Pith. sign in

REVIEW 4 major objections 5 minor 61 references

Efficient Global Neural Architecture Search

T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read This paper claims a global neural architecture search method that ranks candidates with variable training epochs instead of one fixed protocol, making the search 2-4x faster than prior global methods while achieving new state-of-the-art…

desk verdict New empirical results and a clean mini-search-space design, but the 2-4x speedup claim rests on a ranking proxy validated only in-sample on N=50 CIFAR-10 networks with a confound between parameters and epochs. read the letter →

arxiv 2502.03553 v1 pith:CFIMFK2L submitted 2025-02-05 cs.CV

Shahid Siddiqui , Christos Kyrkou , Theocharis Theocharides This is my paper · ORCID
classification cs.CV
keywords neuralarchitecturesearchglobalmacro-microdynamiclearningrankingperformanceevaluationfacerecognitionconvolutionalnetworksAutoML
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 tries to make global neural architecture search, where the entire network including depth, width, and layer operations is discovered automatically, practical on a single GPU. Its central claim is that candidate networks should not all be trained under the same protocol when being ranked; instead, smaller networks should be trained a bit longer and larger ones less, so that relative rankings better predict final, fully trained accuracy. On that idea the authors build a search that first fixes depth and width (macro), then fine-tunes operations and kernels (micro), keeping parameter count roughly constant. They report state-of-the-art accuracy on EMNIST and KMNIST, competitive results on CIFAR and FashionMNIST, and 2-4x faster search than the fastest prior global methods, plus a transfer to face recognition. The paper matters because it claims to close the automation gap left by modular search, which finds building blocks but still requires humans to choose depth and width.

What carries the argument

Two mechanisms carry the argument. The first is the dynamic learning ranking scheme: candidates are trained with epochs adjusted relative to their parameter counts (fewer epochs for larger networks, more for smaller), and the paper reports that this improves rank correlation with final performance to 0.85 versus 0.65 for uniform short training. The second is the macro-micro search strategy: grow depth while accuracy improves, prune width while accuracy holds, then traverse layers to swap separable for plain convolutions and 3x3 for 5x5 kernels, compensating any parameter increase with channel cuts. The search complexity is bounded by evaluating roughly 2*Df + D' + W' architectures instead of the full combinatorial product.

What would settle it

Train 100 architectures from the paper's search space on two different datasets such as CIFAR-100 and FashionMNIST, compute dynamic-learning rankings with the paper's epoch rule, then fully train all candidates and compute the Spearman correlation with final accuracy. If the correlation falls close to the 0.65 static level, or comes out systematically lower than the reported 0.85, the proposed ranking advantage would fail to generalize and the search could be following noise rather than true architecture quality. Repeating the experiment with different epoch schedules, such as one versus two additional epochs, would also test sensitivity to that rule.

Watch

Extended reading notes

Core claim

The paper's core claim is that ranking architectures by short, architecture-aware training, giving more parameters fewer epochs and fewer parameters more epochs, yields Spearman correlations of 0.85 with final performance, roughly 20% higher than the static, same-protocol ranking that gives 0.65. This is presented as the mechanism that makes a genuinely global search feasible: because relative rankings are reliable without full convergence of every candidate, the search can afford to evaluate many candidates. The search then discovers depth and width first, then substitutes plain convolutions or larger kernels only when they improve validation accuracy, reducing channels to keep parameter counts roughly equal so that accuracy gains are attributable to architecture rather than size.

Load-bearing premise

The search's whole ranking philosophy assumes that validation accuracy after a handful of variable-epoch training runs predicts how the network will rank after full training, and this is backed by a single experiment on 50 CIFAR-10 networks, with the search thresholds tuned on the same dataset.

Editorial extensions

If this is right

  • Global NAS could run on a single GPU in hours rather than days, removing the main barrier to practical adoption.
  • The macro-micro split means the search directly outputs a deployable network, eliminating manual depth and width selection after search.
  • If the ranking mechanism generalizes, other global search algorithms could adopt the same variable-epoch ranking to speed themselves up.
  • The face recognition results suggest the framework transfers to other vision tasks, potentially yielding smaller and more accurate networks than hand-designed ResNets.

