is right

  • At an aggressive 98% pruning rate, CTF lands between the extremes on both datasets: on SBU it keeps 76.92% accuracy (versus 49.23% for structured-only and 90.76% for unstructured-only) while delivering a 43x speedup that unstructured-only pruning does not provide.
  • The budget term $\lambda(\sum_\ell \psi(\hat{W}_\ell) - c)^2$ lets the same parametrization hit any targeted pruning rate; once the budget is met, the classification loss dominates the gradient and continues to improve accuracy.
  • CTF produces compact mask tensors with structured zeros plus only a few individual cuts, a configuration the paper argues is the one most likely to yield real inference speedups on standard hardware.
  • In the comparison against regularization-based variational pruning at 98% rate, CTF achieves higher accuracy on both SBU (76.92% versus 75.38% for the best regularizer) and FPHA (70.95% versus 70.78%), and the regularizer baselines report no actual speedup.

Reading between the lines

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

  • A testable extension is to learn $P_r$ and $P_c$ rather than treating them as fixed; if rows and columns do not match a semantically meaningful layout, learned grouping might let CTF apply to weight tensors without a natural 2D structure, such as transformer or recurrent layers.
  • The reported speedups are based on reduced tensor shapes; on hardware that only accelerates dense operations, the 43x and 41x figures would need to be re-measured as wall-clock time, so the practical advantage over other regularizers could be smaller.
  • Since the coarse gate is driven by weight magnitude, CTF implicitly assumes that the smallest-magnitude groups are the least important ones; for layers where that assumption fails, the fine entry-wise gate would have to compensate, and the trade-off could degrade.
  • The composition resembles a two-stage ticket-selection process (coarse mask picks a block-sparse skeleton, fine mask refines survivors); a natural test the paper does not run is whether CTF-pruned subnetworks, re-initialized from scratch, retrain as fast and as accurately as the original network.
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 / 6 minor

Summary. The manuscript proposes a coarse-to-fine (CTF) pruning method for graph convolutional networks (GCNs) on skeleton-based action and hand-gesture recognition. The method defines a mask parametrization ψ(Ŵ)=ψc(ψf(Ŵ))⊙ψf(Ŵ), where ψf is a symmetrized shifted sigmoid that performs fine-grained magnitude pruning and ψc is a coarse-grained parametrization intended to remove rows, columns, and blocks/channels. A variational objective (Eq. 4) combines cross-entropy with a quadratic budget penalty to reach a target pruning rate c. Experiments on SBU and FPHA compare coarse, fine, and CTF pruning, reporting that CTF achieves intermediate accuracy and speedup (e.g., 76.92% accuracy and 43× speedup on SBU at 98% pruning).

Significance. The idea of interpolating group-level and entry-level pruning through a single differentiable mask is interesting and, if made operational, could be useful for deploying GCNs on resource-constrained devices. The empirical pattern across Tables 3 and 4 — CTF lies between coarse and fine pruning in both accuracy and speedup — is consistent. However, the central construction in Eq. 2 is currently non-operational because the matrices P_r and P_c are never specified, no implementation or code is provided, and the speedup numbers lack measurement details. These gaps prevent independent verification of the central claim, so the contribution is best evaluated after substantial revision.

