Pith. sign in

REVIEW 5 major objections 5 minor 41 references

Coflex: Enhancing HW-NAS with Sparse Gaussian Processes for Efficient and Scalable DNN Accelerator Design

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 Coflex claims that sparse Gaussian processes with Pareto-selected inducing points make large-scale software-hardware co-search faster and more accurate than existing multi-objective search methods.

desk verdict Sensible SGP integration for HW-NAS, but the accuracy and EDP claims rest on an unvalidated training-free proxy and an uncalibrated RNN model. read the letter →

arxiv 2507.23437 v2 pith:O6KLSTVH submitted 2025-07-31 cs.LG

Yinhui Ma , Tomomasa Yamasaki , Zhehui Wang , Tao Luo , Bo Wang This is my paper · ORCID
classification cs.LG
keywords hardware-awareneuralarchitecturesearchsparseGaussianprocessmulti-objectiveBayesianoptimizationParetofrontDNNacceleratorenergy-delayproducttraining-freeNASevaluationsoftware-hardwareco-design
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

Coflex is a hardware-aware neural architecture search framework that combines sparse Gaussian processes with multi-objective Bayesian optimization. The paper's central claim is that reducing the Gaussian-process kernel's cost from cubic to near-linear, through Pareto-selected inducing points and dimension decomposition, makes joint neural-network-and-accelerator search over spaces of order $10^{18}$-$10^{19}$ configurations practical without losing optimization quality. Across image classification, semantic segmentation, and language-modeling workloads, the reported top-1 configurations have lower error rate and lower energy-delay product than the compared multi-objective baselines, while wall-clock runtime drops by 1.9x-9.5x. If correct, this would make software-hardware co-design for edge DNN accelerators a routine, simulator-budget-friendly step rather than a prohibitive one.

What carries the argument

The load-bearing object is the sparse Gaussian process surrogate: a small set of inducing points $Z$ defines cross-covariance $K_{XZ}$ and inducing-point covariance $K_{ZZ}$, and the matrix-inversion lemma is used to avoid forming or inverting the full $n \times n$ covariance matrix. The inducing points are selected by filtering for non-dominated trade-off-front candidates, so the surrogate's fidelity is concentrated on the region of objective space that matters for the final recommendation. This is paired with a two-level decomposition that separates software-side and hardware-side hyperparameters, and with a back-end that combines a training-free network-performance evaluator and a cycle-accurate accelerator simulator to supply the two objectives.

What would settle it

Train the specific architectures that Coflex reports as best for ImageNet and CIFAR-100 to convergence and compare their measured error rates with the proxy's predictions and with the baselines' final architectures; if the measured ranking disagrees with the predicted ranking, the claim that Coflex finds more accurate networks is unsupported. On the hardware side, compare simulated energy-delay product against measurements on a fabricated or FPGA-implemented accelerator to check the hardware objective.

Watch

Extended reading notes

Core claim

The paper's central discovery is that a sparse Gaussian process surrogate, whose inducing points are chosen from the current trade-off frontier, preserves multi-objective optimization quality while removing the cubic matrix-inversion bottleneck of standard Gaussian processes. Concretely, the full covariance matrix over $n$ evaluated candidates is replaced by a low-rank factorization built from $m$ inducing points, lowering surrogate cost from $O(n^3)$ to $O(nm^2 + m^3)$ with $m \ll n$. The paper reports that this approximation, combined with a decomposition of the joint space into software and hardware subspaces, lets Coflex reach lower error rates and lower energy-delay products than the compared multi-objective optimizers across three benchmark workloads, while completing the search in substantially less wall-clock time.

Load-bearing premise

The load-bearing premise is that the training-free evaluator's predicted error rates are accurate enough to rank candidate architectures the same way true test accuracy would, since every reported error-rate improvement and final top-1 result rests on that proxy.