Reading between the lines

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

  • The dynamic ranking correlation of 0.85 was measured on only 50 CIFAR-10 networks with a particular pairing of parameters and epochs; on other datasets or with different hyperparameters the correlation might be lower, so the claim that dynamic ranking is truly architecture-aware should be tested more broadly.
  • The search thresholds L+acc+ and L+acc- are empirically tuned on CIFAR-10, so a systematic sensitivity analysis across datasets would clarify whether the method is robust or brittle to threshold choice.
  • If parameter-matched micro search truly isolates architectural gains, the same trick of normalizing for parameter count before accepting an operation change could be adapted to any NAS method.
  • The stated 2-4x speed-up is against a single prior baseline, NASH-Net; a direct runtime comparison with other global methods under identical hardware and hyperparameters would settle how general the speed advantage is.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a global neural architecture search framework with a compact macro-micro search space over depth, width, operation type, and kernel size. To evaluate candidates cheaply, it introduces a dynamic learning ranking scheme in which smaller networks are trained for fewer epochs and larger networks for more epochs. A greedy macro search grows depth and prunes width, followed by a micro search over operations and kernels. The authors report competitive error rates on CIFAR-10 and CIFAR-100, state-of-the-art results on EMNIST and KMNIST, and face verification/identification improvements over AdaFace baselines with smaller models, with claimed search costs 2-4x lower than the fastest global NAS methods.

Significance. If the dynamic ranking mechanism is sound, the work would be a genuinely useful contribution to efficient global NAS: it addresses macro architecture automatically, provides a diverse search space, evaluates candidates cheaply, and backs the search with public code and logs. The paper also follows useful evaluation practices, including a random-sampling baseline and the relative improvement metric, and it explicitly engages with the NAS best-practices literature. However, the central ranking claim rests on a single small in-sample experiment with no variance estimate, and the face-recognition section explicitly stops after macro search only. These issues make the efficiency and transferability claims load-bearing rather than cosmetic, so the current evidence is not yet at the level the paper's conclusions require.

major comments (4)
  1. [Section 3.2] The dynamic learning ranking is the load-bearing mechanism for the claimed 2-4x speedup, but it is validated by a single experiment on 50 CIFAR-10 networks with no variance estimate or out-of-sample check. The reported Spearman value of 0.85 against full training is in-sample with respect to the same 240-network space and the same benchmark used to motivate and tune the mechanism. I ask for validation on additional datasets (e.g., CIFAR-100, FashionMNIST) and repeated subsampling with confidence intervals, or a clear argument why CIFAR-10 alone is sufficient to transfer the ranking property across tasks.
  2. [Section 3.2] The dynamic ranking experiment is confounded: networks are sorted by parameter count and trained for an increasing number of epochs, so the parameter count and training duration increase together. The observed 0.91 correlation between parameters and dynamic validation accuracy, and the 0.85 correlation with final performance, may reflect the trivial effect of longer training rather than an architecture-aware ranking. I request controlled comparisons, for example random assignment of epoch budgets to networks, fixed total training budget across compared schemes, or independent variation of epochs while holding architecture size fixed, to show that the proposed schedule is what drives the higher correlation.
  3. [Section 4.2] The face-recognition experiments are explicitly limited to macro search only: the text states that due to the large number of training samples the authors 'early stop at best networks discovered by macro search only.' Consequently, Tables 6 and 7 demonstrate only the macro-search part of the framework, not the full macro-micro pipeline that is the paper's core contribution. The claim that the framework transfers to face recognition is therefore overstated. Please either run the full search (or a representative micro search) on the face datasets, or rephrase the claims to state clearly that only macro search was evaluated for this application.
  4. [Section 4.1, Table 3] The headline efficiency claim of being '2-4x faster than the fastest global search methods' is based on pooling literature numbers from different papers without reporting a controlled comparison on the same hardware and protocol. For example, NASH-Net and Macro-NAS are compared with their reported GPU-day costs, but the GPU model, training epochs per candidate, and data-loading details are not matched. A same-hardware timing comparison, or at least a transparent table of hardware and training settings for each baseline, is needed before the speedup claim can be taken as established.
minor comments (5)
  1. [Abstract and Section 1] There are repeated typos and spacing errors: 'FasionMNIST' should be 'FashionMNIST', and 'CIF AR' should be 'CIFAR' throughout.
  2. [Section 4.1 and References] The text refers to Macro-NAS using reference [13] in some places and [24] in Table 4; these appear to denote the same method, and the reference numbering should be reconciled.
  3. [Section 3.1] In the definition of the search space X, the width set is written as 'W∈ {Wmin + ne | n ∈ N0, e∈ E}', but E is not defined; please clarify the notation and the role of the step e.
  4. [Section 4.1] The paper does not clearly state which data split is used to make search decisions and to select the final architecture; please specify that validation data are used and that test data are not touched until final evaluation.
  5. [Table 2] The initialization strategy table reports accuracy on ten binary subdatasets sampled from CIFAR-10, but the construction of these binary subdatasets is not described; please provide the sampling procedure and the number of classes per subdataset.

