Pith. sign in

REVIEW 4 major objections 4 minor 51 references

Towards Learning Affine-Invariant Representations via Data-Efficient CNNs

T0 review · 4 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read Near-circular filters make CNNs affine-invariant with few images.

desk verdict Large low-shot gains are real, but the 'rotation-invariant' regularizer contributes only a small part; the architecture carries the result. read the letter →

arxiv 1909.00114 v1 pith:MPWYDU2U submitted 2019-08-31 cs.CV

Xenju Xu , Guanghui Wang , Alan Sullivan , Ziming Zhang This is my paper · ORCID
classification cs.CV
keywords affineinvariancemulti-scalemaxoutrotation-invariantregularizercircularfiltersdataefficiencyfew-shotlearningconvolutionalneuralnetworksaffNIST
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

By folding geometric priors directly into a convolutional network's design and loss, the paper claims that a CNN can become robust to translation, scale, and rotation without needing large transformed training sets. The network uses multi-scale maxout blocks, which select the strongest response across several receptive-field sizes, and a regularizer that pushes each 3x3 spatial filter toward a circular weight pattern. On affNIST, MNIST-rot, and Traffic Sign, the authors report consistent wins over the compared baselines, most dramatically in the low-data regime: with 10 training images per class, the method reaches 84.15 percent on Traffic Sign and 85.06 percent on affNIST. The broader significance is that explicit priors of this kind can substitute for training data, pointing toward a recipe for data-efficient invariance.

What carries the argument

The carrying object is a multi-scale maxout block paired with a rotation-invariant regularizer. In each block, three stacked 3x3 convolutions approximate filters of sizes 3x3, 5x5, and 7x7, and a maxout operator keeps the maximum response per pixel across those scales, absorbing translation and scale variability. The regularizer in Eq. 4 is a variance-style penalty: for each 3x3 filter it measures the squared distance of every non-center weight from the mean of those weights, which is the learnable 'template' corresponding to a circular pattern under the floor-of-radius hash. Because the regularizer is differentiable, it plugs directly into end-to-end training and is what the paper credits for rotation invariance.

What would settle it

On affNIST with 10 images per class, replace Eq. 4 by (a) a hard constraint that all non-center weights in each 3x3 filter are equal, or (b) random 90-degree rotation augmentation at matched training budget. If neither reproduces the reported 85.06 percent, the soft circular-mean penalty itself, not the multi-scale maxout architecture, is the key to the low-data gain.

Watch

Extended reading notes

Core claim

The core claim is that affine invariance decomposes into two learnable priors: multi-scale maxout handles translation and scale, and a rotation-invariant regularizer handles rotation. The proposed regularizer, Eq. 4, is a least-squares penalty that pulls the eight non-center weights of each 3x3 filter toward their mean, making the filter approximate a discretized circle; the template is not fixed but re-estimated as that mean during training. On all three affine-transformed benchmarks the method outperforms ten compared networks under both full-data and 10-image-per-class training, with the reported average margins between 1.96 and 30.37 percentage points, and with learned filters ending up very close to circular patterns. On CIFAR-100 the same network also leads in the low-data regime, suggesting the prior helps beyond synthetic affine transforms.

Load-bearing premise

The load-bearing premise is that penalizing each 3x3 filter's non-center weights toward their mean is a faithful and non-destructive way to make filters circular, so that the rotation prior improves generalization instead of collapsing filters into uninformative flat masks.

Editorial extensions

If this is right

  • With full training data, the method reports 99.08 percent on affNIST, 98.92 percent on MNIST-rot, and 98.87 percent on Traffic Sign, each above the listed comparison networks.
  • With 10 random images per class, reported accuracies are 85.06, 87.49, and 84.15 percent on the same benchmarks, generally with standard deviation under one percent.
  • Adding data augmentation on Traffic Sign with 10 images per class improves the result by 3.69 percent, to 87.84 percent.
  • The rotation-invariant regularizer is nearly neutral with full data but contributes about 1.52 percent average improvement at 10 images per class, most visibly on affNIST.
  • The multi-scale depth choice is not sensitive: replacing three Conv+BN stages with two or four changes full-data accuracy by only about 0.2 percent.

