DANCE: Resource-Efficient Neural Architecture Search with Data-Aware and Continuous Adaptation

Lixin Zou; Maolin Wang; Ruocheng Guo; Shanshan Ye; Sheng Zhang; Tianshuo Wei; Wanyu Wang; Xiangyu Zhao; Xuetao Wei

REVIEW 5 major objections 5 minor 44 references

DANCE: Resource-Efficient Neural Architecture Search with Data-Aware and Continuous Adaptation

T0 review · 5 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 DANCE claims that a single pre-trained supernet, guided by a learned distribution over layer-wise gates, can output an accurate architecture for any new resource constraint at O(1) search cost.

desk verdict The paper's own Table 1 contradicts its 'consistently outperforms' claim, and the O(1) adaptation story is untested for unseen budgets; still, the framework is coherent and the direction is worth pursuing. read the letter →

arxiv 2507.04671 v2 pith:4BMSCN3O submitted 2025-07-07 cs.LG cs.CV

Maolin Wang , Tianshuo Wei , Sheng Zhang , Ruocheng Guo , Wanyu Wang , Shanshan Ye , Lixin Zou , Xuetao Wei

show 1 more author

Xiangyu Zhao

This is my paper · ORCID

classification cs.LGcs.CV

keywords neuralarchitecturesearchcontinuousdistributionSelectGatefeature-levelpruningsupernetresource-adaptivesamplingGumbel-Softmaximageclassification

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

DANCE (Dynamic Architectures with Neural Continuous Evolution) tries to establish that neural architecture search can be replaced by learning one continuous distribution over layer-wise architectural choices, so that a single trained supernet serves every deployment budget. The paper argues that existing NAS methods fail in practice because each scenario demands a separate discrete search, architectures are static once chosen, and performance is inconsistent across platforms. DANCE's answer is a factorized distribution $p(A|D,C)$ over selection gates that read both the data batch and the resource constraint, from which an architecture can be sampled directly. If the claim holds, deploying a model to a new hardware budget would cost one forward pass instead of a full search, while preserving or improving accuracy.

What carries the argument

The load-bearing object is the factorized architecture distribution $p(A|D,C)=\prod_l p(g_l|F_l,D,C)$, implemented by per-layer SelectGate modules. SelectGate fuses a batch embedding with concurrent layer features, maps the combination through a two-layer MLP, applies Gumbel-Softmax for reparameterized exploration, re-weights by the importance score $\mathrm{Score}(A_l)$ (combining static, dynamic, feature, and correlation-penalty terms), and ends with Bernoulli sampling that respects the layer budget $C_l$. This machinery is what turns architecture search into a continuous, differentiable, data- and constraint-aware sampling problem and is what the paper credits for smooth adaptation and low-cost deployment.

What would settle it

Train DANCE with constraints $C_l\in\{0.3,0.4,0.5\}$ as in Appendix C, then sample architectures at unseen constraints such as $0.35$, $0.65$, and $0.85$ and compare their accuracy against the same SuperNet fine-tuned at those exact constraints, or against an oracle that selects the best of many random samples at the same FLOPs; if accuracy at unseen constraints falls sharply or scatters relative to interpolation and retraining, the learned distribution has overfit its training constraints and the continuous-adaptation claim fails.

Watch

Extended reading notes

Core claim

The paper's central discovery is that architecture search can be formulated as continuous evolution over feature dimensions: rather than searching a discrete space of blocks, DANCE learns a distribution $p(A|D,C)=\prod_{l=1}^{L}p(g_l|F_l,D,C)$ over layer-wise binary gates. Each gate is produced by a SelectGate module that combines batch-level and layer-level features, applies Gumbel-Softmax for differentiable exploration, weights components by a four-part importance score (static, dynamic, feature, and correlation), and samples exactly $C_l$ components per layer via Bernoulli sampling. A three-stage training procedure pre-trains a SuperNet, jointly optimizes network weights and gate parameters through dual-loop distribution-guided learning, then fine-tunes scenario-specific SubNets by inheriting the SuperNet weights. The paper reports that direct Stage-2 sampling from the learned distribution, without retraining, already beats several baselines that receive full fine-tuning, and that retrained DANCE subnets achieve the best accuracy across five datasets with far fewer parameters, supporting the claim of $O(1)$ architecture derivation per new scenario.