Circularity Check

1 steps flagged · score 5.0 of 10

Dynamic-learning ranking is validated on the same 50 CIFAR-10 networks used to define it, making the 0.85 correlation an in-sample fit; final benchmark searches remain out-of-sample, so the central claim retains independent content.

  1. fitted input called prediction [Section 3.2, 'Dynamic Learning Rankings']
    "Next we sample 50 networks from 240 and sort w.r.t. their parameters in ascending order. For each network increasing in parameters, we record its validation accuracy against training for an additional epoch. ... We observe a high correlation value of 0.91 between params and corresponding 'validation accuracy with increasing epochs' (dynamic learning rankings). ... In addition, networks ranked using dynamic learning have a high correlation value of 0.85 with their final performance, ie 20% better than static learning rankings."

    The dynamic-ranking protocol is defined from the same 50-network CIFAR-10 sample on which the 0.85 correlation with final performance is then measured; no held-out networks or datasets are used. The 0.85 is therefore a goodness-of-fit on the construction sample, not a predictive validation. Because the paper then adopts this ranking as the evaluation mechanism for the search on CIFAR-10, the reported speed and accuracy evidence for the core claim rests on an in-sample correlation. The result is also confounded, as the paper concedes ('due to increased parameters or better training, or both'): parameter count and training epochs increase together, so the correlation partly encodes training budget rather than architectural quality.

full rationale

The only concrete circularity is the in-sample validation of the dynamic-learning ranking in Section 3.2. The paper samples 240 CIFAR-10 networks, then uses 50 of them both to construct the variable-epoch ranking and to report the 0.85 Spearman correlation with final performance; there is no out-of-sample check that this ranking generalizes to CIFAR-100, EMNIST, KMNIST, FashionMNIST, or CASIA-WebFace. The thresholds L_acc+ = 0.10 and L_acc- = 0.05 in Section 3.3 are also set on CIFAR-10 and applied back to CIFAR-10, which weakens the independence of the CIFAR-10 result, though this is a hyperparameter-selection issue rather than a by-construction equivalence. The final benchmark architectures on EMNIST, KMNIST, and FashionMNIST are produced by full searches on those datasets, so the headline SOTA claims are not forced by the CIFAR-10 fit. Self-citations to [47] are not scored as circular: the paper re-derives and extends the search space and ranking, and prior publication provides external evidence. The face-recognition section explicitly limits itself to macro-search ('we early stop at best networks discovered by macro search only'), which undercuts the full-framework transferability claim but is an acknowledged limitation, not circularity. Overall, partial circularity in one load-bearing validation, with independent content in the final architecture results, supports a moderate score.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The method relies on multiple hand-set thresholds and a data-dependent training schedule. The most consequential free parameters are the accuracy-gain/drop thresholds and the epoch schedule, both of which are tuned on the same datasets used for evaluation. The core assumptions are that short variable-epoch training predicts final rankings and that macro-micro separation is lossless.

free parameters (7)
  • L+_acc+ (depth growth threshold) = 0.10
    Empirically determined in Section 3.3; used to decide whether adding a layer is beneficial.
  • L+_acc- (depth/prune drop threshold) = 0.05
    Empirically determined in Section 3.3; used to stop depth growth or channel pruning after three drops below this value.
  • Emin search epochs = 10 (macro), 2 (micro)
    Chosen for speed; sensitivity not analyzed.
  • Wres channel pruning step = 2
    Fixed pruning stride for width search.
  • Search bounds Dmin/Dmax/Wmin/Wmax = 5-100 depth, 16-128 channels depending on setting
    Defines the search space; chosen to balance navigability and diversity.
  • Dynamic training epoch schedule = add 1 epoch per added layer, add 2 epochs per pruned channel
    Heuristic derived from the 50-network dynamic ranking experiment; not theoretically justified.
  • Initialization strategy = Sep-64-3x3
    Selected from Table 2 based on 10 binary CIFAR-10 subdatasets; a data-dependent choice.
assumptions (4)
  • domain assumption Validation accuracy from short, variable-epoch training correlates with final fully-trained performance across the search space.
    Empirically tested on only 50 CIFAR-10 networks (Section 3.2); no theoretical justification or cross-dataset validation.
  • domain assumption Macro and micro architecture variables can be optimized separately without losing the global optimum.
    The search splits depth/width from operations/kernels (Section 3.3); this disjunction is not proven and could miss architectures where the two interact.
  • domain assumption Trimming the search space (no FC, no pooling, no skip connections, fixed strides, width only in the first layer) preserves enough expressiveness.
    Motivated by FCN and VGG design principles but not compared against including these variables.
  • domain assumption Standard NAS training hyperparameters (SGD, cosine schedule, cutout) are fair across compared methods.
    The paper uses standard settings but does not control for the training protocols of compared methods, which can affect accuracy more than architecture.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient Global Neural Architecture Search." pith.science (2026). https://pith.science/paper/CFIMFK2L