major comments (4)
  1. [Section 4.1, Eq. (2)] The matrices P_r and P_c are described only as 'adjacency matrices that model the neighborhood across rows and columns'; no construction, initialization, normalization, or update rule is given. Consequently, the three terms in Eq. (2) cannot be checked to behave as row-wise, column-wise, and block-wise group masks, and the structured-pruning behavior that underlies the speedups in Tables 3 and 4 is not verifiable. Moreover, if P_r (or P_c) contains multiple ones in a single row, the corresponding mask term is not bounded in [0,1], so ψ=ψc⊙ψf need not be a valid mask. The authors should specify P_r and P_c explicitly, including how groups are defined and how gradients are computed, or provide code.
  2. [Section 4.1, after Eq. (3)] The claim that 'block-wise pruning has the highest priority, followed by column-wise and then row-wise pruning' is not supported by Eq. (3). The Hadamard product is commutative, and any zero factor removes an entry regardless of the order in which the factors are written. A priority ordering requires an additional mechanism, such as sequential masking with re-evaluation of remaining weights or a hierarchical gating that is not commutative. As written, the parametrization cannot implement the described priority.
  3. [Section 4.2, Eq. (4)] The budget term (Σ ψ(Ŵℓ) − c)² with λ=1000 does not by itself guarantee that the final pruning rate equals c; it only penalizes deviation. The paper does not describe how the exact rates in Tables 3 and 4 are obtained (e.g., thresholding after training, re-normalization, or a stopping criterion), nor how the budget is distributed across layers. Since the central comparisons are made at specific rates (70%, 90%, 95%, 98%), this missing mechanism is load-bearing.
  4. [Section 5, Tables 3 and 4] The speedup values ('none', '6 ×', '43 ×', etc.) are not accompanied by any measurement protocol. It is unclear whether these are theoretical FLOP counts, parameter-count ratios, or wall-clock timings on the GeForce GTX 1070, and whether they include memory access patterns. Without a measurement protocol, the reported accuracy-speedup tradeoff cannot be reproduced or compared with other pruning work.
minor comments (6)
  1. [Figure 2 caption] The caption uses the term 'semi-structured pruning' while the text refers to the same setting as coarse-to-fine; please unify the terminology.
  2. [Section 5, first paragraph] The phrase 'according to to tables 3-4' contains a duplicated 'to'.
  3. [Section 4.2, Eq. (4)] The variable c is not defined as a fraction or an absolute count; please clarify its relation to the total number of weights in the network.
  4. [Section 4.1, fine-grained parametrization] The annealing schedule for σ is only described as 'annealed'; the paper should specify the initial value, the schedule, and the stopping criterion for convergence to crisp binary masks.
  5. [Tables 3 and 4] The comparative regularization-based pruning rows report only accuracy and not speedup; reporting speedup for these baselines is necessary for a fair evaluation of the claimed tradeoff.
  6. [General] There is no code or data availability statement; given that the construction in Eq. (2) is not fully specified, releasing code or the exact P_r/P_c construction would greatly improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: CTF accuracy and speedup numbers are measured, and the core parametrization is defined in Eqs. 1-4 without importing a load-bearing self-citation.

full rationale

The paper's central claim, tested in Sec. 5, is that the CTF mask ψ(W)=ψc(ψf(W))⊙ψf(W) gives the best accuracy/speedup tradeoff. That claim is supported by empirically measured accuracies and speedups under the budget loss in Eq. 4, which explicitly enforces the pruning rate; no fitted parameter is relabeled as a prediction. The parametrization in Eqs. 1-3 is defined in the paper itself, and the fine-grained sigmoid ψf(W)=2sigmoid(σW^2)-1 is explicitly stated. No uniqueness theorem, external result, or prior-work ansatz is invoked to force the CTF choice, so the many self-citations in the reference list are not load-bearing. Two non-circularity concerns should be flagged explicitly. First, Sec. 4.1, Eq. 2 defines the coarse mask via matrices P_r and P_c described only as 'adjacency matrices that model the neighborhood across respectively the rows and the columns of W', with no construction, initialization, normalization, or learning rule; this is an omitted-support/reproducibility defect, not circularity, because the reported numbers are not derived from any explicit form of P_r/P_c. Second, the statement 'From Eqs. 2 and 3 ... block-wise pruning has the highest priority, followed by column-wise and then row-wise pruning' is not entailed by the equations, since the Hadamard product is commutative and any zero factor removes the entry; this is a correctness gap, not a circular reduction. Overall, the derivation chain does not reduce to its own inputs, and the accuracy results are externally measured on SBU and FPHA, so the appropriate circularity score is 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on several design choices and assumptions that are not derived from first principles. The free parameters are the pruning rate, the budget weight, the sigmoid scaling schedule, the learning-rate schedule, and the undefined grouping matrices P_r and P_c. The axioms are standard magnitude-pruning heuristics plus the paper-specific assertion that P_r/P_c encode group neighborhoods and that annealed sigmoids become crisp. No new physical entities are introduced; the main gap is that the core coarse-mask operators are both free parameters and axioms.