Editorial extensions

If this is right

  • Search spaces of order $10^{18}$-$10^{19}$ joint configurations become tractable for routine accelerator design within a few hours.
  • Increasing the number of evaluated architectures adds near-linear surrogate cost instead of cubic, so longer and denser searches are affordable at fixed budget.
  • The reported faster convergence to error-rate and energy thresholds, sometimes in as few as two iterations, implies fewer calls to expensive hardware simulators.
  • Because the same pipeline spans CNN and RNN/Transformer workloads, the method transfers across image, dense-prediction, and language tasks without redesign.

Reading between the lines

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

  • Editorial extension: if the training-free proxy's ranking fidelity holds in larger spaces, the sparse-GP treatment transfers to any simulator-dominated co-design problem, such as memory hierarchies or analog front-ends, where query cost rather than architecture count is the bottleneck.
  • Editorial extension: selecting inducing points from the trade-off frontier is a general strategy that ties surrogate capacity to the decision-relevant region of objective space; it could be lifted into other multi-objective Bayesian optimization settings even without the paper's dimension decomposition.
  • Editorial extension: at a fixed wall-clock budget, a near-linear surrogate would permit many more acquisition iterations than a full Gaussian process, so the observed quality gap may understate what the method achieves when both are given equal time.
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

5 major / 5 minor

Summary. The paper proposes Coflex, a hardware-aware neural architecture search framework that combines sparse Gaussian processes with multi-objective Bayesian optimization, dimension decomposition, and Pareto-front-based inducing point selection. It claims to reduce GP kernel complexity from O(n^3) to O(n m^2) and to outperform qNEHVI, qEHVI, qNParEGO, Random, NSGA-II, and PABO in error rate and energy-delay product on three benchmark workloads, with wall-clock speedups of 1.9x-9.5x. The empirical evaluation uses a training-free proxy (RBFlex-NAS) for network error, the DeFiNES simulator for CNN workloads, and an analytical RNN energy model for the NLP workload.

Significance. If the empirical claims were fully supported, Coflex would be a useful contribution: the complexity-reduction argument is coherent, and the idea of using Pareto-front points as inducing points is interesting and well motivated. The use of standard benchmarks and an open-source repository are strengths, and the Section III.D derivation is internally consistent under the standard DTC-style sparse-GP approximation. However, the central empirical claims currently rest on an unvalidated training-free proxy, an absent SGP-versus-full-GP ablation, single-run comparisons, and an unvalidated analytical energy model. The quantitative conclusions therefore need substantial additional evidence before they can be accepted at face value.