@misc{pith2026250203553,
  author       = {Pith},
  title        = {Pith review of: Efficient Global Neural Architecture Search},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CFIMFK2L}},
  note         = {Machine review of arXiv:2502.03553}
}
read the original abstract

Neural architecture search (NAS) has shown promise towards automating neural network design for a given task, but it is computationally demanding due to training costs associated with evaluating a large number of architectures to find the optimal one. To speed up NAS, recent works limit the search to network building blocks (modular search) instead of searching the entire architecture (global search), approximate candidates' performance evaluation in lieu of complete training, and use gradient descent rather than naturally suitable discrete optimization approaches. However, modular search does not determine network's macro architecture i.e. depth and width, demanding manual trial and error post-search, hence lacking automation. In this work, we revisit NAS and design a navigable, yet architecturally diverse, macro-micro search space. In addition, to determine relative rankings of candidates, existing methods employ consistent approximations across entire search spaces, whereas different networks may not be fairly comparable under one training protocol. Hence, we propose an architecture-aware approximation with variable training schemes for different networks. Moreover, we develop an efficient search strategy by disjoining macro-micro network design that yields competitive architectures in terms of both accuracy and size. Our proposed framework achieves a new state-of-the-art on EMNIST and KMNIST, while being highly competitive on the CIFAR-10, CIFAR-100, and FashionMNIST datasets and being 2-4x faster than the fastest global search methods. Lastly, we demonstrate the transferability of our framework to real-world computer vision problems by discovering competitive architectures for face recognition applications.

Figures

Figures reproduced from arXiv: 2502.03553 by the authors.

Figure 1
Figure 1. Tiny (T) models discovered for CIFAR-10, EMNIST, KMNIST and Fashion￾MNIST datasets. Block height and width represents image resolution (R) and number of channels (C) respectively, while depth is represented by the total number of blocks. Different colors in Micro Search Space represent possible operation types and kernel sizes. Due to space limitations, we do not show CIFAR-100 and mobile networks. mean accuracy tha… view at source ↗
Figure 2
Figure 2. Architectural comparison of a. R-18-512 with our b. R-58-256. architecture and training protocol search to improve overall accuracy and search efficiency. 5 Conclusion Contrary to trending modular search that offers partial network discovery, we revisit NAS for end-to-end network discovery. Our search space offers networks of varying complexity and is designed such that it can be efficiently navigated by the propose… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

61 extracted references · 46 canonical work pages

  1. [1]

    ”Very deep convolutional networks for large-scale image recognition.” arXiv preprint arXiv:1409.1556 (2014)

    Simonyan, Karen, and Andrew Zisserman. ”Very deep convolutional networks for large-scale image recognition.” arXiv preprint arXiv:1409.1556 (2014)

    arXiv 2014

  2. [2]

    ”Fully convolutional networks for semantic segmentation.” Proceedings of the IEEE conference on computer vision and pattern recognition

    Long, Jonathan, Evan Shelhamer, and Trevor Darrell. ”Fully convolutional networks for semantic segmentation.” Proceedings of the IEEE conference on computer vision and pattern recognition. 2015

    work page 2015

  3. [3]

    ”Deep residual learning for image recognition.” Proceedings of the IEEE conference on computer vision and pattern recognition

    He, Kaiming, et al. ”Deep residual learning for image recognition.” Proceedings of the IEEE conference on computer vision and pattern recognition. 2016

    2016

  4. [4]