free parameters (5)
  • Pruning rate c = 0.90, 0.95, 0.98 in experiments
    The budget loss explicitly enforces the sum of masks to reach c, so results are reported at these enforced rates. c is an input design choice, not a predicted quantity.
  • Budget weight lambda = 1000
    Set to 1000 in practice to make the budget loss dominate; the tradeoff between accuracy and pruning depends on this choice.
  • Sigmoid scaling sigma = not reported (annealed)
    Controls crispness of masks; the annealing schedule is not specified, so final masks depend on an unreported schedule.
  • Learning rate schedule parameters = multipliers 0.99 and 1/0.99; initial value not stated
    Training dynamics and final accuracy depend on this schedule, and the initial learning rate is not given.
  • Group neighborhood matrices P_r and P_c = not specified
    The row, column, and block grouping in Eq. 2 is defined through these matrices, but their construction or learning is never described. This is both a free design choice and a reproducibility gap.
assumptions (4)
  • domain assumption Weight magnitude is a valid importance proxy for pruning.
    The fine mask removes weights with smallest magnitude; this is the standard magnitude pruning assumption and is not independently justified in the paper.
  • ad hoc to paper P_r and P_c implement row-wise, column-wise, and block-wise neighborhoods.
    Eq. 2 asserts these matrices model the neighborhoods without defining or proving it; this is a load-bearing assumption for the coarse parametrization.
  • ad hoc to paper Annealed sigmoid masks become crisp binary values at convergence.
    The paper relies on psi_f approaching 0/1 values, but no convergence or binarization guarantee is shown.
  • domain assumption Measured speedups reflect the structured mask pattern on standard hardware.
    The speedup numbers assume row, column, and block sparsity translates to actual inference acceleration; the measurement protocol is not described.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Coarse-to-Fine Pruning of Graph Convolutional Networks for Skeleton-based Recognition." pith.science (2026). https://pith.science/paper/2NO3ZZJV

@misc{pith2026241212887,
  author       = {Pith},
  title        = {Pith review of: Learning Coarse-to-Fine Pruning of Graph Convolutional Networks for Skeleton-based Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2NO3ZZJV}},
  note         = {Machine review of arXiv:2412.12887}
}
read the original abstract

Magnitude Pruning is a staple lightweight network design method which seeks to remove connections with the smallest magnitude. This process is either achieved in a structured or unstructured manner. While structured pruning allows reaching high efficiency, unstructured one is more flexible and leads to better accuracy, but this is achieved at the expense of low computational performance. In this paper, we devise a novel coarse-to-fine (CTF) method that gathers the advantages of structured and unstructured pruning while discarding their inconveniences to some extent. Our method relies on a novel CTF parametrization that models the mask of each connection as the Hadamard product involving four parametrizations which capture channel-wise, column-wise, row-wise and entry-wise pruning respectively. Hence, fine-grained pruning is enabled only when the coarse-grained one is disabled, and this leads to highly efficient networks while being effective. Extensive experiments conducted on the challenging task of skeleton-based recognition, using the standard SBU and FPHA datasets, show the clear advantage of our CTF approach against different baselines as well as the related work.

Figures

Figures reproduced from arXiv: 2412.12887 by the authors.

Figure 1
Figure 1. This figure shows the CTF pruning process in Eq. 3; here each diagonal block corresponds to a [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. This figure shows a crop of the mask tensor of the second (multi-head-attention) layer of our [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

102 extracted references · 54 canonical work pages

  1. [1]

    Imagenet classification with deep convolutional neural networks

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84–90, 2017

  2. [2]

    H. Sahbi. Phase-field models for lightweight graph convolutional networks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4643–4649, 2023

  3. [3]

    Semi-supervised classification with graph convolutional networks

    Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016

  4. [4]

    Transductive kernel map learning and its application to image annotation

    P. Vo and H. Sahbi. "Transductive kernel map learning and its application to image annotation." BMVC. 2012

  5. [5]

    Adaptive graph convolutional neural networks

    Ruoyu Li, Sheng Wang, Feiyun Zhu, and Junzhou Huang. Adaptive graph convolutional neural networks. InProceedings of the AAAI, volume 32, 2018

  6. [6]

    Directedacyclicgraphkernelsforactionrecognition

    L.WangandH.Sahbi. Directedacyclicgraphkernelsforactionrecognition. In ProceedingsoftheIEEEInternationalConference on Computer Vision, pages 3168–3175, 2013

  7. [7]

    A new model for learning in graph domains

    Marco Gori, Gabriele Monfardini, and Franco Scarselli. A new model for learning in graph domains. InIJCNN, volume 2, pages 729–734. IEEE, 2005

  8. [8]

    Robust face recognition using dynamic space warping

    H. Sahbi and N. Boujemaa. "Robust face recognition using dynamic space warping." International Workshop on Biometric Authentication. Springer, Berlin, Heidelberg, 2002

Show all 102 references
  1. [9]

    Understanding attention and generalization in graph neural networks

    Boris Knyazev, Graham W Taylor, and Mohamed Amer. Understanding attention and generalization in graph neural networks. Advances in NIPS, 32, 2019

  2. [10]

    Learning attribute representations for remote sensing ship category classification

    Q. Oliveau and H. Sahbi. "Learning attribute representations for remote sensing ship category classification." IEEE JSTARS 10.6 (2017): 2830-2840

  3. [11]

    Improved knowledge distillation via teacher assistant

    Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. InProceedings of the AAAI, volume 34, pages 5191–5198, 2020