Load-bearing premise

The whole O(1)-adaptation story rests on the learned gates working for resource budgets and deployment scenarios they were never trained on, and the paper does not demonstrate that in the main text.

Editorial extensions

If this is right

After one SuperNet pre-training, a new deployment constraint is served by one forward pass through SelectGate instead of a separate NAS run, so marginal search cost per scenario is essentially zero.
Because gates are sampled from a distribution conditioned on $C$, changing the budget changes the architecture smoothly rather than restarting the search, which should keep accuracy stable across hardware tiers.
The reported Stage-2 results imply that architecture quality comes from the learned distribution, not from expensive fine-tuning, so fast deployment no longer depends on per-scenario optimization.
The accuracy gains at 2.1 to 3.0 million parameters on ResNet-18 across five datasets imply that the method can compress heavily while retaining most of the accuracy of much larger pretrained models.

Reading between the lines

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

Because Appendix C trains gates only at $C_l\in\{0.3,0.4,0.5\}$, a natural test the paper does not run is to probe intermediate constraints such as $0.35$, $0.65$, and $0.85$; if those samples underperform, the 'continuous evolution' claim would need to be downgraded to interpolation between trained points.
The factorized form $p(A|D,C)$ treats layers as conditionally independent given features; the same SelectGate idea could be extended to transformers by gating attention heads or FFN dimensions, where cross-layer coupling is stronger and may require a non-factorized gate.
The comparison against score-based pruning suggests the distribution itself, not the importance metric, is the source of gain; an even sharper test would be to feed the same learned scores into an evolutionary or reinforcement-learning search and compare sample efficiency at matched FLOPs.

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

Honestly, the headline claim doesn't survive the paper's own experiments. On CIFAR-10 with ResNet-18, SPOS reaches 89.12% while DANCE (retrained) gets 87.19%; on VGG-16, RecNAS gets 83.80% versus DANCE's 81.32%. The abstract says 'consistently outperforms state-of-the-art NAS approaches' — that's contradicted by Table 1. This is a load-bearing flaw, and the reader's take is right.

What's genuinely useful here is the framework itself. DANCE learns a distribution over layer-wise gates conditioned on data and resource constraint, which unifies NAS and pruning in a clean way. The three-stage training (supernet pre-training, distribution-guided learning, deployment fine-tuning) is sensible, and the ablation of the four importance metrics gives some insight into what matters. The code and appendix are available. That's real effort.

Now the soft spots. First, the baseline comparison is unreliable. The OFA numbers look implausibly low — OFA-Width 1.0 on CIFAR-10/ResNet-18 at 78.23% when the original pre-trained model is 93.31% suggests the baselines weren't given a fair setting. FixMatch is a semi-supervised learning method, not a NAS baseline, so it shouldn't be in that table. Second, search cost — a headline contribution — is never quantified. No GPU hours, no epochs beyond vague statements. Third, the continuous-adaptation claim is untested. Stage 2 trains the SelectGate only with layer-wise budgets {0.3, 0.4, 0.5}, but the paper claims smooth adaptation to any budget and O(1) sampling. Table 2 uses budgets 0.5–0.9, mostly outside the training range, and the promised 0.02-step sweep (0.1–0.5) never appears. The gate's generalization to unseen constraints is exactly what's needed to support the O(1) story, and it's missing. Fourth, the gate loss function is not specified; Appendix C mentions a combination of classification, sparsity, diversity, correlation, and stability terms, but there's no equation. That's a reproducibility gap.

On novelty: the paper extends the authors' own DNS-Rec to image classification at feature level. That's incremental but not necessarily invalid.

Overall: the problems are in the evaluation, not in the core idea. With a fair benchmark, reported search cost, and a held-out constraint test, this could become a useful contribution. As is, I would not trust the reported numbers. I'd send it to peer review — a good referee could force the authors to fix the evaluation — but the letter should make clear that the current version is not acceptable.

Who's this for? Researchers in efficient NAS or runtime-adaptive models who want to see how a learned architecture distribution works. But they should treat the experimental claims with skepticism.

Recommendation: engage with the work as a promising-but-unverified framework, not as evidence for SOTA accuracy.