    Zoph, Barret, and Quoc V. Le. ”Neural architecture search with reinforcement learn- ing.” arXiv preprint arXiv:1611.01578 (2016)

    arXiv 2016

  5. [5]

    ”Designing neural network architectures using reinforcement learning.” arXiv preprint arXiv:1611.02167 (2016)

    Baker, Bowen, et al. ”Designing neural network architectures using reinforcement learning.” arXiv preprint arXiv:1611.02167 (2016)

    arXiv 2016

  6. [6]

    ”Large-scale evolution of image classifiers.” International Con- ference on Machine Learning

    Real, Esteban, et al. ”Large-scale evolution of image classifiers.” International Con- ference on Machine Learning. PMLR, 2017

    work page 2017

  7. [7]

    ”A genetic pro- gramming approach to designing convolutional neural network architectures.” Pro- ceedings of the genetic and evolutionary computation conference

    Suganuma, Masanori, Shinichi Shirakawa, and Tomoharu Nagao. ”A genetic pro- gramming approach to designing convolutional neural network architectures.” Pro- ceedings of the genetic and evolutionary computation conference. 2017

    work page 2017

  8. [8]

    ”Simple and effi- cient architecture search for convolutional neural networks.” arXiv preprint arXiv:1711.04528 (2017)

    Elsken, Thomas, Jan-Hendrik Metzen, and Frank Hutter. ”Simple and effi- cient architecture search for convolutional neural networks.” arXiv preprint arXiv:1711.04528 (2017)

    arXiv 2017

Show all 61 references
  1. [9]

    ”Learning transferable architectures for scalable image recogni- tion.” Proceedings of the IEEE conference on computer vision and pattern recogni- tion

    Zoph, Barret, et al. ”Learning transferable architectures for scalable image recogni- tion.” Proceedings of the IEEE conference on computer vision and pattern recogni- tion. 2018

    2018

  2. [10]

    ”Efficient architecture search by network transformation.” Pro- ceedings of the AAAI Conference on Artificial Intelligence

    Cai, Han, et al. ”Efficient architecture search by network transformation.” Pro- ceedings of the AAAI Conference on Artificial Intelligence. Vol. 32. No. 1. 2018

    2018

  3. [11]

    ”Neural architecture search with bayesian optimi- sation and optimal transport.” Advances in neural information processing systems 31 (2018)

    Kandasamy, Kirthevasan, et al. ”Neural architecture search with bayesian optimi- sation and optimal transport.” Advances in neural information processing systems 31 (2018)

    2018

  4. [12]

    ”Dpp-net: Device-aware progressive search for pareto- optimal neural architectures.” Proceedings of the European Conference on Com- puter Vision (ECCV)

    Dong, Jin-Dong, et al. ”Dpp-net: Device-aware progressive search for pareto- optimal neural architectures.” Proceedings of the European Conference on Com- puter Vision (ECCV). 2018

    2018

  5. [13]

    ”Macro neural architecture search revisited.” 2nd Workshop on Meta-Learning at NeurIPS

    Hu, Hanzhang, et al. ”Macro neural architecture search revisited.” 2nd Workshop on Meta-Learning at NeurIPS. 2018

    2018

  6. [14]

    Efficient neural architecture search via parameters sharing

    Pham H, Guan M, Zoph B, Le Q, Dean J. Efficient neural architecture search via parameters sharing. InInternational conference on machine learning 2018 Jul 3 (pp. 4095-4104). PMLR

    2018

  7. [15]

    ”Darts: Differentiable architec- ture search.” arXiv preprint arXiv:1806.09055 (2018)

    Liu, Hanxiao, Karen Simonyan, and Yiming Yang. ”Darts: Differentiable architec- ture search.” arXiv preprint arXiv:1806.09055 (2018). Efficient Global Neural Architecture Search 17

    2018 arXiv

  8. [16]

    ”Searching for a robust neural architecture in four gpu hours.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Dong, Xuanyi, and Yi Yang. ”Searching for a robust neural architecture in four gpu hours.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019

    2019

  9. [17]

    ”Efficient forward architecture search.” Advances in Neural Information Processing Systems 32 (2019)

    Hu, Hanzhang, et al. ”Efficient forward architecture search.” Advances in Neural Information Processing Systems 32 (2019)

    2019

  10. [18]

    Esperan¸ ca, and Fabio M

    Yang, Antoine, Pedro M. Esperan¸ ca, and Fabio M. Carlucci. ”NAS evaluation is frustratingly hard.” arXiv preprint arXiv:1912.12522 (2019)