  4. [12]

    learning-compression

    Miguel A Carreira-Perpinán and Yerlan Idelbayev. “learning-compression” algorithms for neural net pruning. InProceedings of the IEEE CVPR, pages 8532–8541, 2018

  5. [13]

    Relevance feedback for satellite image change detection

    H. Sahbi. "Relevance feedback for satellite image change detection." IEEE ICASSP, 2013

  6. [14]

    Morphnet: Fast & simple resource-constrained structure learning of deep networks

    Ariel Gordon, Elad Eban, Ofir Nachum, Bo Chen, Hao Wu, Tien-Ju Yang, and Edward Choi. Morphnet: Fast & simple resource-constrained structure learning of deep networks. InProceedings of the IEEE CVPR, pages 1586–1595, 2018

  7. [15]

    Sparse artificial neural networks using a novel smoothed lasso penalization

    Basava Naga Girish Koneru and Vinita Vasudevan. Sparse artificial neural networks using a novel smoothed lasso penalization. IEEE Transactions on Circuits and Systems II: Express Briefs, 66(5):848–852, 2019

  8. [16]

    Entropy-constrained training of deep neural networks

    Simon Wiedemann, Arturo Marban, Klaus-Robert Müller, and Wojciech Samek. Entropy-constrained training of deep neural networks. InIJCNN, 2019

  9. [17]

    Sahbi, J-Y

    H. Sahbi, J-Y. Audibert and R. Keriven. Context-dependent kernels for object classification.IEEE transactions on pattern analysis and machine intelligence, 33(4):699–708, 2011

  10. [18]

    Learning structured sparsity in deep neural networks

    Wei Wen, Chunpeng Wu, Yandan Wang, Yiran Chen, and Hai Li. Learning structured sparsity in deep neural networks. Advances in NIPS, 29, 2016

  11. [19]

    Learning efficient convolutional networks through network slimming

    Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learning efficient convolutional networks through network slimming. InProceedings of the IEEE ICCV, pages 2736–2744, 2017

  12. [20]

    Constrained optical flow for aerial image change detection

    N. Bourdis, D. Marraud and H. Sahbi. "Constrained optical flow for aerial image change detection." in IEEE IGARSS, 2011

  13. [21]

    Learning sparse neural networks throughl_0 regularization

    Christos Louizos, Max Welling, and Diederik P Kingma. Learning sparse neural networks throughl_0 regularization. arXiv preprint arXiv:1712.01312, 2017

  14. [22]

    Searching for mobilenetv3

    Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. InProceedings of the IEEE/CVF ICCV, pages 1314–1324, 2019

  15. [23]

    Wang and H

    L. Wang and H. Sahbi. Bags-of-daglets for action recognition. InIEEE International Conference on Image Processing (ICIP), pages 1550–1554. IEEE, 2014

  16. [24]

    Convolutional two-stream network fusion for video action recognition

    Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Convolutional two-stream network fusion for video action recognition. InIEEE CVPR, 2016

  17. [25]

    Transition forests: Learning discriminative temporal transitions for action recognition and detection

    Guillermo Garcia-Hernando and Tae-Kyun Kim. Transition forests: Learning discriminative temporal transitions for action recognition and detection. InProceedings of the IEEE CVPR, pages 432–440, 2017

  18. [26]

    Bourdis, D

    N. Bourdis, D. Marraud, and H. Sahbi, Spatio-temporal interaction for aerial video change detection, in IGARSS, 2012, pp. 2253–2256

  19. [27]

    First-person hand action benchmark with rgb-d videos and 3d hand pose annotations

    Guillermo Garcia-Hernando, Shanxin Yuan, Seungryul Baek, and Tae-Kyun Kim. First-person hand action benchmark with rgb-d videos and 3d hand pose annotations. InProceedings of the IEEE CVPR, pages 409–419, 2018

  20. [28]

    Mazari and H

    A. Mazari and H. Sahbi. Deep temporal pyramid design for action recognition. InIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2077–2081. IEEE, 2019

  21. [29]

    Mazari and H

    A. Mazari and H. Sahbi. Mlgcn: Multi-laplacian graph convolutional networks for human action recognition. InThe British Machine Vision Conference (BMVC), 2019

  22. [30]

    Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding.arXiv preprint arXiv:1510.00149, 2015

    Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding.arXiv preprint arXiv:1510.00149, 2015

  23. [31]

    Learning both weights and connections for efficient neural network

    Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. Advances in NIPS, 28, 2015. 11

  24. [32]