Referee Report

5 major / 5 minor

Summary. The paper proposes DANCE, a neural architecture search framework that learns a continuous distribution p(A|D,C) over layer-wise selection gates, enabling architecture sampling under varying resource constraints at O(1) cost after training. The method is evaluated on five image datasets with ResNet-18 and VGG-16 backbones, comparing against OFA, RecNAS, SPOS, FixMatch, and score-based pruning, and the authors claim consistent accuracy improvements over state-of-the-art NAS methods with significantly reduced search costs.

Significance. If the central claims held, DANCE would be a valuable contribution: it would provide a single pre-trained supernet from which architectures can be sampled for arbitrary resource constraints without per-scenario searches, bridging NAS and pruning through a feature-dimension perspective. The paper also provides a public code and appendix, which is a strength for reproducibility. However, the headline accuracy claim is directly contradicted by the paper's own Table 1, and the resource-adaptation and search-cost claims are not supported by the reported experiments. As presented, the significance is not established.

major comments (5)

[§3.5, Table 1] The abstract's claim that DANCE 'consistently outperforms state-of-the-art NAS approaches in terms of accuracy' is contradicted by the paper's own results. On CIFAR-10 with ResNet-18, DANCE (Retrained) reaches 87.19% while SPOS reaches 89.12%; on CIFAR-10 with VGG-16, DANCE (Retrained) reaches 81.32% while RecNAS reaches 83.80%. Additionally, on Food-101 with ResNet-18, DANCE (Retrained) obtains 83.80%, below the original pretrained model's 85.01%. The 'consistent' accuracy claim is therefore not supported by the data in Table 1.
[§2.2, Eq. (2); §2.4; §3.2; Appendix C] The load-bearing premise that p(A|D,C) enables smooth, O(1) adaptation to arbitrary resource constraints is never tested at held-out constraints. Stage 2 trains with layer-wise constraints C_l selected only from {0.3, 0.4, 0.5} (Appendix C), yet §2.4 claims O(1) sampling 'for new scenarios' and §3.2 promises a fine-grained 0.02-step sweep from 0.1 to 0.5. That sweep never appears; Table 2 covers constraints 0.9 down to 0.5, of which only 0.5 overlaps the training set. The paper should report results at held-out constraints such as 0.35, 0.65, and 0.75, and compare DANCE's sampled architectures against re-running Stage 2 or fine-tuning, to substantiate the interpolation claim.
[§3.6, Figures 3 and 4] The comparison between DANCE and score-based pruning is not a clean test of the distribution-guided learning mechanism. SelectGate's MLP and the importance weights are trained on the same datasets and resource constraints used for evaluation, whereas the score-based pruning baseline is a fixed rule applied without training. The observed advantage on training distributions is therefore partly a consequence of fitting those distributions, not evidence that the learned distribution generalizes. A fair test would evaluate both methods on held-out data distributions and on held-out constraints, with matched training budgets, to separate memorization from generalization.
[§3.5 and abstract] The claim of 'significantly reducing search costs' is not backed by any measured search-cost comparison. Table 1 reports only parameters and accuracy; there is no wall-clock time, GPU hours, number of epochs, or comparison of total search cost against baselines such as OFA or RecNAS. The O(1) sampling cost in §2.4 is a theoretical statement and does not account for the costs of Stage 1 and Stage 2 training, which are part of the method's total resource consumption. The paper should report measured search costs, including supernet pretraining and distribution-guided training, for a fair efficiency comparison.
[§3.5, Table 1 footnote] The statistical significance marker ('*', p < 0.05) is not substantiated. The paper does not state the number of independent runs, the standard deviations, or the statistical test used (e.g., paired t-test, Wilcoxon). Without this information, the reader cannot verify the claimed 'statistically significant better than the second best' results, and some entries in Table 1 do not appear to be better than the second best at all. The authors should report mean ± std over multiple seeds and specify the test.

minor comments (5)