    2019 arXiv

  11. [19]

    ”Nsga-net: neural architecture search using multi-objective ge- netic algorithm.” Proceedings of the genetic and evolutionary computation confer- ence

    Lu, Zhichao, et al. ”Nsga-net: neural architecture search using multi-objective ge- netic algorithm.” Proceedings of the genetic and evolutionary computation confer- ence. 2019

    2019

  12. [20]

    ”Best practices for scientific research on neu- ral architecture search.” The Journal of Machine Learning Research 21.1 (2020): 9820-9837

    Lindauer, Marius, and Frank Hutter. ”Best practices for scientific research on neu- ral architecture search.” The Journal of Machine Learning Research 21.1 (2020): 9820-9837

    2020

  13. [21]

    ”A comprehensive survey of neural architecture search: Chal- lenges and solutions.” ACM Computing Surveys (CSUR) 54.4 (2021): 1-34

    Ren, Pengzhen, et al. ”A comprehensive survey of neural architecture search: Chal- lenges and solutions.” ACM Computing Surveys (CSUR) 54.4 (2021): 1-34

    2021

  14. [22]

    ”How powerful are performance predictors in neural archi- tecture search?.” Advances in Neural Information Processing Systems 34 (2021): 28454-28469

    White, Colin, et al. ”How powerful are performance predictors in neural archi- tecture search?.” Advances in Neural Information Processing Systems 34 (2021): 28454-28469

    2021

  15. [23]

    ”AGNAS: Attention-Guided Micro and Macro-Architecture Search.” International Conference on Machine Learning

    Sun, Zihao, et al. ”AGNAS: Attention-Guided Micro and Macro-Architecture Search.” International Conference on Machine Learning. PMLR, 2022

    2022

  16. [24]

    Alexandre

    Lopes, Vasco, and Lu ´ ıs A. Alexandre. ”Towards Less Constrained Macro-Neural Architecture Search.” arXiv preprint arXiv:2203.05508 (2022)

    2022 arXiv

  17. [25]

    ”Neural architecture search using progres- sive evolution.” Proceedings of the Genetic and Evolutionary Computation Confer- ence

    Sinha, Nilotpal, and Kuan-Wen Chen. ”Neural architecture search using progres- sive evolution.” Proceedings of the Genetic and Evolutionary Computation Confer- ence. 2022

    2022

  18. [26]

    Evaluating the search phase of neural architecture search

    Kaicheng Yu, Christian Sciuto, Martin Jaggi, Claudiu Musat, and Mathieu Salz- mann. Evaluating the search phase of neural architecture search. In International Conference on Learning Representations, 2019

    2019

  19. [27]

    ”Neural architecture search: Insights from 1000 papers.” arXiv preprint arXiv:2301.08727 (2023)

    White, Colin, Mahmoud Safari, Rhea Sukthanker, Binxin Ru, Thomas Elsken, Ar- ber Zela, Debadeepta Dey, and Frank Hutter. ”Neural architecture search: Insights from 1000 papers.” arXiv preprint arXiv:2301.08727 (2023)

    2023 arXiv

  20. [28]

    Jahs-bench-201: A foundation for research on joint architecture and hyperparameter search

    Bansal A, Stoll D, Janowski M, Zela A, Hutter F. Jahs-bench-201: A foundation for research on joint architecture and hyperparameter search. Advances in Neural Information Processing Systems. 2022 Dec 6;35:38788-802

    2022

  21. [29]

    Baker, B., Gupta, O., Raskar, R., and Naik, N. (2017). Accelerating neural archi- tecture search using performance prediction. arXiv preprint arXiv:1705.10823

    2017 arXiv

  22. [30]

    Zheng, X., Ji, R., Tang, L., Zhang, B., Liu, J., and Tian, Q. (2019). Multinomial distribution learning for effective neural architecture search. In Proceedings of the IEEE International Conference on Computer Vision (pp. 1304-1313)

    2019

  23. [31]

    Neural predictor for neural architecture search

    Wei Wen, Hanxiao Liu, Hai Li, Yiran Chen, Gabriel Bender, and Pieter- Jan Kindermans. Neural predictor for neural architecture search. arXiv preprint arXiv:1912.00848, 2019

    1912 arXiv

  24. [32]

    NAS-Bench-Suite- Zero: Accelerating research on zero cost proxies

    Krishnakumar A, White C, Zela A, Tu R, Safari M, Hutter F. NAS-Bench-Suite- Zero: Accelerating research on zero cost proxies. Advances in Neural Information Processing Systems. 2022 Dec 6;35:28037-51

