Pith. sign in

REVIEW 2 major objections 5 minor 53 references

CoFrNets: Interpretable Neural Architecture Inspired by Continued Fractions

T0 review · 2 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 Continued-fraction ladders can approximate any continuous function on a bounded domain, and their ladder form makes the model directly interpretable.

desk verdict Genuinely new architecture and a clever proof route, but the universality theorem is proved for a different function class than the canonical CoFrNet that is actually trained, so the headline claim needs a fix before it lands. read the letter →

arxiv 2506.05586 v1 pith:QXCVVVWN submitted 2025-06-05 cs.LG cs.AI

Isha Puri , Amit Dhurandhar , Tejaswini Pedapati , Kartikeyan Shanmugam , Dennis Wei , Kush R. Varshney This is my paper
classification cs.LGcs.AI MSC 41A2041A6568T07
keywords continuedfractionsinterpretableneuralarchitectureuniversalapproximationStone-Weierstrasstheoremfeatureattributionsrationalfunctionssparsepolynomiallearning
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

Continued fractions form a ladder of reciprocals: a0 plus one over a1 plus one over a2, and so on. The paper builds a neural network, CoFrNet, whose layers are exactly such ladders with linear functions ak = wk^T x on each rung, and trains the weights with backpropagation. The main theoretical claim is that finite linear combinations of finite-depth continued fractions with linear rungs are dense in the space of bounded continuous functions on [0,1]^p, making the class a universal approximator; the proof goes through Euler's continued-fraction identity rather than the usual width/depth arguments. The same functional form yields interpretability: gradients can be computed in closed form through continuants, and a power-series expansion gives feature and interaction attributions. Experiments on synthetic functions and seven datasets across tabular, text, and image modalities show accuracy comparable to or better than other interpretable models and often multilayer perceptrons.

What carries the argument

The central object is the continued-fraction ladder $f(x;w)=a_0+\frac{1}{a_1+\frac{1}{a_2+\cdots+\frac{1}{a_d}}}$ with $a_k=w_k^T x$, equivalently the ratio of continuants $K_{d+1}(a_0,\ldots,a_d)/K_d(a_1,\ldots,a_d)$, where continuants obey $K_0=1$, $K_1(a_d)=a_d$, and $K_k(a_{d-k+1},\ldots,a_d)=a_{d-k+1}K_{k-1}(a_{d-k+2},\ldots,a_d)+K_{k-2}(a_{d-k+3},\ldots,a_d)$. The load-bearing identity is Euler's continued-fraction formula, Eq. (9): $\frac{a_0}{1+\frac{-a_1}{1+a_1+\frac{-a_2}{1+a_2+\cdots+\frac{-a_d}{1+a_d}}}}=a_0+a_0a_1+\cdots+a_0a_1\cdots a_d$. Applied twice, it rewrites a single monomial as a difference of two ladders, which is what embeds polynomials of linear functions into the linear span of CFL. The reciprocal $1/z$ is the nonlinearity, replaced in practice by $\mathrm{sgn}(z)/\max(|z|,\epsilon)$ to avoid poles, and the continuant recursion gives an O(d) computation of the ladder and its gradient.

What would settle it

For p=2 take g(x)=x1 x2. The CFL representation built in Theorem 2 uses Euler-form ladders with denominators like 1+a1 and a numerator a0, while the canonical Eq. (2) ladders have unit numerators and linear denominators; determine whether any finite linear combination of canonical ladders can approximate x1 x2 to machine precision as depth grows. If a systematic approximation floor persists, the universality guarantee does not transfer to the trained architecture; similarly, comparing the ideal 1/z versus the clipped activation on the same target would test whether the epsilon modification is harmless.

Watch

Extended reading notes

Core claim

On its own terms, the central discovery is Theorem 2: the set CFL of finite-depth continued fractions whose rungs are affine functions of the input, together with all finite linear combinations, is dense in C([0,1]^p,R) with respect to the sup norm. The proof has three steps: polynomials of linear functions form a separating unital subalgebra, so Stone-Weierstrass makes them dense; Euler's continued-fraction formula expresses each product a0 a1 ... ad as the difference of two continued fractions, so every monomial, hence every polynomial of linear functions, lies in a finite linear combination of CFL; therefore CFL is dense as well. The paper further derives Proposition 1, a closed-form gradient of a ladder in terms of continuants, and a power-series correspondence that converts ladders into coefficient attributions for individual features and their interactions. The theorem is stated for CFL, whose rungs are affine; the architecture trained in the experiments is the canonical unit-numerator ladder of Eq. (2).

