Pith. sign in

REVIEW 3 major objections 5 minor 32 references

Through a Steerable Lens: Magnifying Neural Network Interpretability via Phase-Based Extrapolation

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that amplifying a class-conditional gradient in the complex steerable pyramid domain, mainly along phase, produces coherent class-to-class morphs that expose the network's decision geometry.

desk verdict A plausible new method for inter-class morphs via CSP phase gradient extrapolation, but the central claim about decision geometry is unvalidated without quantitative checks. read the letter →

arxiv 2506.02300 v3 pith:MQL2JHBX submitted 2025-06-02 cs.LG

Farzaneh Mahdisoltani , Saeed Mahdisoltani , Roger B. Grosse , David J. Fleet This is my paper
classification cs.LG
keywords neuralnetworkinterpretabilitygradientextrapolationcomplexsteerablepyramidphase-basedmotionmagnificationcounterfactualvisualizationdecisionboundarygeometryWirtingercalculus
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 proposes a visualization method for neural network classifiers: instead of highlighting which pixels matter, it generates a sequence of images that morphs an input from its current class toward a target class. The method treats the network gradient as an infinitesimal push and linearly extrapolates that push over a long distance, but performs the extrapolation in the complex steerable pyramid domain rather than in raw pixel space. Because the pyramid separates local amplitude (feature strength) from local phase (feature position and form), moving primarily along the phase direction produces spatially coherent structural edits, such as a '3' closing into an '8' or a sad mouth turning up. The intended payoff is a dynamic, class-to-class view of the geometry the network actually uses to separate classes, something static saliency maps do not provide.

What carries the argument

The central object is the complex steerable pyramid, a multiscale, orientation-selective image decomposition that separates local amplitude from local phase: $S(\omega)=A(\omega)e^{j\Phi(\omega)}$. The identity that carries the argument is the Wirtinger-chain-rule relation between the complex gradient $\nabla_{S^*}(\omega)L$ and the amplitude/phase gradients, which lets the method translate one gradient step into a phase shift. The phase-focused extrapolation rule $\Phi_k(\omega)=\Phi_0(\omega)+k\Delta\Phi(\omega)$ with $A_k=A_0$ then produces the morphing sequence by reconstruction through the inverse pyramid transform.

What would settle it

Pick a trained classifier and a correctly classified source image, generate the linear phase-extrapolation trajectory toward a target class, and compare it to a multi-step integrated gradient path in the same CSP space where the gradient is recomputed at every step. If the target-class probability along the linear extrapolation plateaus or drops before the target is reached while the integrated path keeps increasing, the fixed-direction assumption is refuted.

Watch

Extended reading notes

Core claim

The central discovery is that a single local gradient step in the complex steerable pyramid (CSP) domain, extrapolated linearly and mostly along phase, yields semantically meaningful and spatially coherent morphing trajectories between classes. Formally, each CSP coefficient is written $S(\omega)=A(\omega)e^{j\Phi(\omega)}$, the Wirtinger gradient with respect to $S^*(\omega)$ is converted into amplitude and phase contributions, and the phase shift $\Delta\Phi(\omega)=\arg\left(1+\frac{\alpha}{2}\frac{\nabla_{A_0}L}{A_0}+j\frac{\alpha}{2}\frac{\nabla_{\Phi_0}L}{A_0^2}\right)$ is applied linearly for $k$ steps while amplitude is held constant. If the paper is right, the transition a classifier implicitly perceives between two classes is visible as a short, coherent deformation sequence, and the directions of most sensitive change in the network's feature space are exposed.

Load-bearing premise

The load-bearing premise is that the single gradient direction computed at the starting image can be stretched linearly over a large distance and still trace a path the classifier would actually take between classes.

Editorial extensions

If this is right

  • A network's class-to-class transition, as seen through phase extrapolation, is localized and often matches human intuition about digit shapes: a 3 closes into an 8 and a 5 grows a loop into a 6.
  • Because the amplitude is held fixed, the morphs isolate structural and positional changes from contrast changes, showing which shape cues the classifier relies on.
  • On color MNIST, the method reveals whether the learned representation encodes color at all: color channels stay unchanged when color is irrelevant, and morphing is steered within the active channel when color defines the class.
  • On FER2013 facial expressions, the trajectories track expression-specific cues such as mouth curvature and smile lines, suggesting that phase directions align with perceptually meaningful expression dimensions.
  • The procedure requires no generative model and no iterative optimization: one autograd pass supplies the direction, and linear extrapolation produces the sequence.