major comments (5)
  1. [Section IV.A; Tables V-VII] The error-rate objective used in Steps 2-4 of Figure 1 is produced by RBFlex-NAS, a training-free estimator from the authors' own group (ref. [15]), and this paper reports no calibration of that estimator against the true oracle values. For NATS-Bench-SSS and NAS-Bench-NLP the ground-truth test error and perplexity of every candidate are available, so the 'error rate' and 'log(perplexity)' entries in Tables V-VII are verifiable only if the proxy is validated. The authors should report the correlation (e.g., Spearman) and calibration between RBFlex-NAS scores and true benchmark objectives on a held-out sample, and should report the true oracle values of the final selected architectures. Without this, the claim of superior accuracy without compromising optimization performance is not empirically grounded.
  2. [Section III.D; Table V] The paper's main contribution is the SGP approximation, but no experiment isolates its effect. The baselines in Table V are labeled 'GP' but they are different optimizers (qNEHVI, qEHVI, qNParEGO), not a full-GP instantiation of Coflex. To support the claim that sparsification preserves optimization performance, the authors should compare Coflex-with-SGP against Coflex-with-full-GP under identical acquisition functions, budgets, and inducing-point-selection rules, and should sweep the number of inducing points m. Such an ablation is load-bearing because both the speedup claim and the 'without compromising optimization performance' claim depend on it.
  3. [Section V.B.1; Tables V-VII] The reported 'Dist.' values are not Euclidean distances of the printed (error rate, EDP) pairs from (0,0). For example, on CIFAR-10 the Coflex row has (9.29, 0.95), whose Euclidean norm is 9.34, not 5.52e-5. If the objectives are normalized before computing the distance, the normalization must be specified; otherwise the top-1 selection criterion is undefined and the rankings in Tables V-VII cannot be reproduced. This affects every headline top-1 comparison.
  4. [Section IV.B; Table VII] The Type-3 EDP values in Table VII are computed with the analytical RNN energy and latency model of Eqs. (16)-(17), which is not validated against any hardware measurement or cycle-accurate simulator in this paper. The model contains free parameters (beta, alpha_i, eta_i, f_ref) whose values are not reported, and the cited references [30]-[33] do not validate this specific model. Because the claimed improvement over baselines in Table VII is several orders of magnitude (2.33e-5 vs. 0.03 uJ*s), the authors must provide parameter values, a sensitivity analysis, and at least a small-scale validation against measured or simulated RNN energy.
  5. [Section V.C; Figure 5] The convergence-speedup claims are not supported as stated. Tables V-VII report single runs with no variance, so the 1.9x-9.5x speedup and all error/EDP differences may reflect noise. Moreover, Section V.C says Coflex 'reaches both thresholds within just two iterations,' but Figure 5(e) sets the EDP cutoff at 1 uJ*s while Table V reports a final EDP of 23.00 uJ*s for Coflex; under that cutoff no listed method reaches the EDP threshold. The thresholds, the convergence iteration counts, and the EDP values need to be reconciled, and multiple random seeds should be reported.
minor comments (5)
  1. [Section II.A] The text contains the typo 'Energy-Dealy-Product'; it should read 'Energy-Delay-Product'.
  2. [Figure 5] Part (b) is labeled 'Log(Pareto optimal region)' in the figure but the caption and surrounding text describe the logarithm of the number of Pareto points; the terminology should be unified.
  3. [Section III.D] Eqs. (14) and (15) use different forms of sparse-GP approximation: Eq. (14) includes a diagonal correction (FITC-like), while Eq. (15) omits it (DTC-like). The authors should state which approximation is actually implemented.
  4. [Section V.A, footnote 1] The footnote says PABO was re-implemented by the authors because the original code is unavailable, but it gives no details of the validation. Please provide the reimplementation code or a comparison on a small problem where the original behavior is documented.
  5. [Tables VI-VII] NSGA-II is present in Table V but absent from Tables VI and VII; the text explains PABO's absence for Type-2 and Type-3 workloads, but the missing NSGA-II rows should be acknowledged or completed.

Circularity Check

1 steps flagged · score 4.0 of 10

Accuracy claims rest on the authors' own RBFlex-NAS proxy (ref [15]) with no in-paper oracle validation; the SGP, EDP, and speedup claims are self-contained, so circularity is partial.

  1. self citation load bearing [Section IV.A 'Training-Free Neural Architecture Evaluation'; Section II.A and Figure 1 Steps 2-4; Abstract; Tables V-VII]
    "We adopt RBFlex-NAS [15], a state-of-the-art NAS algorithm for network performance evaluation in our framework. ... while a training-free NAS algorithm [15] is utilized to predict error rate. ... Our experimental results show that Coflex outperforms state-of-the-art methods in terms of network accuracy and Energy-Delay-Product."