Reading between the lines

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

  • A direct way to test whether the circular-mean penalty is the active ingredient is to add the same penalty to other architectures' convolutional filters; the paper only evaluates it inside its own multi-scale network.
  • Because the regularizer forces filters toward radial symmetry, it may reduce the network's ability to encode orientation itself, a potential cost for tasks where distinguishing rotated instances matters.
  • The same template-matching idea from Eq. 3 could be reused for other symmetries by choosing different hash functions, such as reflection or scale patterns, or by applying the penalty to larger filter sizes.
  • The reported gains could also be compared against matched compute with classic rotation augmentation, which would separate prior-driven data efficiency from augmentation-driven data efficiency.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes a two-part method for learning affine-invariant representations in CNNs: a multi-scale maxout architecture meant to handle translation and scale, and a rotation-invariant regularizer (Eq. 4) that pulls the outer weights of each 3x3 filter toward their mean, thereby supposedly approximating circular patterns. The network is trained end-to-end with a weighted combination of cross-entropy, weight decay, and this regularizer. Experiments on affNIST, MNIST-rot, Traffic Sign, and CIFAR-100 compare the method against nine baselines, reporting large gains in low-data regimes (e.g., 84.15% vs. 54.35% on Traffic Sign with 10 images per class).

Significance. If the empirical results are reproducible, the multi-scale maxout architecture is a credible contribution to low-data image classification under affine transformations. The paper reports three-trial averages with standard deviations for the low-shot experiments and tests on multiple benchmarks, which is a positive aspect. However, the proposed rotation-invariant regularizer is the main novelty in the title and conclusion, and its effect is both theoretically underdeveloped and empirically small. The significance of the work therefore rests on the architecture, not on the regularizer as claimed.

major comments (4)
  1. [Section 3.2.2, Eq. (4)] The formula for the rotation-invariant regularizer is mathematically inconsistent. With p_k = ceil(M_k/2) and q_k = ceil(N_k/2), a 3x3 filter yields p_k = q_k = 2, so the summation range m,n = -2,...,2 contains 25 indices, while the filter has only 3x3 = 9 weights. Moreover, the denominator p_k q_k - 1 equals 3 for a 3x3 filter, not the number of non-center weights (8). As written, the equation cannot be implemented, and this is the central definition of the proposed regularizer.
  2. [Section 3.2.2, Eq. (4) and Fig. 3] The regularizer minimizes the variance of the eight outer weights of each 3x3 filter, whose global minimum is a filter with all outer weights equal. This is not equivalent to 90-degree rotational symmetry on a square grid, where the four corners should be equal and the four edge-centers should be equal but not necessarily all eight equal. The paper provides no filter visualizations, no equivariance measure, and no formal argument that this penalty induces rotation invariance. The reported R2 value of 2.94e-7 is trivially small because it is the training objective; it is not an independent validation of circular patterns. Thus the claim that the regularizer handles rotation is unsupported.
  3. [Section 4.1.2, Fig. 7] The ablation of the regularizer shows only a 1.52% average improvement at 10 images per class and no improvement with full data. Since the reported low-data margins over the best baselines are much larger (e.g., 29.80% on Traffic Sign), the data-efficiency gains appear to be driven by the multi-scale maxout architecture rather than by the regularizer. The paper should isolate the regularizer's effect through controlled rotation experiments or filter analysis, or revise the claims to attribute the gains to the architecture.
  4. [Section 4.1.1] The empirical comparison uses public code implemented in four different deep learning frameworks (Chainer, Keras, TensorFlow, PyTorch), and the paper does not release its own code. The statement that each network was tuned to report its best performance is not verifiable without detailed hyperparameters and seeds. Full-data results are reported without error bars, and no statistical significance tests are given for the low-data comparisons. These factors limit reproducibility of the central benchmark claim.
minor comments (4)
  1. [Section 4.1.2] The sentence 'with a margin of 1.96% or 30.37%' is ambiguous; it is unclear which baselines and which dataset settings yield these two numbers, and they do not appear to match the margins in Table 1.
  2. [Figure 3] The caption and the figure would benefit from axis labels and a clearer explanation of the hash function h and the color coding of equal weights.
  3. [Section 3.2.2] The statement that the templates are 'updated automatically with the mean of the weights' is a re-description of variance reduction; the learnable-template framing adds little beyond a weighted variance penalty.
  4. [Section 1] The novelty claim 'we are the first to propose such regularization for handling rotation in deep learning' is strong; a brief discussion of existing filter-regularization and symmetry-enforcing methods would help place this work.

