Principled Curriculum Learning using Parameter Continuation Methods

Harsh Nilesh Pathak; Randy Paffenroth

REVIEW 4 major objections 5 minor 40 references

Principled Curriculum Learning using Parameter Continuation Methods

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper proposes treating neural network training as tracing a smooth path of minima from an easy problem to the hard one, and shows that arclength continuation generalizes better than ADAM on MNIST.

desk verdict PARC for NNs is a reasonable idea, but this submission doesn't implement it correctly and doesn't verify the theory it rests on. read the letter →

arxiv 2507.22089 v1 pith:BBPP4AKG submitted 2025-07-29 cs.LG cs.AI

Harsh Nilesh Pathak , Randy Paffenroth This is my paper · ORCID

classification cs.LGcs.AI

keywords pseudo-arclengthcontinuationhomotopymethodscurriculumlearningneuralnetworkoptimizationparameternon-convexgeneralizationMNIST

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

This paper argues that the standard way of training neural networks—random initialization plus a direct optimizer such as ADAM—discards the structure that makes hard optimization problems tractable. Instead, the authors embed the target problem in a homotopy that starts from an easy problem and gradually becomes the hard one, and they train by following the resulting curve of critical points. The central move is to parameterize that curve by arclength rather than by the homotopy parameter, which is what pseudo-arclength continuation (PARC) does. On MNIST autoencoder and classifier tasks, PARC achieves better test loss or accuracy than ADAM in four of five settings. The authors claim the approach is theoretically justified by the Implicit Function Theorem, which guarantees the solution curve is smooth and unique locally, while acknowledging that global path structure is not covered.

What carries the argument

The central object is Pseudo-arclength Continuation (PARC), adapted to a first-order, matrix-free form for neural networks. It replaces the fixed homotopy step $\Delta\lambda$ with an arclength step $\Delta s$: a secant predictor advances $(\theta,\lambda)$ along the approximate tangent to the solution curve, and a corrector solves the original loss plus an orthogonality penalty $\gamma(\Delta\theta\cdot\dot\theta + \Delta\lambda\cdot\dot\lambda)$ that keeps iterates on the hyperplane perpendicular to the secant. This lets $\lambda$ adapt during the corrector, so the path can be followed around folds where $\theta(\lambda)$ ceases to be single-valued.

What would settle it

Run PARC on a network while recording the smallest singular value of $\nabla_\theta H$ at every continuation step; if the path reaches a fold where this value drops to zero and PARC's corrector then jumps to a different basin, visible as a sudden change in test loss or failure to converge on a problem where ADAM converges, the central claim that arclength tracking keeps iterates in the basin of attraction would be refuted.

Watch

Extended reading notes

Core claim

The central claim is that neural network training should be seen as path-following, not as a single-shot minimization. Given a hard loss $L(\theta)$, the paper defines a homotopy $\tilde L(\theta,\lambda)=\lambda L(\theta)+(1-\lambda)M(\theta)$ with an easy problem $M$ and considers the critical-point equation $H(\theta,\lambda)=\nabla_\theta \tilde L(\theta,\lambda)=0$. By the Implicit Function Theorem, a regular solution $(\theta_0,0)$ extends locally to a smooth curve $(\theta(s),\lambda(s))$ of critical points. The authors' PARC algorithm follows this curve using a secant predictor and a corrector that enforces orthogonality to the secant, so $\lambda$ is adapted automatically and folds on the solution path do not break the tracking. Empirically, on downsampled 6x6 MNIST, PARC with activation and brightness homotopies reaches lower test loss than ADAM on the three-layer autoencoder and higher test accuracy on the one-layer classifier in four of five task/homotopy combinations.

Load-bearing premise

The load-bearing premise is that there is a smooth, unbroken curve of solutions connecting the easy problem to the hard one, and that following it closely enough keeps every iterate in the basin of the right minimum; the paper's own theory only guarantees this locally, not across the whole training run.