    Nonlinear cross-view sample enrichment for action recognition

    L. Wang and H. Sahbi. "Nonlinear cross-view sample enrichment for action recognition." European Conference on Computer Vision. Springer, Cham, 2014

  25. [33]

    Optimal brain damage.Advances in NIPS, 2, 1989

    Yann LeCun, John Denker, and Sara Solla. Optimal brain damage.Advances in NIPS, 2, 1989

  26. [34]

    Second order derivatives for network pruning: Optimal brain surgeon.Advances in NIPS, 5, 1992

    Babak Hassibi and David Stork. Second order derivatives for network pruning: Optimal brain surgeon.Advances in NIPS, 5, 1992

  27. [35]

    Coarse-to-fine deep kernel networks

    H. Sahbi. "Coarse-to-fine deep kernel networks." IEEE ICCV-W, 2017

  28. [36]

    Jointly learning heterogeneous features for rgb-d activity recognition

    Jian-Fang Hu, Wei-Shi Zheng, Jianhuang Lai, and Jianguo Zhang. Jointly learning heterogeneous features for rgb-d activity recognition. InCVPR, 2015

  29. [37]

    Spatio-temporal graph convolution for skeleton based action recognition

    Chaolong Li, Zhen Cui, Wenming Zheng, Chunyan Xu, and Jian Yang. Spatio-temporal graph convolution for skeleton based action recognition. InAAAI, 2018

  30. [38]

    Global co-occurrence feature learning and active coordinate system conversion for skeleton-based action recognition

    Sheng Li, Tingting Jiang, Tiejun Huang, and Yonghong Tian. Global co-occurrence feature learning and active coordinate system conversion for skeleton-based action recognition. InProceedings of the IEEE/CVF WACV, pages 586–594, 2020

  31. [39]

    Laplacian deep kernel learning for image annotation

    M. Jiu and H. Sahbi. "Laplacian deep kernel learning for image annotation." IEEE ICASSP, 2016

  32. [40]

    Graph cnns with motif and variable temporal block for skeleton-based action recognition

    Yu-Hui Wen, Lin Gao, Hongbo Fu, Fang-Lue Zhang, and Shihong Xia. Graph cnns with motif and variable temporal block for skeleton-based action recognition. InProceedings of the AAAI, volume 33, pages 8989–8996, 2019

  33. [41]

    Topologically-consistentmagnitudepruningforverylightweightgraphconvolutionalnetworks

    H.Sahbi. Topologically-consistentmagnitudepruningforverylightweightgraphconvolutionalnetworks. In IEEEInternational Conference on Image Processing (ICIP), pages 3495–3499. IEEE, 2022

  34. [42]

    Spatial temporal graph convolutional networks for skeleton-based action recognition

    Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, vol 32, 2018

  35. [43]

    A riemannian network for spd matrix learning

    Zhiwu Huang and Luc Van Gool. A riemannian network for spd matrix learning. InProceedings of the AAAI conference on artificial intelligence, volume 31, 2017

  36. [44]

    Sahbi and F

    H. Sahbi and F. Fleuret. Scale-invariance of support vector machines based on the triangular kernel. Diss. INRIA, 2002

  37. [45]

    Jiu and H

    M. Jiu and H. Sahbi. Deep representation design from deep kernel networks.Pattern Recognition, 88:447–457, 2019

  38. [46]

    A novel geometric framework on gram matrix trajectories for human behavior understanding.IEEE TPAMI, 42(1):1–14, 2018

    Anis Kacem, Mohamed Daoudi, Boulbaba Ben Amor, Stefano Berretti, and Juan Carlos Alvarez-Paiva. A novel geometric framework on gram matrix trajectories for human behavior understanding.IEEE TPAMI, 42(1):1–14, 2018

  39. [47]

    Camera pose estimation using visual servoing for aerial video change detection

    N. Bourdis, D. Marraud and H. Sahbi. "Camera pose estimation using visual servoing for aerial video change detection." IEEE IGARSS 2012

  40. [48]

    H. Sahbi. Coarse-to-fine support vector machines for hierarchical face detection. Diss. PhD thesis, Versailles University, 2003

  41. [49]

    Han: An efficient hierarchical self-attention network for skeleton-based gesture recognition.arXiv preprint arXiv:2106.13391, 2021

    Jianbo Liu, Ying Wang, Shiming Xiang, and Chunhong Pan. Han: An efficient hierarchical self-attention network for skeleton-based gesture recognition.arXiv preprint arXiv:2106.13391, 2021

  42. [50]

    Building deep networks on grassmann manifolds

    Zhiwu Huang, Jiqing Wu, and Luc Van Gool. Building deep networks on grassmann manifolds. InProceedings of the AAAI, volume 32, 2018