Reading between the lines

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

  • Beyond the paper, the method's ability to produce coherent morphs in MNIST suggests that CSP phase coordinates act as a partially linearizing chart for convolutional classifiers; if true, one could compare phase-space linear paths across architectures as a diagnostic.
  • Beyond the paper, the Arcade experiment's centroid-following behavior implies these visualizations can expose what a network ignores (here, shape fidelity), not only what it uses; that could be developed into an invariance-discovery tool.
  • Beyond the paper, a quantitative faithfulness metric—such as comparing linear phase extrapolation to a recomputed multi-step gradient path in the same transform space—would let practitioners know when the visualization is trustworthy; the paper leaves such metrics to future work.
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

3 major / 5 minor

Summary. The paper proposes a new interpretability method that visualizes inter-class transitions by extrapolating classifier gradients in the complex steerable pyramid (CSP) domain, focusing on phase components. The method first computes a Wirtinger gradient of the classification loss with respect to complex CSP coefficients, derives a phase update rule, and then linearly extrapolates the phase over several steps while keeping amplitude constant. Experiments on a synthetic Arcade dataset, MNIST (including color variants), and FER2013 show morphing sequences that the authors argue are semantically meaningful and reveal the model's decision boundary geometry. The appendices provide a self-contained derivation of the Wirtinger update and show Fourier-domain comparisons.

Significance. If validated, the paper offers a novel and appealing interpretability lens: a single local gradient step in a structured, invertible transform space, extrapolated linearly, could expose perceptually coherent transformations that highlight what a classifier attends to. The paper's strengths include a clean and internally consistent mathematical derivation in Appendix B, a clear connection to phase-based motion magnification, and a synthetic dataset that allows controlled inspection. However, the central interpretive claim—that these morphs reveal decision-boundary geometry—currently rests almost entirely on qualitative figure inspection. There are no quantitative fidelity metrics, no comparisons to control paths, and no ablation isolating the effect of phase-only versus amplitude updates. The paper itself acknowledges the missing quantitative evaluation in Section 5, which makes the central claim suggestive rather than established.

major comments (3)
  1. [§4.1, §4.2, Eq. (10)–(11)] The central claim that the generated sequences 'reveal the geometry of decision boundaries' is not supported by quantitative evidence. The Arcade experiments (Figure 2) show that the disk moves in the target direction, but any method that shifts the centroid toward the target class would produce the same qualitative result; the experiments do not distinguish a gradient-faithful trajectory from a generic translation. Similarly, the MNIST and FER2013 results are presented as figures without measuring, for example, the target-class logit or probability along the sequence, the distance to the decision boundary, or the gradient alignment at successive steps. The paper should include such quantitative path-fidelity metrics to substantiate the interpretability claim, or explicitly reframe the contribution as a heuristic visualization whose faithfulness is not yet demonstrated.
  2. [§5, 'Conclusion and Future Directions'] The limitation that 'a fixed gradient direction provides a meaningful path over a finite extrapolation distance' is load-bearing because Eqs. (10)–(11) define the morph as a straight-line extrapolation of a single local tangent in phase coordinates. For a nonlinear classifier, there is no guarantee that this line stays near the decision-relevant manifold or that target-class probability increases monotonically along it. The paper should provide at least one concrete test of this assumption, such as plotting target-class logits along the generated sequence, measuring the angle between the initial gradient and the gradient at intermediate points, or comparing against an integrated-gradient path. Without such evidence, the visualization may be perceptually coherent yet geometrically unfaithful.
  3. [§4.2 and Appendix C] The claimed advantages of CSP over pixel-space gradient ascent and Fourier-domain extrapolation are supported only by qualitative statements ('often adds noise', 'tends to produce more global shifts or blurring'). To make the comparison credible, the same source images should be processed with each baseline under matched step sizes and the resulting sequences should be evaluated with quantitative criteria, such as target-class confidence increase, spatial coherence of the edits, or perceptual similarity to the target class. The current comparisons do not control for the amount of total perturbation or the number of steps, so it is unclear whether the observed differences reflect the transform choice or simply different operating points.
minor comments (5)
  1. [Abstract] The abstract contains a spacing typo: 'elucidatehowa' should read 'elucidate how a'.
  2. [§3.3, Eq. (6) and Appendix A] The step-size notation is inconsistent: Appendix A derives a gradient-ascent update with a factor of 2 (Sk+1 = S_k + 2α ∂L/∂S*), while Eq. (6) absorbs the 1/2 into a redefined α. Please state explicitly that the α in the main text is an effective step size, or align the notation between the derivation and the update rule.
  3. [§4.3] The color-MNIST section introduces per-channel amplifier settings and YIQ transformations but does not specify the exact values of α used, nor how the per-channel settings were selected. A short table or sentence describing the hyperparameters would improve reproducibility.
  4. [§4.1] The Arcade dataset description does not specify the exact distribution of the random displacement (e.g., uniform over what range), which makes it difficult to assess how discriminative the task is. Please include the precise generation procedure.
  5. [References] The claim in §5 that 'prior work has shown that subtle phase modifications can induce significant, spatially coherent changes' is supported by references [14,21], but the sentence in §3.3 also makes this point without a citation; please add the citation there for consistency.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the phase-extrapolation morphs are directly constructed from the Wirtinger gradient, with no fitted prediction or self-citation chain doing the work.