Editorial extensions

If this is right

If the method works as claimed, neural network training can be formulated as following a family of minima, so each harder task starts inside the basin of attraction of the previous solution, reducing the role of lucky random initialization.
The first-order PARC avoids Hessian computation, so the arclength strategy can be applied to networks with millions of parameters.
The method improves test loss or accuracy over ADAM on four of five MNIST tasks, suggesting that continuation buys a better critical point, not just better training dynamics.
Because PARC adapts the homotopy parameter during the corrector, it can continue through folds where natural parameter continuation (NPC) would fail.

Reading between the lines

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

Not tested in the paper: the same arclength predictor-corrector scheme could be applied to other curriculum axes, such as data ordering or loss weighting, giving a principled rule for scheduling curricula instead of hand-designed heuristics.
Not tested in the paper: if PARC preserves basins of attraction, it could serve as a fine-tuning or continual-learning tool, tracing from a previously learned solution to a new task's solution rather than restarting from random initialization.
Not tested in the paper: because global path regularity is unproven, monitoring the smallest singular value of $\nabla_\theta H$ during training and adding explicit branch-switching at bifurcations could extend PARC to cases where the easy-to-hard path is not unique.

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.

Desk Editor's Note

read the letter

X, here's my take. The paper's idea is to replace random initialization with continuation along a path of solutions, using arclength parameterization to handle folds. That's a legitimate and interesting direction, and the authors clearly know the classical continuation literature. The two homotopies (activation and brightness) are sensible, and the small experiments on MNIST do show lower test loss than ADAM in 4 of 5 configurations, which is something.

But the paper doesn't hold together. Algorithm 1 is not Keller's pseudo-arclength continuation. A proper PARC step augments the system with an orthogonality condition and corrects both θ and λ. Here the predictor uses a secant but doesn't produce a normalized joint tangent, the corrector updates only θ and never λ, and the orthogonality condition is implemented as a penalty in the loss. The text says λ is 'simultaneously adapted' but the algorithm doesn't do that. That's an internal contradiction, not a minor detail.

The theory section is a restatement of the IFT. Theorem 3.1 is textbook, but the paper never checks its hypotheses for the homotopies used. h-ReLU is not C1, so the theorem's C1 assumption fails outright. For h-sigmoid, no one verifies that ∇θH is nonsingular along the path. The paper even concedes no global path guarantees exist, which is honest but leaves the method with no theoretical backing.

The experiments are too thin to carry the empirical claim: one dataset (6x6 MNIST), small networks, no error bars, no code, no hyperparameter details. And the tables actually contain a counterexample to the abstract's claim: PARC with h-Brightness in Table 2 is substantially worse than ADAM (0.772 vs 0.780 accuracy). The paper mentions this as an exception, but it contradicts 'better generalization performance than state-of-the-art optimization techniques such as ADAM' without qualification.

The novelty is also limited. The authors have already applied continuation to neural networks in their 2019 ICMLA paper and 2024 thesis. The first-order PARC is a minor extension, and as written it's not even a correct one. The citation pattern is self-heavy, which isn't automatically a problem, but here most of the cited prior work is their own, and the current submission looks like a rehash with a new name.

Who is this for? Someone interested in the intersection of numerical continuation and deep learning might skim it for references, but the technical flaws make it unreliable as a guide. I would not send this to peer review in its current form. The right move is to desk reject and suggest that the authors fix the algorithm, verify the IFT conditions or state them as assumptions, and run proper experiments with multiple seeds and a real dataset.

Referee Report

4 major / 5 minor

Summary. The paper proposes to train neural networks by numerically tracing a family of minimizers of a homotopy between an easy objective and the target objective, rather than by direct optimization from a random initialization. The authors introduce two homotopies based on activation functions and a 'brightness' transformation, and present an algorithm called Pseudo-arclength Continuation (PARC) that alternates a secant predictor with a gradient-descent corrector on a loss augmented with an orthogonality penalty. The claimed contributions are a theoretically justified curriculum learning framework and better generalization than ADAM on supervised and unsupervised tasks. The experiments are limited to a downsampled MNIST autoencoder and a one-layer classifier.