Load-bearing premise

The load-bearing premise is that the class CFL for which Theorem 2 proves universality—continued fractions with affine numerators and denominators in Euler form—is effectively the same as the canonical unit-numerator CoFrNet ladders trained in the experiments, with the clipped reciprocal activation standing in for the ideal 1/z, and the paper gives no equivalence or error-bound argument for either step.

Editorial extensions

If this is right

  • Any continuous function on a compact cube can be approximated to arbitrary sup-norm accuracy by a finite linear combination of finite-depth continued fractions with affine rungs.
  • A sparse polynomial with at most d nonzero monomials can be represented with O(d) ladders of depth at most d, in contrast to the p^d-sized dictionary a Lasso-style method would need to search.
  • First-order per-example attributions are available in O(Ldp) time via the continuant gradient formula, and global higher-order attributions are available by extracting power-series coefficients from the trained ladders.
  • CoFrNet-D is an additive model whose per-feature ladders can be plotted directly, while CoFrNet-DL adds full ladders of increasing depth to capture interaction order.
  • On the six real datasets with comparisons, CoFrNet-DL matches or exceeds the tested interpretable models and MLPs, and on ImageNet its accuracy is reported as comparable to ResNet-18-type models.

Reading between the lines

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

  • Editorial inference: if the density theorem extends to the actually trained canonical ladders, the Euler-form embedding suggests a constructive way to initialize or compose ladders for target monomials, potentially turning the architecture into an adaptive rational basis for regression and symbolic regression.
  • Editorial inference: the compact representation of sparse polynomials suggests a testable learning-theoretic prediction: a sparse-penalty variant of CoFrNet should recover sparse polynomial structure with sample complexity depending on the number of nonzero terms rather than p^d, a direction the paper leaves open.
  • Editorial inference: since the proof covers CFL with affine numerators and denominators while the experiments use unit numerators and clipped reciprocals, the universality guarantee for the deployed architecture remains open; closing this gap, or proving an approximation bound in epsilon, would complete the story.
  • Editorial inference: the closed-form continuant gradient might be used not only for attributions but for adversarial robustness analysis, since Lipschitz constants of low-depth rational ladders can be computed analytically; the paper only mentions this as 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

2 major / 5 minor

Summary. The paper proposes CoFrNet, an interpretable neural architecture inspired by continued fractions, with three variants (F, D, DL). The main theoretical contribution is a universal approximation theorem (Theorem 2) for a class of continued fractions with linear layers, proved via Stone-Weierstrass and Euler's continued fraction formula. The paper also derives a gradient formula for interpretability (Proposition 1) and presents experiments on synthetic functions and seven real datasets, showing competitive accuracy with interpretable baselines.

Significance. If the universality theorem applied to the architecture in Eq. (2), the paper would make a valuable contribution: a new proof technique for universal approximation, an interpretable architecture, and a principled way to extract feature attributions. The proof of Proposition 1 and the empirical interpretability results are interesting in their own right. However, the central theoretical claim is not established for the architecture that is actually defined and trained.

major comments (2)
  1. [Section 5, Theorem 2 and Definition 5 vs. Section 4, Eq. (2)] Theorem 2 proves density for the class CFL of Definition 5, whose elements have affine partial numerators v_k^T x + alpha_k and denominators 1 + w_k^T x + beta_k. The CoFrNet ladder in Eq. (2) is the canonical continued fraction a_0 + 1/(a_1 + ...) with unit partial numerators and linear denominators a_k = w_k^T x. The paper offers no inclusion argument showing that the canonical ladders are in CFL or in its linear span, nor that they can approximate the CFL elements used in the proof. Equivalence transformations discussed in Section 3 do not bridge the gap: converting a non-unit partial numerator to 1 introduces a rational function as the new denominator, which the architecture cannot represent with a single linear layer. Consequently, Theorem 2 does not establish the abstract's claim that CoFrNets are universal approximators.
  2. [Section 4, 'Handling Poles' and Section 5] The proof of Theorem 2 uses the ideal reciprocal activation 1/z. The model trained in Section 6 uses the clipped activation sgn(z)/max(|z|, epsilon) with epsilon=0.1. The paper provides no error bound relating the two and no argument that the clipped class is dense in C(chi). Because the clipped activation is bounded on R (by 1/epsilon) while 1/z is unbounded, the function classes are not identical; the universality guarantee, even if established for the canonical form, would not automatically transfer to the trained network.
minor comments (5)
  1. [Section 3] The statement 'We will interchangeably use the different forms in the paper based on convenience' is not justified for the restricted function classes; equivalence transformations hold for continued fractions of real numbers, but they do not preserve the requirement that each layer is a linear function of the input.
  2. [Section 4, 'Handling Poles'] The activation sgn(z)/max(|z|, epsilon) is not differentiable at z = ±epsilon, so the claim that 'All variants are differentiable' is inaccurate as stated.
  3. [Section 6.1] The synthetic experiments use a single ladder, while the universal approximation theorem concerns linear combinations of ladders; the text should clarify that the single-ladder fits are not covered by the theoretical guarantee.
  4. [Section 6.2] The ImageNet accuracy of 0.69 is reported without details of the setup (e.g., input resolution, training schedule), and it is not included in Table 1, which makes the comparison difficult to verify.
  5. [References] Reference [30] is a YouTube video; it is not a rigorous citation for a mathematical claim and should be replaced by a primary source.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the universal approximation proof is a self-contained application of Stone-Weierstrass and Euler's continued-fraction identity, and the interpretation checks are post-hoc comparisons rather than predictions forced by construction.