    2022

  25. [33]

    Nas-bench-zero: A large scale dataset for understanding zero-shot neural architecture search

    Hanlin Chen, Ming Lin, Xiuyu Sun, and Hao Li. Nas-bench-zero: A large scale dataset for understanding zero-shot neural architecture search. Openreview preprint https://openreview.net/forum?id=hP-SILoczR, 2021

    2021

  26. [34]

    Evaluating efficient performance estimators of neu- ral architectures

    Xuefei Ning, Changcheng Tang, Wenshuo Li, Zixuan Zhou, Shuang Liang, Huazhong Yang, and Yu Wang. Evaluating efficient performance estimators of neu- ral architectures. Advances in Neural Information Processing Systems, 34, 2021 18 S. Siddiqui et al

    2021

  27. [35]

    Spearman, C. (1904). The proof and measurement of association between two things. American Journal of Psychology, 15(1), 72-101

    1904

  28. [36]

    Wavemix-lite: A resource-efficient neural net- work for image analysis

    Jeevan P, Viswanathan K, Sethi A. Wavemix-lite: A resource-efficient neural net- work for image analysis. arXiv preprint arXiv:2205.14375. 2022 May 28

    2022 arXiv

  29. [37]

    NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search

    Xuanyi Dong and Yi Yang. NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search. In ICLR, 2020

    2020

  30. [38]

    M. S. Tanveer, M. U. Karim Khan and C. -M. Kyung, ”Fine-Tuning DARTS for Image Classification,” 2020 25th International Conference on Pattern Recognition (ICPR), Milan, Italy, 2021, pp. 4789-4796, doi: 10.1109/ICPR48806.2021.9412221

    2020

  31. [39]

    An evolutionary approach to dynamic introduction of tasks in large-scale multitask learning systems

    Gesmundo A, Dean J. An evolutionary approach to dynamic introduction of tasks in large-scale multitask learning systems. arXiv preprint arXiv:2205.12755. 2022 May 25

    2022 arXiv

  32. [40]

    DeVries, Terrance and Taylor, Graham. (2017). Improved Regularization of Con- volutional Neural Networks with Cutout

    2017

  33. [41]

    https://www.cs.toronto.edu/ kriz/cifar.html

  34. [42]

    Fashion-mnist: a novel image dataset for benchmark- ing machine learning algorithms

    Xiao H, Rasul K, Vollgraf R. Fashion-mnist: a novel image dataset for benchmark- ing machine learning algorithms. arXiv preprint arXiv:1708.07747. 2017 Aug 25

    2017 arXiv

  35. [43]

    EMNIST: Extending MNIST to hand- written letters

    Cohen G, Afshar S, Tapson J, Van Schaik A. EMNIST: Extending MNIST to hand- written letters. In2017 international joint conference on neural networks (IJCNN) 2017 May 14 (pp. 2921-2926). IEEE

    2017

  36. [44]

    Deep learning for classical japanese literature

    Clanuwat T, Bober-Irizar M, Kitamoto A, Lamb A, Yamamoto K, Ha D. Deep learning for classical japanese literature. arXiv preprint arXiv:1812.01718. 2018 Dec 3

    2018 arXiv

  37. [45]

    Imagenet large scale visual recognition challenge

    Russakovsky O, Deng J, Su H, Krause J, Satheesh S, Ma S, Huang Z, Karpathy A, Khosla A, Bernstein M, Berg AC. Imagenet large scale visual recognition challenge. International journal of computer vision. 2015 Dec;115:211-52

    2015

  38. [46]

    Sgas: Sequential greedy architecture search

    Li G, Qian G, Delgadillo IC, Muller M, Thabet A, Ghanem B. Sgas: Sequential greedy architecture search. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition 2020 (pp. 1620-1630)

    2020

  39. [47]

    Siddiqui, S., Kyrkou, C., Theocharides, T. (2023). True Rank Guided Efficient Neural Architecture Search for End to End Low-Complexity Network Discovery. In: Tsapatsoulis, N., et al. Computer Analysis of Images and Patterns. CAIP 2023

    2023

  40. [48]

    Learning face representation from scratch

    Yi D, Lei Z, Liao S, Li SZ. Learning face representation from scratch. arXiv preprint arXiv:1411.7923. 2014 Nov 28