[§3.4 vs. Appendix C] Section 3.4 describes a 'two-stage approach' while the abstract and Appendix C describe a three-stage process; the implementation details should be consistent.
[Figure 4 caption] The caption contains a typo: 'Purning' should be 'Pruning'.
[§2.3 and Eq. (4)] Notation for gates is inconsistent: the text uses g_l, g^{(l)}_f, and g^{(l,t)}_f without explicitly defining how they relate; Eq. (4) uses g_l while earlier equations use g^{(l)}_f.
[References] Several references have formatting issues, e.g., 'Zhaok et al.' (2021) and 'DONG et al.' (2023) are inconsistently capitalized, and 'Howard, 2017' lacks a page range; a careful reference pass is needed.
[Appendix D] The hyperparameter sensitivity analysis covers batch size, α, λ2, and temperature τ, but not λ1, λ3, and λ4; since all four metrics are claimed to be necessary, the omitted λ sensitivity should be reported.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method is a constructive learned-sampling pipeline evaluated against external baselines; the main weakness is unverified generalization to held-out constraints, not circular reasoning.

full rationale

DANCE's derivation chain is constructive rather than circular. The learned distribution p(A|D,C) in Eq. 2 is a trainable conditional sampler, and the SelectGate in Eq. 3 is a concrete MLP with Gumbel-Softmax and Bernoulli sampling under resource constraints. The accuracy claims are empirical comparisons (Table 1, Figures 3-4) against external baselines such as OFA, ProxylessNAS, RecNAS, SPOS, and FixMatch, plus a score-based pruning baseline; they are not algebraic consequences of the training objective. The importance metrics in Appendix A are explicit, independently motivated feature statistics, and the ablation in Table 2 tests their individual contributions rather than assuming them. Self-citations (e.g., DNS-Rec [Zhang et al., 2024]) appear only as background or related work and are not used to justify a uniqueness theorem or to forbid alternative explanations. The genuine weakness is evidential rather than circular: Stage 2 trains only with layer-wise constraints C_l in {0.3, 0.4, 0.5} (Appendix C), while Section 3.2 promises a fine-grained 0.02-step sweep from 0.1 to 0.5 that never appears, and Sections 2.4 and 3.2 assert O(1) adaptation to arbitrary budgets without a held-out-constraint experiment. That is a missing-support or correctness concern, not a circular reduction, so the circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

The central machinery rests on learned gates, importance scores with fitted weights, and a factorization assumption; none of these are derived from first principles, and several are trained on the same data used for evaluation.

free parameters (4)

lambda_1..lambda_4 = not fully reported; sensitivity only for lambda_2
Weights in Score(A_l) combining static, dynamic, feature, and correlation importance. Chosen or learned, with no final values or fitting procedure reported.
alpha (EMA decay) = sensitivity in Fig 5b, value not fixed in main text
Controls update speed of dynamic and feature importance.
temperature tau = chosen from {0.1, 0.5, 1.0}
Gumbel-softmax temperature balancing exploration and sharpness.
layer-wise constraint C_l = selected from {0.3, 0.4, 0.5}
Sparsity/retention target per layer; the paper claims smooth adaptation to other values but does not train on them.

assumptions (4)

domain assumption Layer-wise factorization of architecture distribution p(A|D,C) = product_l p(g_l | F_l, D, C)
Assumes gates are conditionally independent across layers given features, data, and constraints; cross-layer dependencies are ignored (Eq. 2).
standard math Gumbel-softmax provides a valid differentiable relaxation of discrete sampling
Used in Section 2.2 to backpropagate through BernoulliSample with exactly C_l components; the hard sampling itself is not differentiable.
ad hoc to paper The four importance metrics (static, dynamic, feature, correlation) are sufficient to rank architectural components
Appendix A introduces these metrics as 'theoretically grounded' but provides no proof or external validation that they capture component value.
domain assumption SuperNet pre-training transfers to arbitrarily sampled sub-networks
Stage 1 uses progressive path dropping and distillation, but transferability to unseen masks is assumed, not proven.

invented entities (2)

SelectGate module
purpose: Learns to generate layer-wise binary masks conditioned on data batch, layer features, and resource constraint.
A new architectural module introduced in this paper; no independent falsifiable evidence outside the reported experiments.
Continuous architecture distribution p(A|D,C)
purpose: Represented as a product of learned gate distributions to allow O(1) sampling for new scenarios.
A modeling construct; its generalization to unseen constraints is asserted but not demonstrated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of DANCE: Resource-Efficient Neural Architecture Search with Data-Aware and Continuous Adaptation." pith.science (2026). https://pith.science/paper/4BMSCN3O

