Pith. sign in

REVIEW 3 major objections 3 minor 300 references

In graph regression tasks, a deep convolutional GNN layer (GEN) outperforms attention-based layers across a broad battery of benchmarks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 16:37 UTC pith:XCP752IS

load-bearing objection A useful MP-layer benchmark for GNN regression, but the abstract overstates GEN's edge: their own post-hoc tests don't separate GEN from several classical and attention layers. the 3 major comments →

arxiv 2607.26404 v1 pith:XCP752IS submitted 2026-07-29 cs.LG

Examining the Efficacy of Graph Neural Network Message-Passing in Regression Contexts

Keith G. Mills , Aedan J. DeFrates , Joong Ho Kim This is my paper
classification cs.LG
keywords graph neural networksmessage passingregressionneural architecture searchperformance predictionDeeperGCNrank correlationgraph transformers
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Graph neural networks are usually judged on classification benchmarks, yet they are also widely used for scalar regression—predicting numbers such as the accuracy of a neural network architecture. This paper asks whether the choice of message-passing layer matters as much in regression as in classification, and it answers by swapping eleven different GNN layers into four existing regression predictors while changing nothing else. The paper's central result is that a deep convolutional layer called GEN (DeeperGCN) is the most consistently effective choice, with the lowest average rank and the most concentrated top-rank distribution across dozens of trials. Attention-based layers like GATv2 and graph transformers are competitive but are outperformed overall, while classical layers such as k-GNN and PNA offer similar performance at much lower computational cost. The findings matter because regression practitioners often default to attention GNNs from classification; this paper argues that a deep convolution is a stronger default for graph regression.

Core claim

The paper's central claim: deep convolutional message passing—specifically the GEN (DeeperGCN) layer—is the best overall GNN for graph-level regression on neural-architecture-search and diffusion-model datasets. In an aggregate ranking over all experiments, GEN has the lowest mean rank and the most concentrated top-rank distribution, winning more trials than any other layer. A Friedman test rejects equal ranks; a Nemenyi test shows GEN significantly beats MLP, GCN, GIN, and both GATv2 variants. Attention GNNs show no clear edge: they cluster with classical layers k-GNN and PNA, while vanilla GCN and MLP underperform. On graphs over 1,000 nodes, multi-headed GATv2 excels, but deep convolution

What carries the argument

The load-bearing object is the message-passing (MP) layer—the component that defines how a node aggregates information from its neighbors. The protocol is a controlled swap: eleven MP mechanisms (GCN, GraphSAGE, GIN, GEN, k-GNN, PNA, GATv2 with 1 or 4 heads, graph transformer with 1 or 4 heads, plus an MLP baseline) are inserted one at a time into four existing regression predictors, keeping all other training hyperparameters fixed. This isolates the effect of MP design from the rest of the predictor. The central finding is that GEN—a deep convolutional layer designed for stable training at greater depth—is the most robust to this substitution, suggesting its residual connections and learnab

Load-bearing premise

The ranking assumes that keeping every training hyperparameter fixed at the values chosen for the original GNN is a fair way to compare layers; if GEN simply tolerates those settings better, its top rank may be a protocol artifact.