full rationale

The derivation chain is self-contained. Section 3 constructs the morph sequence by definition: Eq. (2) defines S_k = S_0 + k*alpha*g, and Eqs. (10)-(11) define the phase-only extrapolation Phi_k = Phi_0 + k*DeltaPhi with A_k = A_0. DeltaPhi itself is derived in Appendix B from the Wirtinger gradient via the chain rule, so the output morphs are the algorithm's construction rather than a prediction of an independently measured quantity. The amplifier alpha is a hand-chosen step size, not a parameter fitted to make a subsequent claim true; the color-MNIST per-channel amplifiers are described as requiring dataset-specific tuning for visual quality, and that tuning is explicitly presented as a limitation, not as evidence for a prediction. The self-citations involving David Fleet ([27,28,29]) are background references for phase-sensitive processing and are not load-bearing: no uniqueness theorem or forbidden-alternative argument rests on them. The interpretive claim that the morphs reveal decision-boundary geometry is not derived by construction; the paper explicitly labels the fixed-gradient direction as an assumption and calls for quantitative faithfulness metrics in Section 5. Any concern about whether the linear extrapolation faithfully tracks the true decision boundary is a correctness or validation risk, not a circularity. Hence no circular step is present.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical or mathematical entities. It relies on established transforms and standard gradient computation. The main assumptions are domain assumptions about phase semantics and linearity of the gradient path, plus a set of hand-chosen display parameters (step size, number of steps, CSP configuration).

free parameters (3)
  • alpha (amplifier/step size) = not specified, varies per dataset and per channel
    Determines the magnitude of phase extrapolation per step and the number of frames. The paper notes that per-channel amplifiers require careful tuning in the composite color MNIST case. Since the morphs are the output, the choice of alpha shapes the visualization but is not fitted to a quantitative target.
  • CSP hyperparameters = 4 orientations, 3 scales
    The transform configuration is fixed by the authors and affects the localization and orientation selectivity of the morphs. It is a design choice rather than a fitted parameter.
  • number of extrapolation steps = not specified
    The number of frames k in the morphing sequence is chosen by hand for display. This choice affects how far the extrapolation goes and thus the appearance of the trajectory.
assumptions (4)
  • domain assumption Phase in Fourier and steerable pyramid domains carries most of the perceptually relevant structural information, and local phase shifts correspond to coherent spatial deformations.
    This is the foundational assumption of the method, invoked in Sections 3.2 and 3.3. It is supported by the cited signal processing literature, but its validity for the specific morphs shown is not independently verified.
  • domain assumption A single gradient step in CSP coefficient space, extrapolated linearly, stays within a region where the network's decision geometry is approximately linear.
    The paper explicitly acknowledges this assumption in Section 5 as a limitation. No experiment tests whether the straight line in coefficient space corresponds to a plausible path in the network's feature space.
  • domain assumption The steerable pyramid is a sufficiently disentangled basis that its phase directions align with semantically meaningful visual attributes for the tested classifiers.
    The paper argues that CSP provides perceptual feature disentanglement. The visual results are consistent with this, but no quantitative measure of disentanglement or semantic alignment is provided.
  • domain assumption The classifiers used are standard and their training details (architecture, optimization, hyperparameters) are sufficient to produce representative decision boundaries.
    The paper describes the architectures qualitatively but does not provide hyperparameters, training curves, or accuracies. The visualizations depend on the particular trained models, so the lack of model details limits reproducibility.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Through a Steerable Lens: Magnifying Neural Network Interpretability via Phase-Based Extrapolation." pith.science (2026). https://pith.science/paper/MQL2JHBX

@misc{pith2026250602300,
  author       = {Pith},
  title        = {Pith review of: Through a Steerable Lens: Magnifying Neural Network Interpretability via Phase-Based Extrapolation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MQL2JHBX}},
  note         = {Machine review of arXiv:2506.02300}
}
read the original abstract