@misc{pith2026250704671,
  author       = {Pith},
  title        = {Pith review of: DANCE: Resource-Efficient Neural Architecture Search with Data-Aware and Continuous Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4BMSCN3O}},
  note         = {Machine review of arXiv:2507.04671}
}

read the original abstract

Neural Architecture Search (NAS) has emerged as a powerful approach for automating neural network design. However, existing NAS methods face critical limitations in real-world deployments: architectures lack adaptability across scenarios, each deployment context requires costly separate searches, and performance consistency across diverse platforms remains challenging. We propose DANCE (Dynamic Architectures with Neural Continuous Evolution), which reformulates architecture search as a continuous evolution problem through learning distributions over architectural components. DANCE introduces three key innovations: a continuous architecture distribution enabling smooth adaptation, a unified architecture space with learned selection gates for efficient sampling, and a multi-stage training strategy for effective deployment optimization. Extensive experiments across five datasets demonstrate DANCE's effectiveness. Our method consistently outperforms state-of-the-art NAS approaches in terms of accuracy while significantly reducing search costs. Under varying computational constraints, DANCE maintains robust performance while smoothly adapting architectures to different hardware requirements. The code and appendix can be found at https://github.com/Applied-Machine-Learning-Lab/DANCE.

Figures

Figures reproduced from arXiv: 2507.04671 by the authors.

**Figure 2.** Overview of DANCE’s three-stage training process. Stage 1 pre-trains a SuperNet with N stacked layers as backbone. Stage 2 [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Distribution of FLOPs vs. Accuracy across different prun [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Performance comparison between DANCE and Score [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: Parameter sensitivity analysis C. More Details about Implementation Details When implementing the training process, we adopt a threestage approach with carefully selected hyperparameters: Stage 1 (Representation Learning, 100 epochs) focuses on pre-training the SuperN…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 35 canonical work pages

[1]

Un- derstanding and simplifying one-shot architecture search

[Bender et al., 2018] Gabriel Bender, Pieter-Jan Kinder- mans, Barret Zoph, Vijay Vasudevan, and Quoc Le. Un- derstanding and simplifying one-shot architecture search. In Proc. of ICML, pages 550–559,

work page 2018
[3]

Proxylessnas: Direct neural architecture search on tar- get task and hardware

[Cai et al., 2018b] Han Cai, Ligeng Zhu, and Song Han. Proxylessnas: Direct neural architecture search on tar- get task and hardware. arXiv preprint arXiv:1812.00332 ,

arXiv
[4]

Once-for-all: Train one network and specialize it for efficient deployment

[Cai et al., 2019] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment. arXiv preprint arXiv:1908.09791,

arXiv 2019
[9]

Autotrans- fer: Instance transfer for cross-domain recommendations

[Gao et al., 2023] Jingtong Gao, Xiangyu Zhao, Bo Chen, Fan Yan, Huifeng Guo, and Ruiming Tang. Autotrans- fer: Instance transfer for cross-domain recommendations. In Proc. of SIGIR, pages 1478–1487,

work page 2023
[10]

Single path one-shot neural architecture search with uni- form sampling

[Guo et al., 2020] Zichao Guo, Xiangyu Zhang, Haoyuan Mu, Wen Heng, Zechun Liu, Yichen Wei, and Jian Sun. Single path one-shot neural architecture search with uni- form sampling. In Proc. of ECCV, pages 544–560,

work page 2020
[14]

Categorical reparameterization with gumbel-softmax

[Jang et al., 2016] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144,

arXiv 2016
[15]

Automated Self-Supervised Learning for Graphs

[Jin et al., 2021] Wei Jin, Xiaorui Liu, Xiangyu Zhao, Yao Ma, Neil Shah, and Jiliang Tang. Automated self-supervised learning for graphs. arXiv preprint arXiv:2106.05470,

work page Pith review arXiv 2021
[17]

Learning multiple layers of features from tiny im- ages

[Krizhevsky et al., 2009] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny im- ages

2009

Show all 44 references

[19]

Adafs: Adaptive feature selection in deep recommender system

[Lin et al., 2022] Weilin Lin, Xiangyu Zhao, Yejing Wang, Tong Xu, and Xian Wu. Adafs: Adaptive feature selection in deep recommender system. In Proc. of KDD , pages 3309–3317,

2022
[20]