What would settle it

Run a full hyperparameter sweep for each of the eleven GNN layers on one or more of the same datasets (e.g., NAS-Bench-201) and re-rank; if GEN no longer tops the ranking when the others are tuned, the paper's central claim is an artifact. Alternatively, apply the same fixed-hyperparameter swap to a non-DNN regression benchmark such as a molecular dataset; if GEN does not stand out there, the result is specific to network-architecture graphs.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Regression-focused GNN papers should include deep convolutional layers like GEN as a baseline before adopting attention-based mechanisms.
  • The choice of message-passing layer can change performance dramatically on regression benchmarks, so layer ablation should become a standard step in neural-architecture-search predictor development.
  • Classical and theoretical layers (k-GNN, PNA) match attention GNNs on many regression tasks while using substantially less memory and inference time.
  • For graphs with more than 1,000 nodes, multi-headed graph attention (GATv2-H4) is the standout layer, indicating a size-dependent role for attention.
  • The custom attention layers used by the original regression predictors are not always worth their complexity: they excel mainly at low training-data regimes, while simpler layers catch up or surpass them as data grows.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the paper keeps hyperparameters fixed at values chosen for the original GNN, the comparison likely understates the performance of non-original layers; a per-layer hyperparameter search could narrow GEN's lead.
  • A natural extension would be to apply the same swap protocol to molecular or material graph-regression benchmarks to test whether GEN's robustness generalizes beyond DNN-graph domains.
  • The paper's protocol could be turned into a cheap 'layer sanity check' for any new GNN regressor: swap in GEN and k-GNN and compare—if the new layer cannot beat them, the contribution may be elsewhere.
  • The large-graph finding for GATv2 suggests a hybrid design—convolution for small/medium graphs, attention for large ones—could outperform any single layer, though the paper does not test this.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. This paper is an empirical benchmark of GNN message-passing layers for graph-level regression. The authors take four existing GNN regressors (FlowerFormer, PINAT, AutoBuild, Qua2SeDiMo), swap the MP layer with several standard alternatives plus an MLP baseline, keep all other training choices fixed, and evaluate on NAS and diffusion-model quantization datasets across KT, SRCC, MAE, NDCG, and related metrics. The headline result is Fig. 1: GEN (DeeperGCN) has the lowest mean rank and the most concentrated rank-1 distribution, leading the authors to conclude that deep convolutional MP is more effective than attention-based GNNs for regression. The paper also reports a Friedman test, Nemenyi post-hoc p-values, and hardware latency/memory comparisons.

Significance. Regression is an underexplored benchmark setting for GNN layer selection, and this paper provides one of the broader comparisons to date, with multi-repository coverage, 5-seed error bars, a statistical significance analysis, and hardware cost measurements. If the central ranking were robust, the practical guidance would be useful: GEN as a strong default, k-GNN and PNA as efficient competitive alternatives, and attention layers valuable on very large graphs. However, the current protocol has confounds that prevent clean mechanism-level attribution, and some summary claims exceed the statistical and per-dataset evidence.

major comments (3)
  1. [Supplementary Materials, 'Regressor Hyperparameters'; Table 8] The protocol states: 'We do not change these hyperparameters, and instead simply change the form of GNN MP mechanism utilized.' Because each repository's learning rate, batch size, embedding dimension, and training length were chosen for its native layer (DAGFormer, GATSet, or the AutoBuild default), every swapped-in layer is evaluated under another method's training recipe. GEN's rank-1 concentration may therefore reflect robustness to one fixed recipe rather than the superiority of deep convolutional message passing. Table 8 adds a parameter-count confound: on AutoBuild NB-101, GEN has 18.4k parameters vs 4.7k for GCN and 5.6k for GIN. This is load-bearing for the central claim. I recommend at least a sensitivity analysis with per-layer tuned hyperparameters, or matching parameter budgets, before attributing the ranking to the MP mechanism.
  2. [Abstract and Conclusion vs. Table 6 and Figure 4] The abstract and conclusion state that attention-based GNNs are 'significantly outperformed' by GEN. The statistical evidence in Fig. 4 does not support this globally: the Nemenyi post-hoc p-values show GEN is not significantly distinguishable from GraphSAGE (p=0.48), k-GNN (p=0.63), PNA (p=0.36), GT-H1 (p=0.09), or GT-H4 (p=0.28). Table 6 also shows GATv2-H4 as the best method on SDXL across SRCC, NDCG, KT, and MAE, and Table 2 shows DAGFormer winning several FlowerFormer trials. The accurate conclusion is that GEN is top-ranked in aggregate, not that it uniformly dominates attention-based GNNs.
  3. [Figure 1 and Tables 2-3] The overall ranking in Fig. 1 covers only the 11 swap-in methods (MLP through GT-H4) and excludes the native DAGFormer (FlowerFormer) and GATSet (PINAT) layers, even though those layers are reported in Tables 2 and 3 and are competitive on several benchmarks (e.g., DAGFormer is best on NB-101, NB-201, and NB-Graph at low data fractions). The 'overall best GNN' claim is therefore limited to the swap-in set, and the aggregate ranking could change if native layers were included. Please state this scope in the abstract/conclusion.
minor comments (3)
  1. [Results on AutoBuild, first paragraph] The text says 'Immediately, three types of GNNs stand out: Deep Graph Convolutional layers (GEN), k-GNN and multi-headed Graph Transformers (GT-H4). These are the only two GNN types...' — the 'three'/'two' mismatch should be corrected.
  2. [Eq. (5)] In the PINAT loss, the notation I(i) is introduced without definition. Please clarify what I(i) denotes.
  3. [Table 6 / SDXL row] The MLP result on SDXL is NaN; the main text should state this explicitly rather than only in the supplementary Table 24 note.