Understanding the internal representations and decision mechanisms of deep neural networks remains a critical open challenge. While existing interpretability methods often identify influential input regions, they may not elucidate how a model distinguishes between classes or what specific changes would transition an input from one category to another. To address these limitations, we propose a novel framework that visualizes the implicit path between classes by treating the network gradient as a form of infinitesimal motion. Drawing inspiration from phase-based motion magnification, we first decompose images using invertible transforms-specifically the Complex Steerable Pyramid-then compute class-conditional gradients in the transformed space. Rather than iteratively integrating the gradient to trace a full path, we amplify the one-step gradient to the input and perform a linear extrapolation to expose how the model moves from source to target class. By operating in the steerable pyramid domain, these amplified gradients produce semantically meaningful, spatially coherent morphs that highlight the classifier's most sensitive directions, giving insight into the geometry of its decision boundaries. Experiments on both synthetic and real-world datasets demonstrate that our phase-focused extrapolation yields perceptually aligned, semantically meaningful transformations, offering a novel, interpretable lens into neural classifiers' internal representations.

Figures

Figures reproduced from arXiv: 2506.02300 by the authors.

Figure 1
Figure 1. Block diagram of the phase-based gradient extrapolation pipeline. An input image [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Morphing sequences of the Arcade dataset using CSP extrapolation. Each row visualizes [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Transformation trajectories on MNIST: Each row shows a source digit (left frame) morphing [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Joint color-shape morph in the YIQ representation. Target confidence increases throughout [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Transformation trajectories on MNIST with [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Transformation trajectories on MNIST with [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Morphing of FER2013 samples. Images show transitions between emotions by extrapolating [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Examples of Arcade sample transformation using Fourier Transform. [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Examples of MNIST transformations using Fourier Transform phase extrapolation. Com [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 20 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. InAdvances in neural information processing systems (NeurIPS), volume 25, pages 1097–1105, 2012. 9

    work page 2012

  2. [2]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 770–778, 2016

    2016

  3. [3]

    Towards a rigorous science of interpretable machine learning

    Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608, 2017

    arXiv 2017

  4. [4]

    Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead.Nature Machine Intelligence, 1(5):206–215, 2019

    Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead.Nature Machine Intelligence, 1(5):206–215, 2019

    2019

  5. [5]

    Deep inside convolutional networks: Visualising image classification models and saliency maps.arXiv preprint arXiv:1312.6034, 2013

    Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps.arXiv preprint arXiv:1312.6034, 2013

    arXiv 2013

  6. [6]

    Striving for simplicity: The all convolutional net

    Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for simplicity: The all convolutional net. InInternational Conference on Learning Representations (ICLR) Workshop Track, 2015

    work page 2015

  7. [7]

    Smooth- grad: removing noise by adding noise, 2017

    Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. Smooth- grad: removing noise by adding noise, 2017

    work page 2017

  8. [8]

    Axiomatic attribution for deep networks

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In International Conference on Machine Learning (ICML), pages 3319–3328. PMLR, 2017

    2017

Show all 32 references
  1. [9]

    Grad-cam: Visual explanations from deep networks via gradient-based localization

    Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. InProceedings of the IEEE International Conference on Computer Vision (ICCV), pages 61...

    2017

  2. [10]

    Intriguing properties of neural networks

    Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Good- fellow, and Rob Fergus. Intriguing properties of neural networks. InInternational Conference on Learning Representations (ICLR), 2014

    2014

  3. [11]

    Explaining and harnessing adversar- ial examples.arXiv preprint arXiv:1412.6572, 2014

    Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversar- ial examples.arXiv preprint arXiv:1412.6572, 2014

    2014 arXiv

  4. [12]

    Counterfactual explanations without open- ing the black box: Automated decisions and the gdpr.Harvard Journal of Law & Technology, 31(2):841–887, 2018

    Sandra Wachter, Brent Mittelstadt, and Chris Russell. Counterfactual explanations without open- ing the black box: Automated decisions and the gdpr.Harvard Journal of Law & Technology, 31(2):841–887, 2018

    2018

  5. [13]

    Counterfactual visual explanations

    Yash Goyal, Ziyan Wu, Jonas Ernst, Dhruv Batra, Devi Parikh, and Stefan Lee. Counterfactual visual explanations. InInternational Conference on Machine Learning (ICML), pages 2376–

  6. [14]

    Hao-Yu Wu, Michael Rubinstein, Eugene Shih, John Guttag, Frédo Durand, and William T. Freeman. Eulerian video magnification for revealing subtle changes in the world. InACM Transactions on Graphics (TOG), volume 31, pages 1–8. ACM, 2012