Significance. If the central claim were established, the paper would offer a principled way to think about curriculum learning as path following on neural network loss surfaces, and PARC could be a useful alternative to initialization-based training. The paper is also honest about a key difficulty: it states that no global guarantees exist for the solution path and that the arclength parameterization is motivated by local IFT considerations. This framing, and the connection to the numerical continuation literature, is a strength. However, the theoretical justification is not instantiated for the actual networks and homotopies used, the algorithm as written does not implement pseudo-arclength continuation, and the empirical evidence is a single-run comparison on one dataset with no error bars, no hyperparameter details, and one configuration that is worse than ADAM. The significance of the claimed result is therefore not demonstrated in the current manuscript.

major comments (4)

[§3, Theorem 3.1 and Eq. (3)] The theoretical justification consists of stating the Implicit Function Theorem for H(θ,λ)=∇_θ \tilde{L}(θ,λ), but the hypotheses of the theorem are never verified for any of the homotopies used in the experiments. In particular, the h-ReLU activation is not C^1 at the kink, so Theorem 3.1 does not apply to that homotopy at all. For the smooth h-Sigmoid homotopy, the required nonsingularity of ∇_θ H (the Hessian of the loss with respect to θ) is never checked, either analytically or numerically. The paper itself concedes in §3 that no global claims about the solution path exist; that concession is appropriate, but it also means the claimed theoretical justification is restricted to an unverified local statement that does not cover the numerical experiments.
[§4, Algorithm 1] Algorithm 1 as written is not Keller's pseudo-arclength continuation, and the difference is load-bearing. In true pseudo-arclength continuation, the predictor uses a normalized tangent direction of the joint curve (θ(s),λ(s)), and the corrector solves the augmented system consisting of H(θ,λ)=0 together with the arclength/orthogonality constraint, updating both θ and λ. Here, lines 5–6 normalize the θ-increment and the λ-increment separately, rather than normalizing the joint increment; the corrector in lines 9–15 updates only θ and never corrects λ; and the orthogonality condition appears as a penalty term in a modified loss rather than as a constraint on the augmented system. Consequently, even if a regular branch exists, the manuscript provides no argument or diagnostic that the iterates remain on the solution path or that the λ sequence is arclength-parameterized.
[§5, Tables 1-2] The empirical claim of better generalization than ADAM is not supported by the evidence. There is exactly one run per configuration, no error bars or seeds, no reported learning rate, batch size, number of epochs, γ, Δs, or λ schedule, and only one dataset (MNIST downsampled to 6×6) with small networks. Moreover, Table 2 shows that PARC with h-Brightness has test loss 0.731 and test accuracy 0.772, both worse than the standard ADAM baseline (0.675 and 0.780). The conclusion states improvement on '4/5 optimization tasks,' but the two tables report only two tasks, and the tabulated runs are not identified in a way that makes this count verifiable. The paper would need multi-seed results, a complete hyperparameter table, and a resolution of the inconsistent PARC-h-Brightness result in order to substantiate the abstract's claim of better generalization.
[§4, homotopy definitions] The experimental homotopies are not fully specified. Equation (2) defines a general linear homotopy between L and M, and the text gives h(z)=(1−λ)z+λ sigmoid(z) for the h-Sigmoid activation, but the h-ReLU and h-Brightness homotopies are not defined. In particular, 'Brightness Homotopy' appears only in Tables 1–2 and is never stated mathematically. Without these definitions, the experiments cannot be reproduced, and the paper's claimed connection between the theory and the reported results cannot be checked.

minor comments (5)

[§1, Introduction] The citation 'Kathleen et al., 1997' likely refers to Alligood, Sauer, and Yorke's 'Chaos: An Introduction to Dynamical Systems'; the reference list entry should be corrected and completed.
[§4, Algorithm 1] Line 3 uses θ̇ and λ̇ before they are defined; the secant-based tangent should be defined before it appears in the orthogonality penalty.
[§5, Experiments] The phrase 'downsized to 6×6' should specify the interpolation method and whether the same preprocessing was used for all methods, since this affects the comparison.
[General formatting] The paper is a 2025 arXiv submission but is formatted as an ICML 2021 proceedings paper with a copyright notice; the venue/version information should be cleaned up before any submission.
[§6, Conclusion] The statement 'better generalization performance for 4/5 optimization tasks' is ambiguous; the paper reports two tasks with multiple homotopy variants, so the count should be made explicit and tied to specific table rows.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's central claims are empirical and its theoretical scaffolding is an external standard theorem, not a self-referential definition.