Progres- sive neural architecture search

[Liu et al., 2018] Chenxi Liu, Barret Zoph, Maxim Neu- mann, Jonathon Shlens, Wei Hua, Li-Jia Li, Li Fei-Fei, Alan Yuille, Jonathan Huang, and Kevin Murphy. Progres- sive neural architecture search. In Proc. of ECCV , pages 19–34,

2018
[21]

Automated embed- ding size search in deep recommender systems

[Liu et al., 2020] Haochen Liu, Xiangyu Zhao, Chong Wang, Xiaobing Liu, and Jiliang Tang. Automated embed- ding size search in deep recommender systems. In Proc. of SIGIR, pages 2307–2316,

2020
[22]

Autoassign+: Automatic shared embedding assign- ment in streaming recommendation

[Liu et al., 2024] Ziru Liu, Kecheng Chen, Fengyi Song, Bo Chen, Xiangyu Zhao, Huifeng Guo, and Ruiming Tang. Autoassign+: Automatic shared embedding assign- ment in streaming recommendation. Knowledge and In- formation Systems, pages 89–113,

2024
[24]

Efficient neural architecture search via parameters sharing

[Pham et al., 2018] Hieu Pham, Melody Guan, Barret Zoph, Quoc Le, and Jeff Dean. Efficient neural architecture search via parameters sharing. In Proc. of ICML , pages 4095–4104,

2018
[25]

Regularized evolution for im- age classifier architecture search

[Real et al., 2019] Esteban Real, Alok Aggarwal, Yanping Huang, and Quoc V Le. Regularized evolution for im- age classifier architecture search. In Proc. of AAAI, pages 4780–4789,

2019
[26]

A comprehensive survey of neural architecture search: Challenges and solutions

[Ren et al., 2021] Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Xiaojiang Chen, and Xin Wang. A comprehensive survey of neural architecture search: Challenges and solutions. ACM Computing Sur- veys (CSUR), pages 1–34,

2021
[27]

Very deep convolu- tional networks for large-scale image recognition

[Simonyan, 2014] Karen Simonyan. Very deep convolu- tional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556,

2014 arXiv
[29]

Autoassign: Automatic shared embedding assignment in streaming recommenda- tion

[Song et al., 2022] Fengyi Song, Bo Chen, Xiangyu Zhao, Huifeng Guo, and Ruiming Tang. Autoassign: Automatic shared embedding assignment in streaming recommenda- tion. In Proc. of ICDM, pages 458–467,

2022
[30]

Efficient- net: Rethinking model scaling for convolutional neural networks

[Tan and Le, 2019] Mingxing Tan and Quoc Le. Efficient- net: Rethinking model scaling for convolutional neural networks. In Proc. of ICML, pages 6105–6114,

2019
[31]

The caltech- ucsd birds-200-2011 dataset

[Wah et al., 2011] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech- ucsd birds-200-2011 dataset

2011
[33]

Rethink- ing architecture selection in differentiable nas

[Wang et al., 2021] Ruochen Wang, Minhao Cheng, Xiangn- ing Chen, Xiaocheng Tang, and Cho-Jui Hsieh. Rethink- ing architecture selection in differentiable nas. arXiv preprint arXiv:2108.04392,

2021 arXiv
[34]

Fbnet: Hardware-aware efficient convnet design via differentiable neural architecture search

[Wu et al., 2019] Bichen Wu, Xiaoliang Dai, Peizhao Zhang, Yanghan Wang, Fei Sun, Yiming Wu, Yuandong Tian, Peter Vajda, Yangqing Jia, and Kurt Keutzer. Fbnet: Hardware-aware efficient convnet design via differentiable neural architecture search. In Proc. of CVPR , pages 10734–10742,

2019
[35]

Ef- ficient resource-aware neural architecture search with dy- namic adaptive network sampling

[Yang and Sun, 2021] Zhao Yang and Qingshuang Sun. Ef- ficient resource-aware neural architecture search with dy- namic adaptive network sampling. In 2021 IEEE Interna- tional Symposium on Circuits and Systems (ISCAS) , pages 1–5,

2021
[36]

Autostl: Automated spatio-temporal multi-task learning

[Zhang et al., 2023] Zijian Zhang, Xiangyu Zhao, Hao Miao, Chunxu Zhang, Hongwei Zhao, and Junbo Zhang. Autostl: Automated spatio-temporal multi-task learning. In Proc. of AAAI, pages 4902–4910,

2023
[37]

Dns-rec: Data- aware neural architecture search for recommender sys- tems

[Zhang et al., 2024] Sheng Zhang, Maolin Wang, Xiangyu Zhao, Ruocheng Guo, Yao Zhao, Chenyi Zhuang, Jin- jie Gu, Zijian Zhang, and Hongzhi Yin. Dns-rec: Data- aware neural architecture search for recommender sys- tems. In Proceedings of the 18th ACM Conference on Rec- ommender...