  43. [51]

    Decoupled representation learning for skeleton-based gesture recognition

    Jianbo Liu, Yongcheng Liu, Ying Wang, Veronique Prinet, Shiming Xiang, and Chunhong Pan. Decoupled representation learning for skeleton-based gesture recognition. InProceedings of the IEEE/CVF CVPR, pages 5751–5760, 2020

  44. [52]

    H. Sahbi. Kernel-based graph convolutional networks. In25th International Conference on Pattern Recognition (ICPR), pages 4887–4894. IEEE, 2021

  45. [53]

    Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014

  46. [54]

    Structured pruning of neural networks with budget-aware regularization

    Carl Lemaire, Andrew Achkar, and Pierre-Marc Jodoin. Structured pruning of neural networks with budget-aware regularization. InCVPR, pages 9108–9116, 2019

  47. [55]

    Bags-of-daglets for action recognition

    L. Wang and H. Sahbi. "Bags-of-daglets for action recognition." 2014 IEEE International Conference on Image Processing (ICIP). IEEE, 2014

  48. [56]

    Pruning filters for efficient convnets.arXiv preprint arXiv:1608.08710, 2016

    Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets.arXiv preprint arXiv:1608.08710, 2016

  49. [57]

    H. Sahbi. Learning connectivity with graph convolutional networks. In25th International Conference on Pattern Recognition (ICPR), pages 9996–10003. IEEE, 2021

  50. [58]

    Hierarchical recurrent neural network for skeleton based action recognition

    Yong Du, Wei Wang, and Liang Wang. Hierarchical recurrent neural network for skeleton based action recognition. InCVPR, pages 1110–1118, 2015

  51. [59]

    Spatio-temporal lstm with trust gates for 3d human action recognition

    Jun Liu, Amir Shahroudy, Dong Xu, and Gang Wang. Spatio-temporal lstm with trust gates for 3d human action recognition. In ECCV 2016

  52. [60]

    Interactive satellite image change detection with context-aware canonical correlation analysis

    H. Sahbi. "Interactive satellite image change detection with context-aware canonical correlation analysis." IEEE GRSL, (14)5, 2017

  53. [61]

    Skeleton-based human action recognition with global context-aware attention lstm networks.IEEE Transactions on Image Processing, 27(4):1586–1599, 2017

    Jun Liu, Gang Wang, Ling-Yu Duan, Kamila Abdiyeva, and Alex C Kot. Skeleton-based human action recognition with global context-aware attention lstm networks.IEEE Transactions on Image Processing, 27(4):1586–1599, 2017

  54. [62]

    View adaptive recurrent neural networks for high performance human action recognition from skeleton data

    Pengfei Zhang, Cuiling Lan, Junliang Xing, Wenjun Zeng, Jianru Xue, and Nanning Zheng. View adaptive recurrent neural networks for high performance human action recognition from skeleton data. InICCV, pages 2117–2126, 2017

  55. [63]

    H. Sahbi. Kernel pca for similarity invariant shape recognition.Neurocomputing, 70(16-18):3034–3045, 2007

  56. [64]

    Co-occurrence feature learning for skeleton based action recognition using regularized deep lstm networks

    Wentao Zhu, Cuiling Lan, Junliang Xing, Wenjun Zeng, Yanghao Li, Li Shen, and Xiaohui Xie. Co-occurrence feature learning for skeleton based action recognition using regularized deep lstm networks. InAAAI, volume 30, 2016

  57. [65]

    Deepgru: Deep gesture recognition utility

    Mehran Maghoumi and Joseph J LaViola. Deepgru: Deep gesture recognition utility. InISVC 2019

  58. [66]

    Rgb-d-based human motion recognition with deep learning: A survey.CVIU, 2018

    Pichao Wang, Wanqing Li, Philip Ogunbona, Jun Wan, and Sergio Escalera. Rgb-d-based human motion recognition with deep learning: A survey.CVIU, 2018

  59. [67]

    H. Sahbi. Lightweight connectivity in graph convolutional networks for skeleton-based recognition. InIEEE International Conference on Image Processing (ICIP), pages 2329–2333. IEEE, 2021

  60. [68]

    Linear-time online action detection from 3d skeletal data using bags of gesturelets

    Moustafa Meshry, Mohamed E Hussein, and Marwan Torki. Linear-time online action detection from 3d skeletal data using bags of gesturelets. In2016 IEEE WACV, pages 1–9. IEEE, 2016. 12

  61. [69]