Circularity Check

2 steps flagged · score 2.0 of 10

Mild circularity: Eq. 4's template is defined as the filter's own mean, and the near-zero R2 is the minimized objective itself, not independent evidence of rotation invariance; the external benchmark claims are self-contained.

  1. self definitional [Section 3.2.2, Eq. (4) and the paragraph immediately after it]
    "Specifically we define our empirical rotation-invariant regularizer as follows: R2(ω,θ) = Ek∼K[Σ_{m,n≠0}(ωk(m,n) − θk(h(m,n)))^2], where θk(h(m,n)) = Σ_{m′,n′≠0}ωk(m′,n′)/(pkqk−1) is a scalar. Similar to the center loss...we aims to reduce the variance among the weights in each 2D spatial filter with 3×3 pixels, on average."

    The 'learnable template' θ that the regularizer is said to match is, by construction, the mean of the non-center weights of the very filter ω being regularized. Thus the claim that the network 'learns templates' is a re-description of variance reduction: minimizing R2 automatically draws the eight outer weights toward their own average, which is exactly the minimizer of the variance. No independent prior or external circular pattern is imposed; the only constraint is equality of the outer weights, put in by the definition θ = mean(ω).

  2. fitted input called prediction [Section 4.1.2, 'Effect of Rotation-Invariant Regularization' (Fig. 7 discussion)]
    "We also observe that our rotation-invariant regularizer can achieve very small numbers empirically. For instance, on affNIST the value is 2.94×10−7, indicating that our learned filters are very close to the spatial circular patterns."

    R2(ω,θ) is a term in the training objective (Eq. 2) with λ2 = 150, so SGD explicitly minimizes it. Reporting its final value of 2.94×10−7 shows only that the optimizer drove the training loss term to near zero, which is forced by construction. Using this value as evidence that the learned filters 'are very close to the spatial circular patterns' is circular, because a small R2 is the definition of the optimization target, not an independent measurement of rotation invariance or circularity.

full rationale

The paper's headline results are empirical accuracies on affNIST, MNIST-rot, Traffic Sign, and CIFAR-100 compared against external baselines. These comparisons are self-contained: the multi-scale maxout architecture and the regularizer are trained on public datasets and evaluated on held-out test sets, so the main data-efficiency claim does not reduce to an input or to a self-citation chain. The authors cite their own prior work ([44], [50]) only as related background, not as a load-bearing uniqueness or correctness argument. The only circular elements are internal to the presentation of the regularizer. In Eq. 4, the 'learnable template' θ is defined as the mean of the current non-center weights of the same filter, so the claim that the network learns templates is a re-description of variance reduction rather than matching against an independent prior. Similarly, the observation that R2 reaches 2.94×10−7 is used as evidence that filters approximate circular patterns, but R2 is the very term being minimized, so its small value is forced by the training objective rather than an independent check of rotation invariance. These issues do not invalidate the external benchmark comparisons; they affect only the interpretation of what the regularizer contributes, especially since the ablation in Fig. 7 shows only about 1.52% average gain at 10 images per class. Overall circularity is mild.

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

The central empirical claim rests on five assumptions: standard CNN translation equivariance, maxout over scales selecting the best scale, the specific circular-pattern hash being sufficient for rotation invariance, differentiability of the regularizer, and faithful tuned baselines. All except differentiability are heuristic or unverified in the manuscript. The only free parameters are ordinary training hyperparameters and architecture choices, plus the regularizer weight lambda2=150. No new physical or conceptual entities are introduced.