2024
[38]

Adaptive and automated deep recommender systems

[Zhao, 2022] Xiangyu Zhao. Adaptive and automated deep recommender systems. ACM SIGWEB Newsletter , pages 1–4,

2022
[39]

Au- toemb: Automated embedding dimensionality search in streaming recommendations

[Zhaok et al., 2021] Xiangyu Zhaok, Haochen Liu, Wenqi Fan, Hui Liu, Jiliang Tang, Chong Wang, Ming Chen, Xudong Zheng, Xiaobing Liu, and Xiwang Yang. Au- toemb: Automated embedding dimensionality search in streaming recommendations. In Proc. of ICDM , pages 896–905,

2021
[40]

Autogen: An automated dy- namic model generation framework for recommender sys- tem

[Zhu et al., 2023] Chenxu Zhu, Bo Chen, Huifeng Guo, Hang Xu, Xiangyang Li, Xiangyu Zhao, Weinan Zhang, Yong Yu, and Ruiming Tang. Autogen: An automated dy- namic model generation framework for recommender sys- tem. In Proc. of WSDM, pages 598–606,

2023
[41]

Semi-supervised learning lit- erature survey

[Zhu, 2005] Xiaojin Jerry Zhu. Semi-supervised learning lit- erature survey

2005
[43]

Details of Component Scoring Mechanism The Score(Al) function establishes a theoretically grounded framework for evaluating architectural components

Appendix of DANCE A. Details of Component Scoring Mechanism The Score(Al) function establishes a theoretically grounded framework for evaluating architectural components. Our key insight is that component importance should be assessed from multiple complementary perspectives, ...

2016
[44]

CIFAR-10/100 [Krizhevsky et al

and 2 back- bones (ResNet-18 [He et al., 2016] and VGG-16 [Simonyan, 2014]). CIFAR-10/100 [Krizhevsky et al. , 2009 ]: A basic com- puter vision dataset with small (32×32) images. CIFAR-10 has 10 classes with 6,000 images each, while CIFAR-100 has 100 classes with 600 images e...

2016
[2005]

Neural architecture search with re- inforcement learning

[Zoph, 2016] B Zoph. Neural architecture search with re- inforcement learning. arXiv preprint arXiv:1611.01578 ,

2016 arXiv
[2009]

Automlp: Automated mlp for sequential recommen- dations

[Li et al., 2023] Muyang Li, Zijian Zhang, Xiangyu Zhao, Wanyu Wang, Minghao Zhao, Runze Wu, and Ruocheng Guo. Automlp: Automated mlp for sequential recommen- dations. In Proceedings of the ACM web conference 2023, pages 1190–1198,

2023
[2011]

Haq: Hardware-aware automated quantization with mixed precision

[Wang et al., 2019] Kuan Wang, Zhijian Liu, Yujun Lin, Ji Lin, and Song Han. Haq: Hardware-aware automated quantization with mixed precision. In Proc. of CVPR , pages 8612–8620,

2019
[2014]

Fixmatch: Simplifying semi-supervised learning with con- sistency and confidence

[Sohn et al., 2020] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with con- sistency and confidence. Proc. of NeurIPS , pages 596– 608,

2020
[2015]

Deep residual learning for image recog- nition

[He et al., 2016] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog- nition. In Proc. of CVPR, pages 770–778,

2016
[2016]

Mobilenets: Efficient convolutional neural networks for mobile vision applica- tions

[Howard, 2017] Andrew G Howard. Mobilenets: Efficient convolutional neural networks for mobile vision applica- tions. arXiv preprint arXiv:1704.04861,

2017 arXiv
[2017]

Neural architecture search: A survey

[Elsken et al., 2019] Thomas Elsken, Jan Hendrik Metzen, and Frank Hutter. Neural architecture search: A survey. Journal of Machine Learning Research, pages 1–21,