    2014 arXiv

  41. [49]

    Labeled faces in the wild: A database forstudying face recognition in unconstrained environments

    Huang GB, Mattar M, Berg T, Learned-Miller E. Labeled faces in the wild: A database forstudying face recognition in unconstrained environments. InWorkshop on faces in’Real-Life’Images: detection, alignment, and recognition 2008

    2008

  42. [50]

    Frontal to profile face verification in the wild

    Sengupta S, Chen JC, Castillo C, Patel VM, Chellappa R, Jacobs DW. Frontal to profile face verification in the wild. In2016 IEEE winter conference on applications of computer vision (W ACV) 2016 Mar 7 (pp. 1-9). IEEE

    2016

  43. [51]

    Cross-pose lfw: A database for studying cross-pose face recog- nition in unconstrained environments

    Zheng T, Deng W. Cross-pose lfw: A database for studying cross-pose face recog- nition in unconstrained environments. Beijing University of Posts and Telecommu- nications, Tech. Rep. 2018 Feb;5(7):5

    2018

  44. [52]

    Agedb: the first manually collected, in-the-wild age database

    Moschoglou S, Papaioannou A, Sagonas C, Deng J, Kotsia I, Zafeiriou S. Agedb: the first manually collected, in-the-wild age database. Inproceedings of the IEEE conference on computer vision and pattern recognition workshops 2017 (pp. 51-59)

    2017

  45. [53]

    Cross-age lfw: A database for studying cross-age face recognition in unconstrained environments

    Zheng T, Deng W, Hu J. Cross-age lfw: A database for studying cross-age face recognition in unconstrained environments. arXiv preprint arXiv:1708.08197. 2017 Aug 28. Efficient Global Neural Architecture Search 19

    2017 arXiv

  46. [54]

    Iarpa janus benchmark-b face dataset

    Whitelam C, Taborsky E, Blanton A, Maze B, Adams J, Miller T, Kalka N, Jain AK, Duncan JA, Allen K, Cheney J. Iarpa janus benchmark-b face dataset. In- proceedings of the IEEE conference on computer vision and pattern recognition workshops 2017 (pp. 90-98)

    2017

  47. [55]

    Iarpa janus benchmark-c: Face dataset and proto- col

    Maze B, Adams J, Duncan JA, Kalka N, Miller T, Otto C, Jain AK, Niggel WT, Anderson J, Cheney J, Grother P. Iarpa janus benchmark-c: Face dataset and proto- col. In 2018 international conference on biometrics (ICB) 2018 Feb 20 (pp. 158-165). IEEE

    2018

  48. [56]

    Low-resolution face recognition

    Cheng Z, Zhu X, Gong S. Low-resolution face recognition. InComputer Vi- sion–ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, De- cember 2–6, 2018, Revised Selected Papers, Part III 14 2019 (pp. 605-621). Springer International Publishing

    2018

  49. [57]

    Arcface: Additive angular margin loss for deep face recognition

    Deng J, Guo J, Xue N, Zafeiriou S. Arcface: Additive angular margin loss for deep face recognition. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition 2019 (pp. 4690-4699)

    2019

  50. [58]

    Adaface: Quality adaptive margin for face recognition

    Kim M, Jain AK, Liu X. Adaface: Quality adaptive margin for face recognition. InProceedings of the IEEE/CVF conference on computer vision and pattern recog- nition 2022 (pp. 18750-18759)

    2022

  51. [59]

    Rethinking bias mitigation: Fairer architectures make for fairer face recognition

    Dooley S, Sukthanker R, Dickerson J, White C, Hutter F, Goldblum M. Rethinking bias mitigation: Fairer architectures make for fairer face recognition. Advances in Neural Information Processing Systems. 2024 Feb 13;36

    2024

  52. [60]

    Teacher guided neural architecture search for face recognition

    Wang X. Teacher guided neural architecture search for face recognition. InPro- ceedings of the AAAI Conference on Artificial Intelligence 2021 May 18 (Vol. 35, No. 4, pp. 2817-2825)

    2021

  53. [61]

    Pocketnet: Extreme lightweight face recognition network using neural architecture search and multistep knowledge distillation

    Boutros F, Siebke P, Klemt M, Damer N, Kirchbuchner F, Kuijper A. Pocketnet: Extreme lightweight face recognition network using neural architecture search and multistep knowledge distillation. IEEE Access. 2022 Apr 26;10:46823-33

    2022

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.