full rationale

The chain from Lemma 1 (PL is a unital subalgebra and separating) through Theorem 2 (PL is contained in finite linear combinations of CFL, so CFL is dense) is derived from external classical results — Stone-Weierstrass [45] and Euler's continued-fraction identity (9)-(10) — and not from the CoFrNet architecture's fitted parameters. The monomial construction in Theorem 2 explicitly sets v0=0, alpha0=c, wk=-vk=uk, and alphak=betak=0, so the density claim is not defined in terms of the conclusion. The IPS and IC interpretation sections compute power series and continuant-based gradients from a fitted ladder and then compare them to known ground-truth functions; these are empirical checks, not self-justifying predictions. Self-citations such as [13], [37], and [40] appear only in related-work context or as SOTA/baseline comparisons and do not carry the proof. The paper itself flags a limitation in Section 7: the epsilon-clipped activation 'may also limit the expressiveness of the function,' which is a scope caveat, not a circular step. One non-circular gap does remain: Theorem 2 is proved for the generalized CFL class of Definition 5, whose terms have affine numerators and denominators, whereas Eq. (2) and Figure 1 define the canonical unit-numerator ladder used in the experiments, and the paper gives no inclusion argument between the two; that is a correctness/rigor issue, not a reduction of the derivation to its own inputs.

Assumptions & free parameters 1 free parameters · 6 assumptions · 0 invented entities

The only substantive axioms are classical theorems. The paper's own unstated assumption that the proved CFL class coincides with the trained canonical architecture is the main unsupported premise. The pole-avoidance epsilon and the ideal-versus-clipped activation mismatch add further unproved assumptions.

free parameters (1)
  • epsilon (pole-avoidance threshold) = 0.1
    Chosen by hand in Section 6; the trained activation sgn(z)/max(|z|,epsilon) differs from the 1/z used in the theoretical analysis, and no bound is given for the approximation error introduced.
assumptions (6)
  • standard math Stone-Weierstrass theorem
    Invoked as Theorem 1 to conclude that P_L is dense in C([0,1]^p,R).
  • standard math Euler's continued fraction formula
    Used in Theorem 2, Eq. (9), to express polynomials of linear functionals as linear combinations of continued fractions.
  • standard math Rank-nullity theorem
    Used in Lemma 1 to find a vector u that separates two distinct points x and y.
  • domain assumption Domain assumption chi=[0,1]^p
    The universality proof is stated only on the compact cube; extension to other compact sets or all of R^p is asserted without proof.
  • ad hoc to paper Equivalence between proved class CFL and trained canonical architecture
    The paper does not establish that the canonical CoFrNet with unit numerators is in or dense in the CFL class of Definition 5, which has affine numerators and denominators. This is the main unsupported premise in the theoretical claim.
  • ad hoc to paper Ideal reciprocal activation in theory
    The universal approximation theorem and Proposition 1 assume exact 1/z, while experiments use sgn(z)/max(|z|,epsilon). The paper assumes the theoretical results extend to the clipped activation without proof.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CoFrNets: Interpretable Neural Architecture Inspired by Continued Fractions." pith.science (2026). https://pith.science/paper/QXCVVVWN