2019
[2018]

Food-101–mining discriminative com- ponents with random forests

[Bossard et al., 2014] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101–mining discriminative com- ponents with random forests. In Proc. of ECCV , pages 446–461,

2014
[2019]

Automated machine learning for deep recommender systems: A sur- vey

[Chen et al., 2022] Bo Chen, Xiangyu Zhao, Yejing Wang, Wenqi Fan, Huifeng Guo, and Ruiming Tang. Automated machine learning for deep recommender systems: A sur- vey. arXiv preprint arXiv:2204.01390,

2022 arXiv
[2020]

Learning both weights and connections for efficient neural network

[Han et al., 2015] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. Proc. of NeurIPS,

2015
[2021]

Lsun-stanford car dataset: enhancing large-scale car image datasets using deep learning for usage in gan training

[Kramberger and Potoˇcnik, 2020] Tin Kramberger and Boˇzidar Poto ˇcnik. Lsun-stanford car dataset: enhancing large-scale car image datasets using deep learning for usage in gan training. Applied Sciences, page 4913,

2020
[2022]

Review of one-shot neural architec- ture search

[DONG et al., 2023] Pei-jie DONG, Xin NIU, Zi-mian WEI, and Xue-hui CHEN. Review of one-shot neural architec- ture search. Computer Engineering & Science , page 191,

2023
[2023]

Scalable multitask policy gradient reinforcement learning

[El Bsat et al., 2017] Salam El Bsat, Haitham Bou Ammar, and Matthew Taylor. Scalable multitask policy gradient reinforcement learning. In Proc. of AAAI,

2017
[2024]

Recnas: Resource-constrained neural architecture search based on differentiable anneal- ing and dynamic pruning

[Peng et al., 2022] Cheng Peng, Yangyang Li, Ronghua Shang, and Licheng Jiao. Recnas: Resource-constrained neural architecture search based on differentiable anneal- ing and dynamic pruning. IEEE Transactions on Neural Networks and Learning Systems , pages 2805–2819,

2022

Pith tools

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

[1] [1]

Un- derstanding and simplifying one-shot architecture search

[Bender et al., 2018] Gabriel Bender, Pieter-Jan Kinder- mans, Barret Zoph, Vijay Vasudevan, and Quoc Le. Un- derstanding and simplifying one-shot architecture search. In Proc. of ICML, pages 550–559,

work page 2018

[2] [3]

Proxylessnas: Direct neural architecture search on tar- get task and hardware

[Cai et al., 2018b] Han Cai, Ligeng Zhu, and Song Han. Proxylessnas: Direct neural architecture search on tar- get task and hardware. arXiv preprint arXiv:1812.00332 ,

arXiv

[3] [4]

Once-for-all: Train one network and specialize it for efficient deployment

[Cai et al., 2019] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment. arXiv preprint arXiv:1908.09791,

arXiv 2019

[4] [9]

Autotrans- fer: Instance transfer for cross-domain recommendations

[Gao et al., 2023] Jingtong Gao, Xiangyu Zhao, Bo Chen, Fan Yan, Huifeng Guo, and Ruiming Tang. Autotrans- fer: Instance transfer for cross-domain recommendations. In Proc. of SIGIR, pages 1478–1487,

work page 2023

[5] [10]

Single path one-shot neural architecture search with uni- form sampling

[Guo et al., 2020] Zichao Guo, Xiangyu Zhang, Haoyuan Mu, Wen Heng, Zechun Liu, Yichen Wei, and Jian Sun. Single path one-shot neural architecture search with uni- form sampling. In Proc. of ECCV, pages 544–560,

work page 2020

[6] [14]

Categorical reparameterization with gumbel-softmax

[Jang et al., 2016] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144,

arXiv 2016

[7] [15]

Automated Self-Supervised Learning for Graphs

[Jin et al., 2021] Wei Jin, Xiaorui Liu, Xiangyu Zhao, Yao Ma, Neil Shah, and Jiliang Tang. Automated self-supervised learning for graphs. arXiv preprint arXiv:2106.05470,

work page Pith review arXiv 2021

[8] [17]

Learning multiple layers of features from tiny im- ages

[Krizhevsky et al., 2009] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny im- ages

2009