full rationale

The paper's central claim is that PARC generalizes better than ADAM on supervised and unsupervised tasks. This is an empirical claim supported by the tables in Section 5, which compare continuation-trained networks against a standard ADAM baseline on held-out test data; it is not derived from the cited prior work. The theoretical justification in Section 3 is the standard Implicit Function Theorem quoted from Allgower and Georg, an external numerics textbook, and the paper explicitly disclaims global path guarantees in Section 3 ('to the best of our knowledge there are no such claims on the global structure of the solution path'), so the invocation of IFT is not a self-citation that forces the result. The paper does cite several works by the same authors for homotopy-based training and the arclength parameterization, but those citations are not load-bearing for the empirical comparison, and the algorithmic content in Algorithm 1 is stated in the paper itself. There is no equation in which a predicted quantity is defined in terms of the fitted output, and no fitted parameter is renamed as a prediction. Concerns raised by a skeptical reader, such as the unverified C1 regularity of h-ReLU homotopies or the discrepancy between Algorithm 1 and Keller's pseudo-arclength continuation, are correctness and rigor risks rather than circularity: they do not make the claimed result equivalent to its inputs by construction. The paper even reports in Table 2 one PARC configuration that is worse than ADAM, which is inconsistent with a forced or definitional outcome. Accordingly, no circular step is identified.

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

The paper's method depends on several handpicked hyperparameters and unverified assumptions about the loss landscape. No new entities are introduced.

free parameters (3)

gamma = not reported
Weight of the orthogonality constraint in the loss in Algorithm 1; chosen by hand, no value given.
delta_s (arclength step) = not reported
Step size for predictor in Algorithm 1; affects tracing accuracy and convergence.
homotopy type and lambda schedule = not reported
The choice among h-sigmoid, h-ReLU, and h-Brightness and the discretization of lambda are handpicked; h-Brightness is not defined in the text.

assumptions (4)

standard math Implicit Function Theorem (IFT) holds for the homotopy H(theta,lambda)=0 at the starting point.
Used to guarantee local smooth solution path, cited from Allgower & Georg; but conditions not checked for neural networks.
domain assumption A global solution path of minimizers exists and is traceable for neural network homotopies.
The paper assumes continuation can connect easy to hard tasks without crossing singularities that break the arclength parameterization; Section 3 admits no global theory.
domain assumption The corrector with an orthogonality penalty tracks the solution path accurately.
Algorithm 1 uses a penalty term to enforce orthogonality, but no proof that this substitutes for the exact pseudo-arclength constraint.
domain assumption Loss surfaces are C^1 (H is C^1) in parameters and lambda.
Needed for IFT; neural network losses with ReLU are not C^1, a point not addressed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Principled Curriculum Learning using Parameter Continuation Methods." pith.science (2026). https://pith.science/paper/BBPP4AKG

@misc{pith2026250722089,
  author       = {Pith},
  title        = {Pith review of: Principled Curriculum Learning using Parameter Continuation Methods},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BBPP4AKG}},
  note         = {Machine review of arXiv:2507.22089}
}

read the original abstract

In this work, we propose a parameter continuation method for the optimization of neural networks. There is a close connection between parameter continuation, homotopies, and curriculum learning. The methods we propose here are theoretically justified and practically effective for several problems in deep neural networks. In particular, we demonstrate better generalization performance than state-of-the-art optimization techniques such as ADAM for supervised and unsupervised learning tasks.

Figures

Figures reproduced from arXiv: 2507.22089 by the authors.