@misc{pith2026250605586,
  author       = {Pith},
  title        = {Pith review of: CoFrNets: Interpretable Neural Architecture Inspired by Continued Fractions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QXCVVVWN}},
  note         = {Machine review of arXiv:2506.05586}
}
read the original abstract

In recent years there has been a considerable amount of research on local post hoc explanations for neural networks. However, work on building interpretable neural architectures has been relatively sparse. In this paper, we present a novel neural architecture, CoFrNet, inspired by the form of continued fractions which are known to have many attractive properties in number theory, such as fast convergence of approximations to real numbers. We show that CoFrNets can be efficiently trained as well as interpreted leveraging their particular functional form. Moreover, we prove that such architectures are universal approximators based on a proof strategy that is different than the typical strategy used to prove universal approximation results for neural networks based on infinite width (or depth), which is likely to be of independent interest. We experiment on nonlinear synthetic functions and are able to accurately model as well as estimate feature attributions and even higher order terms in some cases, which is a testament to the representational power as well as interpretability of such architectures. To further showcase the power of CoFrNets, we experiment on seven real datasets spanning tabular, text and image modalities, and show that they are either comparable or significantly better than other interpretable models and multilayer perceptrons, sometimes approaching the accuracies of state-of-the-art models.

Figures

Figures reproduced from arXiv: 2506.05586 by the authors.

Figure 1
Figure 1. Single ladder (depth d) CoFrNet archi￾tecture on the left. On the right we see the corre￾sponding function computed at each stage. The proposed architecture is “simple”: there are only p weights to be learned in each layer as opposed to a quadratic number for standard ar￾chitectures such as multilayer perceptron (MLP) or those with densely connected layers.3 A key differentiation from other neural architec￾tures is … view at source ↗
Figure 2
Figure 2. Three variants of the CoFrNet architecture. In all three variants, the output (top triangle) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Original function (OF; in yellow) and the corre [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: In the first row for the three tabular datasets (a-c), we see the behavior of the most important [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Above (https://en.wikipedia.org/wiki/Coffer) we see an example of a coffer in building architecture, which is a series of (square) sunken panels [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Above we see 24 randomly chosen CIFAR10 test images (in grey scale) and to the [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: Above we see plots of the functions that represent the three most important variables for [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]
Figure 8
Figure 8. Figure 8: Above we see plots of the functions that represent the three most important variables for [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]
Figure 9
Figure 9. Figure 9: Above we see plots of the functions that represent the three most important variables for [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 45 canonical work pages

  1. [1]

    Agarwal, N

    R. Agarwal, N. Frosst, X. Zhang, R. Caruana, and G. E. Hinton. Neural additive models: Interpretable machine learning with neural nets. InarXiv:2004.13912, 2020

    arXiv 2004

  2. [2]

    Alvarez-Melis and T

    D. Alvarez-Melis and T. Jaakkola. Towards robust interpretability with self-explaining neural networks. InAdvances in Neural Information Processing Systems, pages 7775–7784. 2018

    work page 2018

  3. [3]

    A. P. Austin, M. Krishnamoorthy, S. Leyffer, S. Mrenna, J. M¨uller, and H. Schulz. Practical algorithms for multivariate rational approximation.Computer Physics Communications, 261, 2021

    work page 2021

  4. [4]

    Bourbaki.Elements of mathematics

    N. Bourbaki.Elements of mathematics. Springer-Verlag, 1989

    work page 1989

  5. [5]

    Breiman, J

    L. Breiman, J. Friedman, C. J. Stone, and R. Olshen.Classification and Regression Trees. Chapman and Hall, 1984

    work page 1984

  6. [6]

    Caruana, Y

    R. Caruana, Y . Lou, J. Gehrke, P. Koch, M. Sturm, and N. Elhadad. Intelligible models for healthcare: Predicting pneumonia risk and hospital 30-day readmission. InProceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’15, pages 1721–1730, New York, NY , USA, 2015. ACM

    work page 2015

  7. [7]

    Chrysos, S

    G. Chrysos, S. Moschoglou, G. Bouritsas, J. Deng, Y . Panagakis, and S. Zafeiriou. Deep polynomial neural networks. InIntl. Conference on Computer Vision and Pattern Recognition (CVPR), 2020

    work page 2020

  8. [8]

    Springer-Verlag, 1989

    Churchill, Brown, and Verhey.Complex Variables and Applications. Springer-Verlag, 1989

    work page 1989

Show all 53 references
  1. [9]

    G. Cybenko. Approximation by superpositions of a sigmoidal function.Mathematics of control, signals and systems, 2(4):303–314, 1989