    RBFlex-NAS [15] supplies the error-rate objective optimized in Figure 1 Steps 2-4 and the Err(%)/log(perplexity) entries in Tables V-VII; its five authors include four of this paper's authors (Yamasaki, Wang, Luo, Wang). The paper never validates RBFlex-NAS against ground-truth test error or perplexity, although NATS-Bench-SSS, TransNAS-Bench-101, and NAS-Bench-NLP expose oracles at zero training cost. The headline claim that Coflex outperforms SOTA 'in terms of network accuracy' therefore reduces, on this paper's evidence, to 'Coflex optimizes the authors' own training-free scorer better than baselines do,' and the only support equating that scorer with network accuracy is the overlapping self-citation [15].

full rationale

The SGP contribution is derived in-paper from standard, externally checkable mathematics: the Woodbury identity (Eq. 11) applied with A = sigma^2 I, U = K_XZ, C = K_ZZ^{-1}, V = K_ZX, together with the Matern-3/2 kernel (Eq. 9), gives O(nm^2 + m^3); the Pareto-front inducing-point filter (Eq. 7) is described by equations. The EDP half of the evaluation rests mainly on DeFiNES [14], an external simulator quantitatively validated against hardware (3% latency, 6% energy error), so that part is independent evidence, and the 1.9x-9.5x speedups are measured wall-clock times. Two limitations were weighed and found non-circular: (a) the Type-3 EDP uses the analytical RNN model (Eqs. 16-17) because 'the DeFiNES simulator is limited to supporting CNN-based architectures only'; that model is unvalidated for these RNNs, so the two-orders-of-magnitude EDP gap in Table VII is a correctness risk, not a circularity, since the model parameters are not fitted to the reported winners; and (b) the PABO baseline is the authors' re-implementation from pseudo-code because 'PABO is not publicly available,' a baseline-fidelity risk, but PABO is not this group's method, so it is not self-referential. The one load-bearing circular chain is the accuracy metric: the objective optimized (RBFlex-NAS error) and the metric reported as 'network accuracy' are the same function, and the legitimacy of that function rests on self-citation [15] with no in-paper check against benchmark oracles that are freely available. Because all methods are scored by the same proxy and the SGP/EDP/speedup claims have independent content, the score is 4: one load-bearing self-citation, with the central claim only partially reduced.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

Coflex introduces no new physical or algorithmic entities. The central claims rest on standard GP/SGP mathematics plus several domain assumptions about the accuracy of the training-free predictor and the analytical RNN energy model. The free parameters, especially the undisclosed m and RNN model coefficients, are not independently verified.

free parameters (4)
  • Number of inducing points m
    Controls SGP accuracy and complexity; never specified, yet it is the core of the near-linear complexity claim.
  • GP kernel hyperparameters (length scale, variance, noise)
    Likely fitted by marginal likelihood; values not reported, so posterior quality is not reproducible.
  • Initial sample budget = 100
    LHS design of 100 initial samples chosen by hand; affects warm start and comparison fairness.
  • RNN energy and latency model parameters (beta, alpha_i, eta_i, f_ref)
    Equations (16)-(17) depend on unstated efficiency and scaling coefficients; not validated against hardware.
assumptions (5)
  • domain assumption Sparse GP approximation of K_XX by K_XZ K_ZZ^{-1} K_ZX plus diagonal preserves enough posterior fidelity for optimization
    Used in Section III-D to justify O(nm^2) complexity and to claim no compromise in optimization performance.
  • domain assumption RBFlex-NAS training-free score is a valid proxy for true classification error rate
    Section IV.A; all accuracy objectives and Table V use this proxy instead of ground-truth benchmark labels.
  • domain assumption Analytical RNN energy and latency equations (16)-(17) reflect real hardware behavior for NAS-Bench-NLP
    Section IV.B; DeFiNES cannot simulate RNNs, so a hand-written model produces EDP values; no validation reported.
  • domain assumption Dimension decomposition into 'error-wise' and 'energy-wise' subspaces preserves the multi-objective landscape
    Section III.B; inherited from PABO; if objectives interact strongly across software and hardware dimensions, decomposition may miss optima.
  • standard math Woodbury identity and Matérn kernel give valid GP inference
    Standard results used in Sections III.C and III.D.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Coflex: Enhancing HW-NAS with Sparse Gaussian Processes for Efficient and Scalable DNN Accelerator Design." pith.science (2026). https://pith.science/paper/O6KLSTVH