Circularity Check

0 steps flagged

No circularity: the GEN ranking is an empirical benchmark outcome, not a fitted, self-referential, or citation-derived prediction.

full rationale

The paper's central claim is an empirical ranking produced by running fixed GNN layer implementations across several regression repositories; there is no derivation chain in which a predicted quantity is defined in terms of the fitted quantity, and no parameter is fit to a subset and then 'predicted' on the same target. The GNN layers are external, plug-and-play implementations from PyTorch Geometric, and no Ansatz or uniqueness theorem is imported from the authors' prior work to force the outcome. The authors do cite their own prior works (AutoBuild, Qua2SeDiMo) as two of the four benchmark environments, but these citations supply datasets and repositories, not the conclusion that GEN is best; the conclusion is obtained by running those benchmarks, not by invoking the prior papers' claims. The paper itself flags the main limitation in the Supplementary Materials: 'We do not change these hyperparameters, and instead simply change the form of GNN MP mechanism utilized.' This is a genuine protocol-fairness confound—all swapped-in layers run at hyperparameters chosen for the original repository layers—and Table 8 shows unequal parameter counts across GNNs (e.g., GEN 18.4k vs GCN 4.7k on AutoBuild NB-101). These are validity threats to the causal attribution of GEN's success to its message-passing mechanism, but they are not circularity: the measured rankings could change under a fairer protocol without any step reducing to the paper's own inputs by construction. No self-justifying equation, no fitted-quantity-renamed-as-prediction, and no load-bearing self-citation chain were found. Therefore the appropriate circularity score is 0.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 0 invented entities

The central claim is an empirical ranking; it depends on the evaluation protocol. The main protocol-level free choice is per-repository hyperparameters taken from source papers; no invented entities, forces, or fitted constants are introduced.

free parameters (1)
  • Per-repository training hyperparameters (epochs, learning rate, batch size, embedding dimension) = NB-101 FlowerFormer: 300 eps, 8e-4, batch 128, d=64; PINAT: 300 eps, 1e-4, batch 10, d=256; AutoBuild: 200 eps, 1e-4, ba
    Chosen by original repositories, not tuned for each swapped GNN; the comparison's fairness depends on these fixed values.
axioms (4)
  • domain assumption Swapping only the message-passing layer while keeping every other component fixed isolates the effect of layer choice.
    Stated as the controlled variable in the introduction and repeated in the supplementary; if hyperparameters are tuned for the original layer, this fails.
  • domain assumption The four repositories' tasks (NAS accuracy prediction and diffusion-model quantization FID) are representative of graph regression.
    All datasets in Table 1 are DNN-architecture graphs; no molecule or physical-graph datasets appear. The abstract calls this 'a slew of regression contexts'.
  • domain assumption PyTorch-Geometric's implementations of GCN/SAGE/GIN/GEN/k-GNN/PNA/GATv2/GT are fair and comparable implementations.
    All non-custom layers are taken 'plug-and-play from PyTorch-Geometric'; implementation quality is assumed uniform.
  • domain assumption Friedman/Nemenyi tests treat each dataset/metric/train-fraction trial as an exchangeable observation.
    Used in the Statistical Significance section; trials share data and are not independent replicates, which may inflate significance.