free parameters (5)
  • lambda2 (rotation regularizer weight) = 150
    Global strength of the circular-filter penalty in Eq. 2, set by hand and used for all datasets; Fig. 7 shows accuracy varies with it.
  • multi-scale maxout setting = 3x[Conv+BN] with effective filter sizes 3x3, 5x5, 7x7
    Architecture choice selected after comparing 2, 3, and 4 sequential Conv+BN blocks in Table 2.
  • training schedule and learning rates = 0.01 (full data) or 0.0001 (10-shot), decayed by 0.1 at 20k and 30k iterations; 42k iterations, batch size 100
    Standard hyperparameters set for all networks; baseline tuning may have differed.
  • network width and channels = [32,64,128,256,512] with FC 1024
    Default capacity chosen for the accuracy/efficiency trade-off; parameter study in Fig. 8.
  • lambda1 (weight decay) = 0.0005
    Standard L2 regularization, not the paper's main contribution.
assumptions (5)
  • domain assumption Convolution is translation equivariant, so translation invariance is handled by standard convolution and pooling.
    Central to the claim that only scale and rotation need special treatment; invoked in Section 1 and Fig. 1.
  • domain assumption Maxout over multi-scale receptive fields yields scale-invariant features.
    The paper assumes the max response across scales selects the best scale; no proof is provided. Section 3.1.
  • ad hoc to paper A 3x3 filter whose outer weights are all equal approximates a circular, rotation-invariant pattern in discrete images.
    Load-bearing for the regularizer in Eq. 4; the paper acknowledges this is an approximation in Section 3.2.1 and the Discussion.
  • standard math The regularizer R2 is differentiable, so SGD can train the network end-to-end.
    Required for backpropagation, stated in Section 3.2.2.
  • domain assumption Public implementations of baseline networks faithfully represent the original methods and were tuned to their best performance.
    Needed for the comparisons in Table 1 and Figs. 4-10; not verifiable from the manuscript.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Learning Affine-Invariant Representations via Data-Efficient CNNs." pith.science (2026). https://pith.science/paper/MPWYDU2U

@misc{pith2026190900114,
  author       = {Pith},
  title        = {Pith review of: Towards Learning Affine-Invariant Representations via Data-Efficient CNNs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MPWYDU2U}},
  note         = {Machine review of arXiv:1909.00114}
}
read the original abstract

In this paper we propose integrating a priori knowledge into both design and training of convolutional neural networks (CNNs) to learn object representations that are invariant to affine transformations (i.e., translation, scale, rotation). Accordingly we propose a novel multi-scale maxout CNN and train it end-to-end with a novel rotation-invariant regularizer. This regularizer aims to enforce the weights in each 2D spatial filter to approximate circular patterns. In this way, we manage to handle affine transformations in training using convolution, multi-scale maxout, and circular filters. Empirically we demonstrate that such knowledge can significantly improve the data-efficiency as well as generalization and robustness of learned models. For instance, on the Traffic Sign data set and trained with only 10 images per class, our method can achieve 84.15% that outperforms the state-of-the-art by 29.80% in terms of test accuracy.

Figures

Figures reproduced from arXiv: 1909.00114 by the authors.