    Eshed Ohn-Bar and Mohan Manubhai Trivedi. Hand gesture recognition in real time for automotive interfaces: A multimodal vision-based approach and evaluations.IEEE transactions on intelligent transportation systems, 15(6):2368–2377, 2014

  62. [70]

    Applying interest operators in semi-fragile video watermarking

    S. Thiemert, H. Sahbi, and M. Steinebach. "Applying interest operators in semi-fragile video watermarking." Security, Steganography, and Watermarking of Multimedia Contents VII. Vol. 5681. SPIE, 2005

  63. [71]

    Dropneuron: Simplifying the structure of deep neural networks.arXiv preprint arXiv:1606.07326, 2016

    Wei Pan, Hao Dong, and Yike Guo. Dropneuron: Simplifying the structure of deep neural networks.arXiv preprint arXiv:1606.07326, 2016

  64. [72]

    Using entropy for image and video authentication watermarks

    S. Thiemert, H. Sahbi, and M. Steinebach. "Using entropy for image and video authentication watermarks." Security, Steganography, and Watermarking of Multimedia Contents VIII. Vol. 6072. SPIE, 2006

  65. [73]

    An end-to-end spatio-temporal attention model for human action recognition from skeleton data

    Sijie Song, Cuiling Lan, Junliang Xing, Wenjun Zeng, and Jiaying Liu. An end-to-end spatio-temporal attention model for human action recognition from skeleton data. InAAAI, volume 31, 2017

  66. [74]

    Human action recognition by representing 3d skeletons as points in a lie group

    Raviteja Vemulapalli, Felipe Arrate, and Rama Chellappa. Human action recognition by representing 3d skeletons as points in a lie group. InCVPR, 2014

  67. [75]

    From coarse to fine skin and face detection

    H. Sahbi and N. Boujemaa. "From coarse to fine skin and face detection." Proceedings of the eighth ACM international conference on Multimedia. 2000

  68. [76]

    Regularization of neural networks using dropconnect

    Li Wan, Matthew Zeiler, Sixin Zhang, Yann Le Cun, and Rob Fergus. Regularization of neural networks using dropconnect. In ICML, pages 1058–1066. PMLR, 2013

  69. [77]

    Effective 3d action recognition using eigenjoints.Journal of Visual Communication and Image Representation, 25(1):2–11, 2014

    Xiaodong Yang and YingLi Tian. Effective 3d action recognition using eigenjoints.Journal of Visual Communication and Image Representation, 25(1):2–11, 2014

  70. [78]

    Yuan, G-S

    F. Yuan, G-S. Xia, H. Sahbi, and V. Prinet. Mid-level features and spatio-temporal context for activity recognition.Pattern Recognition, 45(12):4182–4191, 2012

  71. [79]

    Interactive body part contrast mining for human interaction recognition

    Yanli Ji, Guo Ye, and Hong Cheng. Interactive body part contrast mining for human interaction recognition. InICMEW, pages 1–6. IEEE, 2014

  72. [80]

    Category-blind human action recognition: A practical recognition system

    Wenbo Li, Longyin Wen, Mooi Choo Chuah, and Siwei Lyu. Category-blind human action recognition: A practical recognition system. InIEEE ICCV, 2015

  73. [81]

    Sahbi and F

    H. Sahbi and F. Fleuret. Kernel methods and scale invariance using the triangular kernel. Technical report, INRIA, 2004

  74. [82]

    Hon4d: Histogram of oriented 4d normals for activity recognition from depth sequences

    Omar Oreifej and Zicheng Liu. Hon4d: Histogram of oriented 4d normals for activity recognition from depth sequences. In CVPR, pages 716–723, 2013

  75. [83]

    3d action recognition from novel viewpoints

    Hossein Rahmani and Ajmal Mian. 3d action recognition from novel viewpoints. InCVPR, pages 1506–1515, 2016

  76. [84]

    Sahbi and D

    H. Sahbi and D. Geman. A hierarchy of support vector machines for pattern detection.Journal of Machine Learning Research, 7(10), 2006

  77. [85]

    Two-person interaction detection using body-pose features and multiple instance learning

    Kiwon Yun, Jean Honorio, Debaleena Chattopadhyay, Tamara L Berg, and Dimitris Samaras. Two-person interaction detection using body-pose features and multiple instance learning. In CVPRW, pages 28–35. IEEE, 2012

  78. [86]

    Jiu and H

    M. Jiu and H. Sahbi. Nonlinear deep kernel learning for image annotation.IEEE Transactions on Image Processing, 26(4):1820–1832, 2017

  79. [87]