**Figure 1.** Pseudo-arclength Continuation of singularities such as folds (points which cannot be parameterized by λ) and bifurcations. To mitigate this issue we propose a more principled predictor-corrector framework to provide a robust tracking around singularities in solution paths. Pseudo-arclength Continuation (PARC) for Neural Networks is the main contribution of this paper. Originally, PARC use second and third order deri… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 29 canonical work pages

[1]

and Georg, K

Allgower, E. and Georg, K. Introduction to Numerical Continuation Methods. Society for Industrial and Applied Mathematics, 2003. doi:10.1137/1.9780898719154. URL https://epubs.siam.org/doi/abs/10.1137/1.9780898719154

doi:10.1137/1.9780898719154 2003
[2]

Sur les courbes définies par des équations différentielles

Bendixson, I. Sur les courbes définies par des équations différentielles . Acta Mathematica, 24 0 (none): 0 1 -- 88, 1901. doi:10.1007/BF02403068. URL https://doi.org/10.1007/BF02403068

work page doi:10.1007/bf02403068 1901
[3]

Curriculum learning, 2009

Bengio, Y., Louradour, J., Collobert, R., and Weston, J. Curriculum learning, 2009

work page 2009
[4]

N., and Yano, T

Chandrasekaran, R., Pathak, H. N., and Yano, T. Deep neural query understanding system at expedia group. In 2020 IEEE International Conference on Big Data (Big Data), pp.\ 1476--1484. IEEE, 2020

work page 2020
[5]

H., Sherwood, W., LaMar, M., and Guckenheimer, J

Clewley, R. H., Sherwood, W., LaMar, M., and Guckenheimer, J. Pydstool, a software environment for dynamical systems modeling. URL http://pydstool. sourceforge. net, 2007

work page 2007
[6]

Dhooge, A., Govaerts, W., and Kuznetsov, Y. A. Matcont: A matlab package for numerical bifurcation analysis of odes. SIGSAM Bull., 38 0 (1): 0 21–22, March 2004. ISSN 0163-5824. doi:10.1145/980175.980184. URL https://doi.org/10.1145/980175.980184

work page arXiv 2004
[7]

J., Fairgrieve, T

Doedel, E. J., Fairgrieve, T. F., Sandstede, B., Champneys, A. R., Kuznetsov, Y. A., and Wang, X. Auto-07p: Continuation and bifurcation software for ordinary differential equations, 2007

work page 2007
[8]

Adaptive subgradient methods for online learning and stochastic optimization

Duchi, J., Hazan, E., and Singer, Y. Adaptive subgradient methods for online learning and stochastic optimization. Journal of machine learning research, 12: 0 2121--2159, 2011. ISSN 1532-4435. URL http://dl.acm.org/citation.cfm?id=1953048.2021068

work page arXiv 2011

Show all 40 references

[9]

Deep Learning

Goodfellow, I., Bengio, Y., and Courville, A. Deep Learning. MIT Press, 2016. http://www.deeplearningbook.org

2016
[10]

Noisy activation functions

G \" u l c ehre, C ., Moczulski, M., Denil, M., and Bengio, Y. Noisy activation functions. CoRR, abs/1603.00391, 2016. URL http://arxiv.org/abs/1603.00391

2016 arXiv
[11]

and Weinshall, D

Hacohen, G. and Weinshall, D. On the power of curriculum learning in training deep networks. CoRR, abs/1904.03626, 2019. URL http://arxiv.org/abs/1904.03626

1904 arXiv
[12]

Deep residual learning for image recognition

He, K., Zhang, X., Ren, S., and Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 770--778, 2016

2016
[13]

Exploring neural network structure through sparse recurrent neural networks: A recasting and distillation of neural network hyperparameters

Hershey, Q., Paffenroth, R., and Pathak, H. Exploring neural network structure through sparse recurrent neural networks: A recasting and distillation of neural network hyperparameters. In 2023 International Conference on Machine Learning and Applications (ICMLA), pp.\ 128--135...

2023
[14]

Rethinking the relationship between recurrent and non-recurrent neural networks: A study in sparsity, 2024

Hershey, Q., Paffenroth, R., Pathak, H., and Tavener, S. Rethinking the relationship between recurrent and non-recurrent neural networks: A study in sparsity, 2024. URL https://arxiv.org/abs/2404.00880

2024 arXiv
[15]

Rmsprop: Divide the gradient by a running average of its recent magnitude

Hinton, G., Srivastava, N., and Swersky, K. Rmsprop: Divide the gradient by a running average of its recent magnitude. Neural networks for machine learning, Coursera lecture 6e, 2012

2012
[16]