pith-pipeline@v1.3.0-daily-deepseek · 33391 in / 11260 out tokens · 106911 ms · 2026-08-01T16:37:40.746133+00:00 · methodology

0 comments
read the original abstract

Graph Neural Networks (GNN) facilitate effective prediction on graph data such as molecules, media networks and neural network blueprints. GNNs facilitate prediction through message passing techniques which define how information flows from a node to its neighbors. Due to the ubiquity of the graph data type, the development of newer and better GNNs has garnered much interest in the machine learning community. However, GNN evaluation and benchmarking is primarily driven by classification tasks. Thus, prospective GNN message passing layers are evaluated on their ability to outperform prior work in classification contexts. In contrast, GNNs are equally capable of performing scalar regression prediction, yet this class of problem is often overlooked when proposing new GNNs while the best classification GNNs are utilized in an a priori or off-the-shelf manner for regression problems. In response, this paper studies the efficacy of GNN layers in a slew of regression contexts from rank ordering, error minimization and insight extraction. Results show that deep convolutional GNNs, particularly GEN, are more effective at these tasks than attention-based GNNs, while other classical, theoretically-inspired GNNs remain competitive and efficient.

Figures

Figures reproduced from arXiv: 2607.26404 by Aedan J. DeFrates, Joong Ho Kim, Keith G. Mills.

Figure 1
Figure 1. Figure 1: Violin plot of the overall GNN ranking distributions across experiments in the main body of this manuscript (primary [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Latency vs. KT plots for FlowerFormer NB-101 5% [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Latency vs. Hop SRCC for AutoBuild NB-101. Icon [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Pairwise Nemenyi post-hoc p-values for the Fried [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

300 extracted references · 4 canonical work pages · 1 internal anchor

  1. [1]

    , author=

    Neural architecture search: A survey. , author=. J. Mach. Learn. Res. , volume=

  2. [2]

    Gonzalez , title =

    Alvin Wan and Xiaoliang Dai and Peizhao Zhang and Zijian He and Yuandong Tian and Saining Xie and Bichen Wu and Matthew Yu and Tao Xu and Kan Chen and Peter Vajda and Joseph E. Gonzalez , title =. 2020. 2020 , doi =

    2020

  3. [3]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Mobilenetv2: Inverted residuals and linear bottlenecks , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  4. [4]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops , pages=

    Fast hardware-aware neural architecture search , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops , pages=

  5. [5]

    arXiv preprint arXiv:1707.06347 , year=

    Proximal policy optimization algorithms , author=. arXiv preprint arXiv:1707.06347 , year=

    Pith/arXiv arXiv

  6. [6]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Weisfeiler and leman go neural: Higher-order graph neural networks , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  7. [7]

    arXiv preprint arXiv:1412.3555 , year=

    Empirical evaluation of gated recurrent neural networks on sequence modeling , author=. arXiv preprint arXiv:1412.3555 , year=

    Pith/arXiv arXiv

  8. [8]

    ICML deep learning workshop , volume=

    Siamese neural networks for one-shot image recognition , author=. ICML deep learning workshop , volume=. 2015 , organization=

    2015

  9. [9]

    arXiv preprint arXiv:1807.00734 , year=

    The relativistic discriminator: a key element missing from standard GAN , author=. arXiv preprint arXiv:1807.00734 , year=

    Pith/arXiv arXiv

  10. [10]

    2012 , publisher=

    Machine learning: a probabilistic perspective , author=. 2012 , publisher=

    2012

  11. [11]

    2013 , publisher=

    The cross-entropy method: a unified approach to combinatorial optimization, Monte-Carlo simulation and machine learning , author=. 2013 , publisher=

    2013

  12. [12]

    2002 , publisher=

    Rare event estimation for static models via cross-entropy and importance sampling , author=. 2002 , publisher=

    2002

  13. [13]