    The moving pose: An efficient 3d kinematics descriptor for low-latency action recognition and detection

    Mihai Zanfir, Marius Leordeanu, and Cristian Sminchisescu. The moving pose: An efficient 3d kinematics descriptor for low-latency action recognition and detection. InICCV, pages 2752–2759, 2013

  80. [88]

    Graph-cut transducers for relevance feedback in content based image retrieval

    H. Sahbi, Jean-Yves Audibert, and Renaud Keriven. "Graph-cut transducers for relevance feedback in content based image retrieval." 2007 IEEE 11th International Conference on Computer Vision. IEEE, 2007

  81. [89]

    Efficient temporal sequence comparison and classification using gram matrix embeddings on a riemannian manifold

    Xikang Zhang, Yin Wang, Mengran Gou, Mario Sznaier, and Octavia Camps. Efficient temporal sequence comparison and classification using gram matrix embeddings on a riemannian manifold. InCVPR, pages 4498–4507, 2016

  82. [90]

    Context-dependent kernel design for object matching and recognition

    H. Sahbi, Jean-Yves Audibert, Jaonary Rabarisoa, and Renaud Keriven. "Context-dependent kernel design for object matching and recognition." In 2008 IEEE Conference on Computer Vision and Pattern Recognition, pp. 1-8. IEEE, 2008

  83. [91]

    Convolutional neural networks and long short-term memory for skeleton-based human activity and hand gesture recognition.Pattern Recognition, 76:80–94, 2018

    Juan C Nunez, Raul Cabido, Juan J Pantrigo, Antonio S Montemayor, and Jose F Velez. Convolutional neural networks and long short-term memory for skeleton-based human activity and hand gesture recognition.Pattern Recognition, 76:80–94, 2018

  84. [92]

    Pruning-as-search: Efficient neural architecture search via channel pruning and structural reparameterization

    LI, Yanyu, ZHAO, Pu, YUAN, Geng, et al. Pruning-as-search: Efficient neural architecture search via channel pruning and structural reparameterization. arXiv preprint arXiv:2206.01198, 2022

  85. [93]

    H. Sahbi. Learning laplacians in chebyshev graph convolutional networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 2064-2075)

  86. [94]

    TELECOM ParisTech at ImageClefphoto 2008: Bi-Modal Text and Image Retrieval with Diversity Enhancement

    M. Ferecatu and H. Sahbi. "TELECOM ParisTech at ImageClefphoto 2008: Bi-Modal Text and Image Retrieval with Diversity Enhancement." CLEF (Working Notes). 2008

  87. [95]

    CNRS-TELECOM ParisTech at ImageCLEF 2013 Scalable Concept Image Annotation Task: Winning Annotations with Context Dependent SVMs

    H. Sahbi. "CNRS-TELECOM ParisTech at ImageCLEF 2013 Scalable Concept Image Annotation Task: Winning Annotations with Context Dependent SVMs." CLEF (Working Notes). 2013

  88. [96]

    Coarse-to-fine support vector classifiers for face detection

    H. Sahbi and N. Boujemaa. "Coarse-to-fine support vector classifiers for face detection." Object recognition supported by user interaction for service robots. Vol. 3. IEEE, 2002

  89. [97]

    Visual content extraction for automatic semantic annotation of video news

    N. Boujemaa, F. Fleuret, V. Gouet, and H. Sahbi. "Visual content extraction for automatic semantic annotation of video news." In the proceedings of the SPIE Conference, San Jose, CA, vol. 6. 2004

  90. [98]

    Misalignment resilient cca for interactive satellite image change detection

    H. Sahbi. "Misalignment resilient cca for interactive satellite image change detection." 2016 23rd International Conference on Pattern Recognition (ICPR). IEEE, 2016

  91. [99]

    From 2D silhouettes to 3D object retrieval: contributions and benchmarking

    T. Napoléon and H. Sahbi. "From 2D silhouettes to 3D object retrieval: contributions and benchmarking." EURASIP Journal on Image and Video Processing 2010 (2010): 1-17

  92. [100]

    Semi supervised deep kernel design for image annotation

    M. Jiu and H. Sahbi. "Semi supervised deep kernel design for image annotation." 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2015

  93. [101]

    Deep kernel map networks for image annotation

    M. Jiu and H. Sahbi. "Deep kernel map networks for image annotation." 2016 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2016

  94. [102]

    Multi-view object matching and tracking using canonical correlation analysis

    M. Ferecatu and H. Sahbi. "Multi-view object matching and tracking using canonical correlation analysis." 2009 16th IEEE International Conference on Image Processing (ICIP). IEEE, 2009

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.