Figure 1
Figure 1. To learn affine-invariant representations, we propose (a) a multi￾scale maxout convolutional network block to handle translation and scale, and (b) a regularizer to handle rotation. We use (a) for constructing our network, and embed (b) into our learning. to the transformations. For instance, TI-Pooling [26] assem￾bles all the transformed instances from the same data point in a pool and takes the maximal response fo… view at source ↗
Figure 3
Figure 3. Examples of weight patterns, defined by hash function h, that can be used to approximate circular patterns for rotation invariance. In each subfigure the same color denotes the same weight. encode such properties. Therefore, we propose the follow￾ing general formulation for rotation-invariant regularizers: R2(ω, θ) (3) = Ek∼K " Xpk m=−pk Xqk n=−qk d  ωk(m, n), θk(h(m, n)) # , where k ∈ K denotes the index of a 2D … view at source ↗
Figure 2
Figure 2. Illustration of the network we use in our experiments for learn￾ing affine-invariant features. Each dashed block is a multi-scale maxout block accounting for scale invariance, and the numbers here denote the default dimensions of inputs for the corresponding blocks and layers. object. In fact we use two and three 3 × 3 convolutions to approximate the responses with filter sizes of 5×5 and 7×7, respectively, for effi… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Test accuracy comparison of different networks on the three data sets. “Full” here indicates that we use all the training images. Our approach significantly outperforms the state-of-the-art, especially with small numbers of training images. 1 2 3 4 5 6 7 8 9 10 # Train…
Figure 5
Figure 5. Figure 5: Data augmentation comparison on Traffic Sign. 4.1.2 Results Better Generalization, Data-Efficiency, & Robustness. We summarize the test accuracy comparison in [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Illustration of training/testing behavior of different networks on affNIST. Full 10 [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Illustration of the effect of λ2 in Eq. 2 on test accuracy [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Test accuracy comparison with the others using different numbers of parameters. Best viewed in color. 10 4 10 5 10 6 10 7 10 8 # Parameters 10 -2 10 -1 10 0 10 1 Training Time Per Iteration (s) RotEqNet Harmonics TI-Pooling GCNN STN ResNet-32 CapsNet GoogLeNet DCN Ours…
Figure 9
Figure 9. Figure 9: Training/Test time comparison with the others using different numbers of parameters. Best viewed in color. 10 20 50 100 full # Training Images Per Class 0 10 20 30 40 50 60 70 80 Test Accuracy (%) RotEqNet Harmonics TI-Pooling ResNet-32 GoogleNet WRN-28-10 Ours [PITH_…
Figure 10
Figure 10. Figure 10: Test accuracy comparison of different networks on CIFAR￾100. “Full” here indicates that we use all the training images. Again our approach significantly outperforms the state-of-the-art, especially with small numbers of training images. can simply remove one Conv+BN i…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 43 canonical work pages

  1. [1]

    B. Amos, I. Jimenez, J. Sacks, B. Boots, and J. Z. Kolter. Differentiable mpc for end-to-end planning and control. In NIPS, pages 8299–8310, 2018. 2

    work page 2018

  2. [2]

    Andreas, M

    J. Andreas, M. Rohrbach, T. Darrell, and D. Klein. Neural module networks. In CVPR, pages 39–48, 2016. 1, 2

    work page 2016

  3. [3]

    Audebert, B

    N. Audebert, B. Le Saux, and S. Lefèvre. Semantic segmen- tation of earth observation data using multimodal and multi- scale deep networks. In ACCV, pages 180–196, 2016. 2

    work page 2016

  4. [4]

    Belkin, P

    M. Belkin, P. Niyogi, and V . Sindhwani. Manifold regular- ization: A geometric framework for learning from labeled and unlabeled examples. JMLR, 7(Nov):2399–2434, 2006. 4

    work page 2006

  5. [5]

    M. M. Bronstein, J. Bruna, Y . LeCun, A. Szlam, and P. Van- dergheynst. Geometric deep learning: going beyond eu- clidean data. IEEE Signal Processing Magazine, 34(4):18– 42, 2017. 4

    work page 2017

  6. [6]

    Cohen and M

    T. Cohen and M. Welling. Group equivariant convolutional networks. In ICML, pages 2990–2999, 2016. 2, 5

    work page 2016

  7. [7]

    Crandall, P

    D. Crandall, P. Felzenszwalb, and D. Huttenlocher. Spatial priors for part-based recognition using statistical models. In CVPR, volume 1, pages 10–17, 2005. 1

    work page 2005

  8. [8]

    J. Dai, H. Qi, Y . Xiong, Y . Li, G. Zhang, H. Hu, and Y . Wei. Deformable convolutional networks. In CVPR, pages 764– 773, 2017. 2, 5

    work page 2017

Show all 51 references
  1. [9]

    de Avila Belbute-Peres, K

    F. de Avila Belbute-Peres, K. Smith, K. Allen, J. Tenenbaum, and J. Z. Kolter. End-to-end differentiable physics for learn- ing and control. In NIPS, pages 7178–7189, 2018. 2

    2018

  2. [10]

    P. F. Felzenszwalb and D. P. Huttenlocher. Distance trans- forms of sampled functions. Theory Of Computing, 8:415– 428, 2012. 1

    2012

  3. [11]

    M. A. Fischler and R. A. Elschlager. The representation and matching of pictorial structures. IEEE Transactions on com- puters, 100(1):67–92, 1973. 1