    1989

  2. [10]

    S. Dash, O. G¨unl¨uk, and D. Wei. Boolean decision rules via column generation. InAdvances in Neural Information Processing Systems, pages 4655–4665, 2018

    2018

  3. [11]

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009

    2009

  4. [12]

    Dheeru and E

    D. Dheeru and E. Karra Taniskidou. UCI machine learning repository, 2017

    2017

  5. [13]

    Dhurandhar, P.-Y

    A. Dhurandhar, P.-Y . Chen, R. Luss, C.-C. Tu, P. Ting, K. Shanmugam, and P. Das. Explanations based on the missing: Towards contrastive explanations with pertinent negatives. InAdvances in Neural Information Processing Systems, pages 592–603, 2018

    2018

  6. [14]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. De- hghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. InIntl. Conference on Learning Represen...

    2021

  7. [15]

    Keras MNIST MLP implementation

    Fchollet, Matsuyamax, and Kemaswill. Keras MNIST MLP implementation. In https://github.com/fchollet/keras/blob/master/examples/mnist mlp.py, 2017

    2017

  8. [16]

    Keras MNIST CNN implementations

    Fchollet, Matsuyamax, Smerity, and Kemaswill. Keras MNIST CNN implementations. In https://github.com/fchollet/keras/blob/master/examples/mnist cnn.py, 2017

    2017

  9. [17]

    Foret, A

    P. Foret, A. Kleiner, H. Mobahi, and B. Neyshabur. Sharpness-aware minimization for efficiently improving generalization. InIntl Conference on Learning Representations (ICLR), 2021

    2021

  10. [18]

    T. J. Hastie and R. J. Tibshirani.Generalized Additive Models. Chapman and Hall/CRC, 1990

    1990

  11. [19]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. InIntl. Conference on Computer Vision and Pattern Recognition (CVPR), 2015

    2015

  12. [20]

    M. Hind, D. Wei, M. Campbell, N. C. F. Codella, A. Dhurandhar, A. Mojsilovic, K. N. Rama- murthy, and K. R. Varshney. TED: Teaching AI to explain its decisions. InProceedings of the AAAI/ACM Conference on AI, Ethics, and Society, pages 123–129, 2019. 11

    2019

  13. [21]

    K. Hornik. Some new results on neural network approximation.Neural networks, 6(8):1069– 1072, 1993

    1993

  14. [22]

    Huang, Z

    G. Huang, Z. Liu, L. V . D. Maaten, and K. Q. Weinberger. Densely connected convolutional networks. In2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2261–2269, 2017

    2017

  15. [23]

    W. B. Jones and W. Thron.Continued fractions. Analytic theory and applications. Encyclopedia of Mathematics and its Applications. Addison-Wesley, 1980

    1980

  16. [24]

    Quora insincere question dataset, 2019

    Kaggle. Quora insincere question dataset, 2019

    2019

  17. [25]

    Krizhevsky

    A. Krizhevsky. Learning multiple layers of features from tiny images, 2009

    2009

  18. [26]

    Lemhadri, L

    I. Lemhadri, L. A. Feng Ruan, and R. Tibshirani. LassoNet: A neural network with feature sparsity. InAISTATS, 2021

    2021

  19. [27]

    Z. Lu, H. Pu, F. Wang, Z. Hu, and L. Wang. The expressive power of neural networks: A view from the width. InAdvances in Neural Information Processing Systems, 2017

    2017

  20. [28]

    S. M. Lundberg, G. Erion, H. Chen, A. DeGrave, J. M. Prutkin, B. Nair, R. Katz, J. Himmelfarb, N. Bansal, and S.-I. Lee. From local explanations to global understanding with explainable ai for trees.Nature Mach. Intl., 2020

    2020

  21. [29]

    A. L. Maas, R. E. Daly, P. T. Pham, D. Huang, A. Y . Ng, and C. Potts. Learning word vectors for sentiment analysis. InThe 49th Annual Meeting of the Association for Computational Linguistics (ACL), 2020

    2020

  22. [30]

    How Ramanujan solved the Strand puzzle

    Mathologer. How Ramanujan solved the Strand puzzle. https://www.youtube.com/watch?v=V2BybLCmUzs, 2020

    2020

  23. [31]

    K. Milton. Summation techniques, Pad ´e approximants, and continued fractions. 2011. http: //www.nhn.ou.edu/~milton/p5013/chap8.pdf

    2011

  24. [32]