    Annals of Operations Research , volume=

    Probabilistic bounds (via large deviations) for the solutions of stochastic programming problems , author=. Annals of Operations Research , volume=. 1995 , publisher=

    1995

  14. [14]

    1993 , publisher=

    Discrete event systems: Sensitivity analysis and stochastic optimization by the score function method , author=. 1993 , publisher=

    1993

  15. [15]

    2014 , publisher=

    Lectures on stochastic programming: modeling and theory , author=. 2014 , publisher=

    2014

  16. [16]

    SIAM journal on optimization , volume=

    On the rate of convergence of optimal solutions of Monte Carlo approximations of stochastic programs , author=. SIAM journal on optimization , volume=. 2000 , publisher=

    2000

  17. [17]

    Advances in Neural Information Processing Systems , pages=

    D-VAE: A variational autoencoder for directed acyclic graphs , author=. Advances in Neural Information Processing Systems , pages=

  18. [18]

    Advances in neural information processing systems , pages=

    Generative adversarial nets , author=. Advances in neural information processing systems , pages=

  19. [19]

    Advances in neural information processing systems , pages=

    f-gan: Training generative neural samplers using variational divergence minimization , author=. Advances in neural information processing systems , pages=

  20. [20]

    Advances in neural information processing systems , pages=

    Graph convolutional policy network for goal-directed molecular graph generation , author=. Advances in neural information processing systems , pages=

  21. [21]

    arXiv preprint arXiv:1802.08773 , year=

    Graphrnn: Generating realistic graphs with deep auto-regressive models , author=. arXiv preprint arXiv:1802.08773 , year=

    Pith/arXiv arXiv

  22. [22]

    Advances in neural information processing systems , pages=

    Neural architecture optimization , author=. Advances in neural information processing systems , pages=

  23. [23]

    2019 , eprint=

    Balanced One-shot Neural Architecture Optimization , author=. 2019 , eprint=

    2019

  24. [24]

    ICLR , year =

    Hanxiao Liu and Karen Simonyan and Yiming Yang , title =. ICLR , year =

  25. [25]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages=

    Mnasnet: Platform-aware neural architecture search for mobile , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages=

  26. [26]

    arXiv preprint arXiv:2001.08437 , year=

    Multi-objective Neural Architecture Search via Non-stationary Policy Gradient , author=. arXiv preprint arXiv:2001.08437 , year=

    Pith/arXiv arXiv 2001

  27. [27]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages=

    Fbnet: Hardware-aware efficient convnet design via differentiable neural architecture search , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages=

  28. [28]

    International Conference on Learning Representations , year=

    PC-DARTS: Partial Channel Connections for Memory-Efficient Architecture Search , author=. International Conference on Learning Representations , year=

  29. [29]

    Proceedings of the IEEE International Conference on Computer Vision , pages=

    Progressive differentiable architecture search: Bridging the depth gap between search and evaluation , author=. Proceedings of the IEEE International Conference on Computer Vision , pages=

  30. [30]

    2019 , eprint=

    RC-DARTS: Resource Constrained Differentiable Architecture Search , author=. 2019 , eprint=

    2019

  31. [31]

    Proceedings of the IEEE Conference on computer vision and pattern recognition , pages=

    Searching for a robust neural architecture in four gpu hours , author=. Proceedings of the IEEE Conference on computer vision and pattern recognition , pages=

  32. [32]

    7th International Conference on Learning Representations,

    Sirui Xie and Hehui Zheng and Chunxiao Liu and Liang Lin , title =. 7th International Conference on Learning Representations,. 2019 , url =

    2019

  33. [33]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Hournas: Extremely fast neural architecture search through an hourglass lens , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  34. [34]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Learning transferable architectures for scalable image recognition , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  35. [35]

    Proceedings of the aaai conference on artificial intelligence , volume=

    Regularized evolution for image classifier architecture search , author=. Proceedings of the aaai conference on artificial intelligence , volume=

  36. [36]

    ICML , pages=

    Nas-bench-101: Towards reproducible neural architecture search , author=. ICML , pages=