@misc{pith2026250723437,
  author       = {Pith},
  title        = {Pith review of: Coflex: Enhancing HW-NAS with Sparse Gaussian Processes for Efficient and Scalable DNN Accelerator Design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O6KLSTVH}},
  note         = {Machine review of arXiv:2507.23437}
}
read the original abstract

Hardware-Aware Neural Architecture Search (HW-NAS) is an efficient approach to automatically co-optimizing neural network performance and hardware energy efficiency, making it particularly useful for the development of Deep Neural Network accelerators on the edge. However, the extensive search space and high computational cost pose significant challenges to its practical adoption. To address these limitations, we propose Coflex, a novel HW-NAS framework that integrates the Sparse Gaussian Process (SGP) with multi-objective Bayesian optimization. By leveraging sparse inducing points, Coflex reduces the GP kernel complexity from cubic to near-linear with respect to the number of training samples, without compromising optimization performance. This enables scalable approximation of large-scale search space, substantially decreasing computational overhead while preserving high predictive accuracy. We evaluate the efficacy of Coflex across various benchmarks, focusing on accelerator-specific architecture. Our experimental results show that Coflex outperforms state-of-the-art methods in terms of network accuracy and Energy-Delay-Product, while achieving a computational speed-up ranging from 1.9x to 9.5x.

Figures

Figures reproduced from arXiv: 2507.23437 by the authors.

Figure 1
Figure 1. Overview of the Coflex framework with modular components for candidate generation, performance evaluation, and Pareto-based optimization. the surrogate model in the front-end (Step 6). Through iterative optimization, Coflex refines its search and progressively converges towards Pareto-optimal design configurations. In Step 7, the final output is generated, including optimal neural network architecture and hardware c… view at source ↗
Figure 2
Figure 2. Illustration of Coflex with dimension decomposition and sparse Gaussian inducing strategies. SW1 SW2 SW3 SW4 SW5 PEs x-axis: [ 1-64 ] PEs y-axis: [ 1-64 ] I memory: [ 10-512 ] W memory: [ 10-512 ] O memory: [ 10-512 ] Bandwidth: [ 10-512 ] … … … … … … … … … … … … … SW𝑛 [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Illustration of a NAS benchmark-driven, hardware-aware search space. The number of software configuration n is defined as n = 5 for NATS-Bench￾SSS, n = 6 for TransNAS-Bench-101, and n = 8 for NAS-Bench-NLP. As we wish to predict the function output y∗ at a new test point x∗, we then assume the objective function f(x) satisfies a GP prior as f(x) ∼ GP(µ(x), k(x, x ′ )) (1) The function values at any finite set of inp… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Overview of the accelerator prototype for hardware space search. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Performance Evaluation under Workload1 on ImageNet: (a) Pareto front in the EDP (uJ.s) vs error rate domain, illustrating the trade-off between [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Comparison of runtime (in minutes) for different optimization methods [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 37 canonical work pages

  1. [15]

    Rbflex-nas: Training-free neural architecture search using radial basis function kernel and hyperparameter detection,

    T. Yamasaki, Z. Wang, T. Luo, N. Chen, and B. Wang, “Rbflex-nas: Training-free neural architecture search using radial basis function kernel and hyperparameter detection,” IEEE Transactions on Neural Networks and Learning Systems , 2025

    work page 2025

  2. [30]

    Elsa: A throughput-optimized design of an lstm accelerator for energy-constrained devices,

    E. Azari and S. Vrudhula, “Elsa: A throughput-optimized design of an lstm accelerator for energy-constrained devices,” ACM Transactions on Embedded Computing Systems (TECS) , vol. 19, no. 1, pp. 1–21, 2020

    work page 2020

  3. [33]

    Vau da muntanialas: Energy-efficient multi-die scalable acceleration of rnn inference,