    1973

  4. [12]

    Girshick, F

    R. Girshick, F. Iandola, T. Darrell, and J. Malik. Deformable part models are convolutional neural networks. In CVPR, pages 437–446, 2015. 1

    2015

  5. [13]

    I. J. Goodfellow, D. Warde-Farley, M. Mirza, A. Courville, and Y . Bengio. Maxout networks. InICML, pages III–1319,

  6. [14]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In CVPR, pages 770–778, 2016. 2, 5

    2016

  7. [15]

    G. Hinton. Taking inverse graphics seriously. https://www.cs.toronto.edu/ hinton/csc2535/notes/lec6b.pdf. 1

  8. [16]

    Hinton, N

    G. Hinton, N. Frosst, and S. Sabour. Matrix capsules with em routing. In ICLR, 2018. 1

    2018

  9. [17]

    G. E. Hinton, A. Krizhevsky, and S. D. Wang. Transform- ing auto-encoders. In International Conference on Artificial Neural Networks, pages 44–51. Springer, 2011. 1

    2011

  10. [18]

    Hoogeboom, J

    E. Hoogeboom, J. W. Peters, T. S. Cohen, and M. Welling. Hexaconv. arXiv preprint arXiv:1803.02108, 2018. 2

    2018 arXiv

  11. [19]

    Huang, D

    G. Huang, D. Chen, T. Li, F. Wu, L. van der Maaten, and K. Q. Weinberger. Multi-scale dense networks for resource efficient image classification. arXiv preprint arXiv:1703.09844, 2017. 2

    2017 arXiv

  12. [20]

    Ioffe and C

    S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, pages 448–456, 2015. 3

    2015

  13. [21]

    L. Isik, E. M. Meyers, J. Z. Leibo, and T. Poggio. The dy- namics of invariant object recognition in the human visual system. Journal of neurophysiology, 111(1):91–102, 2013. 1

    2013

  14. [22]

    Jaderberg, K

    M. Jaderberg, K. Simonyan, A. Zisserman, et al. Spatial transformer networks. In NIPS, pages 2017–2025, 2015. 2, 3, 5

    2017

  15. [23]

    X. Jia, B. De Brabandere, T. Tuytelaars, and L. V . Gool. Dy- namic filter networks. In NIPS, pages 667–675, 2016. 2

    2016

  16. [24]

    Kanazawa, A

    A. Kanazawa, A. Sharma, and D. Jacobs. Locally scale- invariant convolutional neural networks. arXiv preprint arXiv:1412.5104, 2014. 2

    2014 arXiv

  17. [25]

    Krizhevsky, V

    A. Krizhevsky, V . Nair, and G. Hinton. Cifar-100 (canadian institute for advanced research). 8

  18. [26]

    Laptev, N

    D. Laptev, N. Savinov, J. M. Buhmann, and M. Pollefeys. Ti- pooling: transformation-invariant pooling for feature learn- ing in convolutional neural networks. In CVPR, pages 289– 297, 2016. 2, 3, 5

    2016

  19. [27]

    Larochelle, D

    H. Larochelle, D. Erhan, A. Courville, J. Bergstra, and Y . Bengio. An empirical evaluation of deep architectures on problems with many factors of variation. In ICML, pages 473–480, 2007. 5

    2007

  20. [28]

    Larsson, M

    G. Larsson, M. Maire, and G. Shakhnarovich. Fractalnet: Ultra-deep neural networks without residuals. arXiv preprint arXiv:1605.07648, 2016. 2

    2016 arXiv

  21. [29]

    Y . LECUN. The mnist database of handwritten digits. http://yann.lecun.com/exdb/mnist/. 5

  22. [30]

    Liao and G

    Z. Liao and G. Carneiro. Competitive multi-scale convolu- tion. arXiv preprint arXiv:1511.05635, 2015. 2

    2015 arXiv

  23. [31]

    T.-Y . Lin, P. Dollár, R. Girshick, K. He, B. Hariharan, and S. Belongie. Feature pyramid networks for object detection. In CVPR, pages 2117–2125, 2017. 2

    2017

  24. [32]