  37. [37]

    International Conference on Learning Representations , year=

    NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search , author=. International Conference on Learning Representations , year=

  38. [38]

    Uncertainty in Artificial Intelligence , pages=

    Random search and reproducibility for neural architecture search , author=. Uncertainty in Artificial Intelligence , pages=. 2020 , organization=

    2020

  39. [39]

    2019 , eprint=

    Evaluating the Search Phase of Neural Architecture Search , author=. 2019 , eprint=

    2019

  40. [40]

    International Conference on Learning Representations , year=

    Understanding Architectures Learnt by Cell-based Neural Architecture Search , author=. International Conference on Learning Representations , year=

  41. [41]

    ICLR , year=

    Once for All: Train One Network and Specialize it for Efficient Deployment , author=. ICLR , year=

  42. [42]

    Le and Mark Sandler and Bo Chen and Weijun Wang and Liang

    Andrew Howard and Ruoming Pang and Hartwig Adam and Quoc V. Le and Mark Sandler and Bo Chen and Weijun Wang and Liang. Searching for MobileNetV3 , booktitle =. 2019 , url =. doi:10.1109/ICCV.2019.00140 , timestamp =

    arXiv 2019

  43. [43]

    Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part VII 16 , pages=

    Bignas: Scaling up neural architecture search with big single-stage models , author=. Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part VII 16 , pages=. 2020 , organization=

    2020

  44. [44]

    7th International Conference on Learning Representations,

    Thomas Elsken and Jan Hendrik Metzen and Frank Hutter , title =. 7th International Conference on Learning Representations,. 2019 , url =

    2019

  45. [45]

    arXiv preprint arXiv:1708.05344 , year=

    Smash: one-shot model architecture search through hypernetworks , author=. arXiv preprint arXiv:1708.05344 , year=

    Pith/arXiv arXiv

  46. [46]

    arXiv preprint arXiv:1810.05749 , year=

    Graph hypernetworks for neural architecture search , author=. arXiv preprint arXiv:1810.05749 , year=

    Pith/arXiv arXiv

  47. [47]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Deep residual learning for image recognition , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  48. [48]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Densely connected convolutional networks , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  49. [49]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Rethinking the inception architecture for computer vision , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  50. [50]

    arXiv preprint arXiv:1704.04861 , year=

    Mobilenets: Efficient convolutional neural networks for mobile vision applications , author=. arXiv preprint arXiv:1704.04861 , year=

    Pith/arXiv arXiv

  51. [51]

    2013 , publisher=

    Handbook of combinatorial optimization , author=. 2013 , publisher=

    2013

  52. [52]

    arXiv preprint arXiv:1912.12522 , year=

    NAS evaluation is frustratingly hard , author=. arXiv preprint arXiv:1912.12522 , year=

    Pith/arXiv arXiv 1912

  53. [53]

    Advances in neural information processing systems , pages=

    Neural architecture search with bayesian optimisation and optimal transport , author=. Advances in neural information processing systems , pages=

  54. [54]

    Williams, R. J. , journal =

  55. [55]

    9th International Conference on Learning Representations,

    Liam Li and Mikhail Khodak and Nina Balcan and Ameet Talwalkar , title =. 9th International Conference on Learning Representations,. 2021 , url =

    2021

  56. [56]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Can Weight Sharing Outperform Random Architecture Search? An Investigation With TuNAS , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  57. [57]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Neural architecture search using deep neural networks and monte carlo tree search , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  58. [58]

    International Conference on Machine Learning , pages=

    Stabilizing differentiable architecture search via perturbation-based regularization , author=. International Conference on Machine Learning , pages=. 2020 , organization=

    2020

  59. [59]

    8th International Conference on Learning Representations,

    Arber Zela and Thomas Elsken and Tonmoy Saikia and Yassine Marrakchi and Thomas Brox and Frank Hutter , title =. 8th International Conference on Learning Representations,. 2020 , url =