    G. Paulin, F. Conti, L. Cavigelli, and L. Benini, “Vau da muntanialas: Energy-efficient multi-die scalable acceleration of rnn inference,” IEEE Transactions on Circuits and Systems I: Regular Papers , vol. 69, no. 1, pp. 244–257, 2021

    work page 2021

  4. [1]

    Pruning neural networks without any data by iteratively conserving synaptic flow,

    H. Tanaka, D. Kunin, D. L. Yamins, and S. Ganguli, “Pruning neural networks without any data by iteratively conserving synaptic flow,” Advances in neural information processing systems , vol. 33, pp. 6377– 6389, 2020

    work page 2020

  5. [2]

    Neural architecture search: Insights from 1000 papers,

    C. White, M. Safari, R. Sukthanker, B. Ru, T. Elsken, A. Zela, D. Dey, and F. Hutter, “Neural architecture search: Insights from 1000 papers,” arXiv preprint arXiv:2301.08727 , 2023

    arXiv 2023

  6. [3]

    Multiobjective tree-structured parzen estimator,

    Y . Ozaki, Y . Tanigaki, S. Watanabe, M. Nomura, and M. Onishi, “Multiobjective tree-structured parzen estimator,” Journal of Artificial Intelligence Research, vol. 73, pp. 1209–1250, 2022

    work page 2022

  7. [4]

    Edgeyolo: An edge-real- time object detector,

    S. Liu, J. Zha, J. Sun, Z. Li, and G. Wang, “Edgeyolo: An edge-real- time object detector,” in 2023 42nd Chinese Control Conference (CCC) . IEEE, 2023, pp. 7507–7512

    work page 2023

  8. [5]

    Edge computing for real-time internet of things applications: Future internet revolution,

    N. M. Quy, L. A. Ngoc, N. T. Ban, N. V . Hau, and V . K. Quy, “Edge computing for real-time internet of things applications: Future internet revolution,” Wireless Personal Communications , vol. 132, no. 2, pp. 1423–1452, 2023

    work page 2023

Show all 41 references
  1. [6]

    Making accurate object detection at the edge: Review and new approach,

    Z. Huang, S. Yang, M. Zhou, Z. Gong, A. Abusorrah, C. Lin, and Z. Huang, “Making accurate object detection at the edge: Review and new approach,” Artificial Intelligence Review , vol. 55, no. 3, pp. 2245– 2274, 2022

    2022

  2. [7]

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

    H. Cai, C. Gan, T. Wang, Z. Zhang, and S. Han, “Once-for-all: Train one network and specialize it for efficient deployment,” arXiv preprint arXiv:1908.09791, 2019

    1908 arXiv

  3. [8]

    Pabo: Pseudo agent-based multi-objective bayesian hyperparameter optimization for efficient neural accelerator design,

    M. Parsa, A. Ankit, A. Ziabari, and K. Roy, “Pabo: Pseudo agent-based multi-objective bayesian hyperparameter optimization for efficient neural accelerator design,” in 2019 IEEE/ACM International Conference on Computer-Aided Design (ICCAD) . IEEE, 2019, pp. 1–8

    2019

  4. [9]

    Bayesian multi-objective hyperparameter optimization for accurate, fast, and efficient neural network accelerator design,

    M. Parsa, J. P. Mitchell, C. D. Schuman, R. M. Patton, T. E. Potok, and K. Roy, “Bayesian multi-objective hyperparameter optimization for accurate, fast, and efficient neural network accelerator design,” Frontiers in neuroscience, vol. 14, p. 667, 2020

    2020

  5. [10]

    Neural architecture search for in-memory computing-based deep learning accelerators,

    O. Krestinskaya, M. E. Fouda, H. Benmeziane, K. El Maghraoui, A. Sebastian, W. D. Lu, M. Lanza, H. Li, F. Kurdahi, S. A. Fahmy et al., “Neural architecture search for in-memory computing-based deep learning accelerators,” Nature Reviews Electrical Engineering , vol. 1, no. 6, ...