    S. Luan, C. Chen, B. Zhang, J. Han, and J. Liu. Gabor con- volutional networks. TIP, 2018. 2

    2018

  25. [33]

    Marcos, M

    D. Marcos, M. V olpi, N. Komodakis, and D. Tuia. Rota- tion equivariant vector field networks. InICCV, pages 5048– 5057, 2017. 2, 5

    2017

  26. [34]

    Sabour, N

    S. Sabour, N. Frosst, and G. E. Hinton. Dynamic routing between capsules. In NIPS, pages 3859–3869, 2017. 1, 5

    2017

  27. [35]

    Snell, K

    J. Snell, K. Swersky, and R. Zemel. Prototypical networks for few-shot learning. In NIPS, pages 4077–4087, 2017. 3

    2017

  28. [36]

    Stallkamp, M

    J. Stallkamp, M. Schlipsing, J. Salmen, and C. Igel. The German Traffic Sign Recognition Benchmark: A multi-class classification competition. In IJCNN, pages 1453–1460,

  29. [37]

    Szegedy, W

    C. Szegedy, W. Liu, Y . Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V . Vanhoucke, and A. Rabinovich. Going deeper with convolutions. In CVPR, pages 1–9, 2015. 2, 5

    2015

  30. [38]

    Takahashi, T

    R. Takahashi, T. Matsubara, and K. Uehara. Scale-invariant recognition by weight-shared cnns in parallel. In ACML, pages 295–310, 2017. 2

    2017

  31. [39]

    J. Wang, Z. Wei, T. Zhang, and W. Zeng. Deeply-fused nets. arXiv preprint arXiv:1605.07716, 2016. 2

    2016 arXiv

  32. [40]

    Weiler, F

    M. Weiler, F. A. Hamprecht, and M. Storath. Learning steer- able filters for rotation equivariant cnns. In CVPR, pages 849–858, 2018. 2

    2018

  33. [41]

    Y . Wen, K. Zhang, Z. Li, and Y . Qiao. A discriminative fea- ture learning approach for deep face recognition. In ECCV, pages 499–515, 2016. 3, 5

    2016

  34. [42]

    D. E. Worrall, S. J. Garbin, D. Turmukhambetov, and G. J. Brostow. Harmonic networks: Deep translation and rotation equivariance. In CVPR, volume 2, 2017. 2, 5, 6

    2017

  35. [43]

    J. Wu, D. Li, Y . Yang, C. Bajaj, and X. Ji. Dynamic filtering with large sampling field for convnets. In ECCV, pages 185– 200, 2018. 2

    2018

  36. [44]

    Y . Xu, T. Xiao, J. Zhang, K. Yang, and Z. Zhang. Scale- invariant convolutional neural networks. arXiv preprint arXiv:1411.6369, 2014. 2

    2014 arXiv

  37. [45]

    Yu and V

    F. Yu and V . Koltun. Multi-scale context aggregation by di- lated convolutions. arXiv preprint arXiv:1511.07122, 2015. 2

    2015 arXiv

  38. [46]

    Zagoruyko and N

    S. Zagoruyko and N. Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 4

    2016 arXiv

  39. [47]

    Zhang, Y

    Q. Zhang, Y . Nian Wu, and S.-C. Zhu. Interpretable convo- lutional neural networks. In CVPR, pages 8827–8836, 2018. 2

    2018

  40. [48]

    Zhang, G.-J

    T. Zhang, G.-J. Qi, B. Xiao, and J. Wang. Interleaved group convolutions. In CVPR, 2017. 2

    2017

  41. [49]

    Zhang, L

    X. Zhang, L. Liu, Y . Xie, J. Chen, L. Wu, and M. Pietikäinen. Rotation invariant local binary convolution neural networks. In ICCV Workshops, pages 1210–1219, 2017. 2

    2017

  42. [50]

    Zhang and M

    Z. Zhang and M. Brand. Convergent block coordinate de- scent for training tikhonov regularized deep neural networks. In NIPS, pages 1721–1730. 2017. 1

    2017

  43. [51]

    Y . Zhou, Q. Ye, Q. Qiu, and J. Jiao. Oriented response net- works. In CVPR, pages 4961–4970, 2017. 2

    2017

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.