    2012

  7. [15]

    Freeman and Edward H

    William T. Freeman and Edward H. Adelson. The design and use of steerable filters.IEEE Transactions on Pattern Analysis and Machine Intelligence, 13(9):891–906, 1991

    1991

  8. [16]

    Oppenheim and Ronald W

    Alan V . Oppenheim and Ronald W. Schafer.Discrete-Time Signal Processing. Prentice Hall, 2nd edition, 1999

    1999

  9. [17]

    The steerable pyramid: A flexible architecture for multi-scale derivative computation

    Eero P Simoncelli and William T Freeman. The steerable pyramid: A flexible architecture for multi-scale derivative computation. InProceedings of 2nd IEEE International Conference on Image Processing, volume 3, pages 444–447. IEEE, 1995

    1995

  10. [18]

    Academic Press, 1999

    Stéphane Mallat.A Wavelet Tour of Signal Processing. Academic Press, 1999

    1999

  11. [19]

    A parametric texture model based on joint statistics of complex wavelet coefficients.International Journal of Computer Vision, 40(1):49–70, 2000

    Javier Portilla and Eero P Simoncelli. A parametric texture model based on joint statistics of complex wavelet coefficients.International Journal of Computer Vision, 40(1):49–70, 2000. 10

    2000

  12. [20]

    Phase-based video motion processing

    Neal Wadhwa, Michael Rubinstein, Frédo Durand, and William T Freeman. Phase-based video motion processing. InACM Transactions on Graphics (TOG), volume 32, pages 1–10. ACM, 2013

    2013

  13. [21]

    Neal Wadhwa, Michael Rubinstein, Frédo Durand, and William T. Freeman. Phase-based motion magnification. InCommunications of the ACM, volume 60, pages 87–95, 2017

    2017

  14. [22]

    Feature visualization.Distill, 2(11):e7, 2017

    Chris Olah, Alexander Mordvintsev, and Ludwig Schubert. Feature visualization.Distill, 2(11):e7, 2017

    2017

  15. [23]

    Network dissec- tion: Quantifying interpretability of deep visual representations

    David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. Network dissec- tion: Quantifying interpretability of deep visual representations. InProceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 6541–6549, 2017

    2017

  16. [24]

    Interpretation of neural networks is fragile

    Amirata Ghorbani, Abubakar Abid, and James Zou. Interpretation of neural networks is fragile. InProceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 3681–3688, 2019

    2019

  17. [25]

    C. D. Kuglin and D. C. Hines. The phase correlation image alignment method. InProceedings of the IEEE International Conference on Cybernetics and Society, pages 163–165, 1975

    1975

  18. [26]

    Oppenheim and Jae S

    Alan V . Oppenheim and Jae S. Lim. The importance of phase in signals.Proceedings of the IEEE, 69(5):529–541, may 1981

    1981

  19. [27]

    Computation of component image velocity from local phase information.International Journal of Computer Vision, 5(1):77–104, 1990

    David J Fleet and Allan D Jepson. Computation of component image velocity from local phase information.International Journal of Computer Vision, 5(1):77–104, 1990

    1990

  20. [28]

    J. L. Barron, D. J. Fleet, S. S. Beauchemin, and T. A. Burkitt. Performance of optical flow techniques. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 236–242, Champaign, IL, June 1992

    1992

  21. [29]

    Fleet, Allan D

    David J. Fleet, Allan D. Jepson, and Michael R. M. Jenkin. Phase-based disparity measurement. CVGIP: Image Understanding, 53(2):198–210, February 1991

    1991

  22. [30]

    Brandwood

    David H. Brandwood. A complex gradient operator and its application in adaptive array theory. IEE Proceedings H (Microwaves, Optics and Antennas), 130(1):11–16, 1983

    1983

  23. [31]

    Gradient-based learning applied to document recognition.Proceedings of the IEEE, 86(11):2278–2324, 1998

    Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition.Proceedings of the IEEE, 86(11):2278–2324, 1998

    1998

  24. [32]

    Ian J. Goodfellow, Dumitru Erhan, Pierre Luc Carrier, Aaron Courville, Mehdi Mirza, Ben Hamner, Will Cukierski, Yichuan Tang, David Thaler, Dong-Hyun Lee, Yingbo Zhou, Chetan Ramaiah, Fangxiang Feng, Ruifan Li, Xiaojie Wang, Dimitris Athanasakis, John Shawe- Taylor, Maxim Mila...

    2013 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.