    Mitra and R

    S. Mitra and R. Sherwood. Canonic realizations of digital filters using the continued fraction expansion.IEEE Transactions on Audio and Electroacoustics, 20(3):185–194, 1972

    1972

  25. [33]

    Molina, P

    A. Molina, P. Schramowski, and K. Kersting. Pad ´e activation units: End-to-end learning of flexible activation functions in deep networks. InIntl. Conference on Learning Representations (ICLR), 2020

    2020

  26. [34]

    Molnar.Interpretable Machine Learning

    C. Molnar.Interpretable Machine Learning. https://christophm.github.io/interpretable-ml-book, 2019

    2019

  27. [35]

    Negahban and D

    S. Negahban and D. Shah. Learning sparse Boolean polynomials. In2012 50th Annual Allerton Conference on Communication, Control, and Computing (Allerton), pages 2032–2036. IEEE, 2012

    2012

  28. [36]

    H. Nori, S. Jenkins, P. Koch, and R. Caruana. InterpretML: A unified framework for machine learning interpretability. InarXiv:1909.09223, 2019

    1909 arXiv

  29. [37]

    Pedapati, A

    T. Pedapati, A. Balakrishnan, K. Shanmugam, and A. Dhurandhar. Learning global transparent models consistent with local contrastive explanations. InAdvances in Neural Information Processing Systems, 2020

    2020

  30. [38]

    Pennington, R

    J. Pennington, R. Socher, and C. D. Manning. Glove: Global vectors for word representation. InEmpirical Methods in Natural Language Processing (EMNLP), pages 1532–1543, 2014

    2014

  31. [39]

    M. W. Pownall.Functions and Graphs: Calculus Preparatory Mathematics. Prentice-Hall, 1983

    1983

  32. [40]

    K. N. Ramamurthy, B. Vinzamuri, Y . Zhang, and A. Dhurandhar. Model agnostic multilevel explanations. InAdvances in Neural Information Processing Systems, 2020. 12

    2020

  33. [41]

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

    2019

  34. [42]

    Sabour, N

    S. Sabour, N. Frosst, and G. E. Hinton. Dynamic routing between capsules. InAdvances in Neural Information Processing Systems, 2017

    2017

  35. [43]

    Sandler, A

    M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen. MobileNetV2: Inverted residuals and linear bottlenecks. In2018 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4510–4520, 2018

    2018

  36. [44]

    Srivastava, G

    N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting.Journal of Machine Learning Research, 15(56):1929–1958, 2014

    1929

  37. [45]

    M. H. Stone. Applications of the theory of Boolean rings to general topology.Transactions of the American Mathematical Society, 41(3):375–481, 1937

    1937

  38. [46]

    Swapna Rekha, J

    H. Swapna Rekha, J. Nayak, and H. S. Behera. Pi-sigma neural network: Survey of a decade progress. In A. K. Das, J. Nayak, B. Naik, S. Dutta, and D. Pelusi, editors,Computational Intelligence in Pattern Recognition, pages 429–441, Singapore, 2020. Springer Singapore

    2020

  39. [47]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, and I. Polosukhin. Attention is all you need. In I. Guyon, U. V . Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors,Advances in Neural Information Processing...

    2017

  40. [48]

    T.-W. Weng, H. Zhang, P.-Y . Chen, J. Yi, D. Su, Y . Gao, C.-J. Hsieh, and L. Daniel. Evaluating the robustness of neural networks: An extreme value theory approach. InProceedings of the International Conference on Learning Representations, 2018

    2018

  41. [49]

    Coffer, 2021

    Wikipedia. Coffer, 2021

    2021

  42. [50]

    Test functions for optimization, 2021

    Wikipedia. Test functions for optimization, 2021

    2021

  43. [51]

    Wolfram language function, 2020

    Wolfram Research. Wolfram language function, 2020

    2020

  44. [52]

    Z. Yang, Z. Dai, Y . Yang, J. Carbonell, R. Salakhutdinov, and Q. V . Le. XLNet: Generalized autoregressive pretraining for language understanding. InAdvances in Neural Inf. Processing Systems, 2019

    2019

  45. [53]

    some”, “people

    M. Yu, S. Chang, Y . Zhang, and T. S. Jaakkola. Rethinking cooperative rationalization: Intro- spective extraction and complement control. InEMNLP, 2019. 13 Figure 5: Above (https://en.wikipedia.org/wiki/Coffer) we see an example of a coffer in building architecture, which is ...

    2019

Pith tools

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