    2020

  60. [60]

    arXiv preprint arXiv:1701.00160 , year=

    NIPS 2016 tutorial: Generative adversarial networks , author=. arXiv preprint arXiv:1701.00160 , year=

    Pith/arXiv arXiv 2016

  61. [61]

    arXiv preprint arXiv:1701.04862 , year=

    Towards principled methods for training generative adversarial networks , author=. arXiv preprint arXiv:1701.04862 , year=

    Pith/arXiv arXiv

  62. [62]

    Advances in Neural Information Processing Systems , volume=

    Does unsupervised architecture representation learning help neural architecture search? , author=. Advances in Neural Information Processing Systems , volume=

  63. [63]

    Advances in Neural Information Processing Systems , volume=

    Cream of the crop: Distilling prioritized paths for one-shot neural architecture search , author=. Advances in Neural Information Processing Systems , volume=

  64. [64]

    arXiv preprint arXiv:2007.03331 , year=

    Gold-nas: Gradual, one-level, differentiable , author=. arXiv preprint arXiv:2007.03331 , year=

    Pith/arXiv arXiv 2007

  65. [65]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Squeeze-and-excitation networks , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  66. [66]

    , author=

    Semantic Segmentation of Satellite Images using a Modified CNN with Hard-Swish Activation Function. , author=. VISIGRAPP (4: VISAPP) , pages=

  67. [67]

    Icml , year=

    Rectified linear units improve restricted boltzmann machines , author=. Icml , year=

  68. [68]

    MCUNet: Tiny Deep Learning on IoT Devices , booktitle =

    Ji Lin and Wei. MCUNet: Tiny Deep Learning on IoT Devices , booktitle =. 2020 , url =

    2020

  69. [69]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    On network design spaces for visual recognition , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  70. [70]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Designing network design spaces , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  71. [71]

    Advances in Neural Information Processing Systems , volume=

    Design space for graph neural networks , author=. Advances in Neural Information Processing Systems , volume=

  72. [72]

    Lawrence and Doll \'a r, Piotr

    Lin, Tsung-Yi and Maire, Michael and Belongie, Serge and Bourdev, Lubomir and Girshick, Ross and Hays, James and Perona, Pietro and Ramanan, Deva and Zitnick, C. Lawrence and Doll \'a r, Piotr. Microsoft COCO: Common Objects in Context. Computer Vision -- ECCV 2014. 2014

    2014

  73. [73]

    The IEEE International Conference on Computer Vision (ICCV) , month =

    Zhou, Xingyi and Huang, Qixing and Sun, Xiao and Xue, Xiangyang and Wei, Yichen , title =. The IEEE International Conference on Computer Vision (ICCV) , month =

  74. [74]

    Chemometrics and intelligent laboratory systems , volume=

    Principal component analysis , author=. Chemometrics and intelligent laboratory systems , volume=. 1987 , publisher=

    1987

  75. [75]

    Interpretable Deep Convolutional Fuzzy Classifier , year=

    Yeganejou, Mojtaba and Dick, Scott and Miller, James , journal=. Interpretable Deep Convolutional Fuzzy Classifier , year=

  76. [76]

    Applied Soft Computing , volume=

    A genetic reduction of feature space in the design of fuzzy models , author=. Applied Soft Computing , volume=. 2012 , publisher=

    2012

  77. [77]

    Surrogate

    Arber Zela and Julien Niklas Siems and Lucas Zimmer and Jovita Lukasik and Margret Keuper and Frank Hutter , booktitle=. Surrogate

  78. [78]

    International Conference on Machine Learning , pages=

    EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks , author=. International Conference on Machine Learning , pages=. 2019 , organization=

    2019

  79. [79]

    , booktitle=

    Fey, Matthias and Lenssen, Jan E. , booktitle=. Fast Graph Representation Learning with

  80. [80]

    International Conference on Machine Learning , pages=

    Efficient Neural Architecture Search via Parameters Sharing , author=. International Conference on Machine Learning , pages=. 2018 , organization=

    2018

Showing first 80 references.