Progressive growing of gans for improved quality, stability, and variation

Karras, T., Aila, T., Laine, S., and Lehtinen, J. Progressive growing of gans for improved quality, stability, and variation. CoRR, abs/1710.10196, 2017. URL http://arxiv.org/abs/1710.10196

2017 arXiv
[17]

Chaos: An introduction to dynamical systems

Kathleen, T., Tim, D., and James, A. Chaos: An introduction to dynamical systems. Physics Today, 50: 0 67--68, 1997

1997
[18]

Keller, H. B. Numerical solution of bifurcation and nonlinear eigenvalue problems . In Rabinowitz, P. H. (ed.), Applications of Bifurcation Theory, pp.\ 359--384, New York, 1977. Academic Press

1977
[19]

Keller, H. B. Global homotopies and newton methods. In DE BOOR , C. and GOLUB, G. H. (eds.), Recent Advances in Numerical Analysis, pp.\ 73--94. Academic Press, 1978. ISBN 978-0-12-208360-0. doi:https://doi.org/10.1016/B978-0-12-208360-0.50009-7. URL https://www.sciencedirect....

1978 doi
[20]

Kingma, D. P. and Ba, J. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. URL http://arxiv.org/abs/1412.6980

2014 arXiv
[21]

Neue beitr \"a ge zur riemann'schen functionentheorie

Klein, F. Neue beitr \"a ge zur riemann'schen functionentheorie. Mathematische Annalen, 21 0 (2): 0 141--218, 1883
[22]

Une m \'e thode de r \'e solution d'une cat \'e gorie d' \'e quations transcendantes

Lahaye, E. Une m \'e thode de r \'e solution d'une cat \'e gorie d' \'e quations transcendantes. Comptes rendus des s \'e ances de l'Acad \'e mie des sciences. Vie acad \'e mique , 197: 0 1840--1842, 1934

1934
[23]

Deep learning

LeCun, Y., Bengio, Y., and Hinton, G. Deep learning. Nature, 2015. URL https://www.nature.com/articles/nature14539

2015
[24]

and Schauder, J

Leray, J. and Schauder, J. Topologie et \'e quations fonctionnelles. In Annales scientifiques de l' \'E cole normale sup \'e rieure , volume 51, pp.\ 45--78, 1934

1934
[25]

May , R. M. Simple mathematical models with very complicated dynamics . Nature, 261 0 (5560): 0 459--467, June 1976. doi:10.1038/261459a0

1976 doi
[26]

Curriculum dropout

Morerio, P., Cavazza, J., Volpi, R., Vidal, R., and Murino, V. Curriculum dropout. In Proceedings of the IEEE International Conference on Computer Vision, pp.\ 3544--3552, 2017

2017
[27]

and Clinton Paffenroth, R

Nilesh Pathak, H. and Clinton Paffenroth, R. Non-convex optimization using parameter continuation methods for deep neural networks. In Deep Learning Applications, Volume 2, pp.\ 273--298. Springer, 2020

2020
[28]

and Paffenroth, R

Nilesh Pathak, H. and Paffenroth, R. Parameter continuation methods for the optimization of deep neural networks. In 2019 18th IEEE International Conference On Machine Learning And Applications (ICMLA), pp.\ 1637--1643, 2019. doi:10.1109/ICMLA.2019.00268

2019
[29]

Pathak, H. N. Parameter continuation with secant approximation for deep neural networks. PhD thesis, Worcester Polytechnic Institute, 2018

2018
[30]

Pathak, H. N. Advancing Neural Network Optimization and Design Through the Lens of Continuation Methods and Iterative Dynamical Systems. PhD thesis, Worcester Polytechnic Institute, 2024

2024
[31]

Pathak, H. N. and Paffenroth, R. Solo connection: A parameter efficient fine-tuning technique for transformers, 2025. URL https://arxiv.org/abs/2507.14353

2025 arXiv
[32]

N., Li, X., Minaee, S., and Cowan, B

Pathak, H. N., Li, X., Minaee, S., and Cowan, B. Efficient super resolution for large-scale images using attentional gan. In 2018 IEEE International Conference on Big Data (Big Data), pp.\ 1777--1786, 2018. doi:10.1109/BigData.2018.8622477

2018
[33]

N., Paffenroth, R., and Hershey, Q

Pathak, H. N., Paffenroth, R., and Hershey, Q. Sequentia12d: Organizing center of skip connections for transformers. In 2023 International Conference on Machine Learning and Applications (ICMLA), pp.\ 362--368. IEEE, 2023

2023
[34]

Curriculum by smoothing

Sinha, S., Garg, A., and Larochelle, H. Curriculum by smoothing. Advances in Neural Information Processing Systems, 33, 2020

2020
[35]

T., Rota, P., and Sebe, N

Soviany, P., Ionescu, R. T., Rota, P., and Sebe, N. Curriculum learning: A survey. arXiv preprint, 2021

2021
[36]

T., Rota, P., and Sebe, N

Soviany, P., Ionescu, R. T., Rota, P., and Sebe, N. Curriculum learning: A survey. International Journal of Computer Vision, 130 0 (6): 0 1526--1565, 2022

2022
[37]

The T rilinos P roject W ebsite

T rilinos P roject T eam, T. The T rilinos P roject W ebsite
[38]

Wang, R., Lehman, J., Clune, J., and Stanley, K. O. Paired open-ended trailblazer (POET): endlessly generating increasingly complex and diverse learning environments and their solutions. CoRR, abs/1901.01753, 2019. URL http://arxiv.org/abs/1901.01753

1901 arXiv
[39]

and Cohen, G

Weinshall, D. and Cohen, G. Curriculum learning by transfer learning: Theory and experiments with deep networks. CoRR, abs/1802.03796, 2018. URL http://arxiv.org/abs/1802.03796

2018 arXiv
[40]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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

[1] [1]

and Georg, K

Allgower, E. and Georg, K. Introduction to Numerical Continuation Methods. Society for Industrial and Applied Mathematics, 2003. doi:10.1137/1.9780898719154. URL https://epubs.siam.org/doi/abs/10.1137/1.9780898719154

doi:10.1137/1.9780898719154 2003

[2] [2]

Sur les courbes définies par des équations différentielles

Bendixson, I. Sur les courbes définies par des équations différentielles . Acta Mathematica, 24 0 (none): 0 1 -- 88, 1901. doi:10.1007/BF02403068. URL https://doi.org/10.1007/BF02403068

work page doi:10.1007/bf02403068 1901

[3] [3]

Curriculum learning, 2009

Bengio, Y., Louradour, J., Collobert, R., and Weston, J. Curriculum learning, 2009

work page 2009

[4] [4]

N., and Yano, T

Chandrasekaran, R., Pathak, H. N., and Yano, T. Deep neural query understanding system at expedia group. In 2020 IEEE International Conference on Big Data (Big Data), pp.\ 1476--1484. IEEE, 2020

work page 2020

[5] [5]

H., Sherwood, W., LaMar, M., and Guckenheimer, J

Clewley, R. H., Sherwood, W., LaMar, M., and Guckenheimer, J. Pydstool, a software environment for dynamical systems modeling. URL http://pydstool. sourceforge. net, 2007

work page 2007

[6] [6]

Dhooge, A., Govaerts, W., and Kuznetsov, Y. A. Matcont: A matlab package for numerical bifurcation analysis of odes. SIGSAM Bull., 38 0 (1): 0 21–22, March 2004. ISSN 0163-5824. doi:10.1145/980175.980184. URL https://doi.org/10.1145/980175.980184

work page arXiv 2004

[7] [7]

J., Fairgrieve, T

Doedel, E. J., Fairgrieve, T. F., Sandstede, B., Champneys, A. R., Kuznetsov, Y. A., and Wang, X. Auto-07p: Continuation and bifurcation software for ordinary differential equations, 2007

work page 2007

[8] [8]

Adaptive subgradient methods for online learning and stochastic optimization

Duchi, J., Hazan, E., and Singer, Y. Adaptive subgradient methods for online learning and stochastic optimization. Journal of machine learning research, 12: 0 2121--2159, 2011. ISSN 1532-4435. URL http://dl.acm.org/citation.cfm?id=1953048.2021068

work page arXiv 2011