    2024

  6. [11]

    Pymoo: Multi-objective optimization in python,

    J. Blank and K. Deb, “Pymoo: Multi-objective optimization in python,” Ieee access, vol. 8, pp. 89 497–89 509, 2020

    2020

  7. [12]

    A fast and elitist multiobjective genetic algorithm: Nsga-ii,

    K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitist multiobjective genetic algorithm: Nsga-ii,” IEEE transactions on evolu- tionary computation, vol. 6, no. 2, pp. 182–197, 2002

    2002

  8. [13]

    Sparse gaussian processes using pseudo-inputs,

    E. Snelson and Z. Ghahramani, “Sparse gaussian processes using pseudo-inputs,” Advances in neural information processing systems , vol. 18, 2005

    2005

  9. [14]

    Defines: En- abling fast exploration of the depth-first scheduling space for dnn accel- erators through analytical modeling,

    L. Mei, K. Goetschalckx, A. Symons, and M. Verhelst, “Defines: En- abling fast exploration of the depth-first scheduling space for dnn accel- erators through analytical modeling,” in 2023 IEEE International Sym- posium on High-Performance Computer Architecture (HPCA) . IEEE, 2...

    2023

  10. [16]

    Nats-bench: Benchmarking nas algorithms for architecture topology and size,

    X. Dong, L. Liu, K. Musial, and B. Gabrys, “Nats-bench: Benchmarking nas algorithms for architecture topology and size,” IEEE transactions on pattern analysis and machine intelligence, vol. 44, no. 7, pp. 3634–3646, 2021

    2021

  11. [17]

    Nas-bench-201: Extending the scope of repro- ducible neural architecture search,

    X. Dong and Y . Yang, “Nas-bench-201: Extending the scope of repro- ducible neural architecture search,” arXiv preprint arXiv:2001.00326 , 2020

    2001 arXiv

  12. [18]

    Introduction to gaussian processes,

    D. J. MacKay et al. , “Introduction to gaussian processes,” NATO ASI series F computer and systems sciences , vol. 168, pp. 133–166, 1998

    1998

  13. [19]

    Improving the expected improvement algorithm,

    C. Qin, D. Klabjan, and D. Russo, “Improving the expected improvement algorithm,” Advances in Neural Information Processing Systems, vol. 30, 2017

    2017

  14. [20]

    Transnas-bench-101: Improving transferability and generalizability of cross-task neural architecture search,

    Y . Duan, X. Chen, H. Xu, Z. Chen, X. Liang, T. Zhang, and Z. Li, “Transnas-bench-101: Improving transferability and generalizability of cross-task neural architecture search,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021, pp. 5251–5260

    2021

  15. [21]

    Nas-bench-nlp: neural architecture search benchmark for natural language processing,

    N. Klyuchnikov, I. Trofimov, E. Artemova, M. Salnikov, M. Fedorov, A. Filippov, and E. Burnaev, “Nas-bench-nlp: neural architecture search benchmark for natural language processing,” IEEE Access, vol. 10, pp. 45 736–45 747, 2022

    2022

  16. [22]

    Gaussian processes for machine learning,

    M. Seeger, “Gaussian processes for machine learning,” International journal of neural systems , vol. 14, no. 02, pp. 69–106, 2004

    2004

  17. [23]

    Sur la r ´esolution num ´erique des syst `emes d’ ´equations lin´eaires,

    A.-L. Cholesky, “Sur la r ´esolution num ´erique des syst `emes d’ ´equations lin´eaires,” Bulletin de la Sabix. Soci ´et´e des amis de la Biblioth `eque et de l’Histoire de l’ ´Ecole polytechnique, no. 39, pp. 81–95, 2005

    2005

  18. [24]

    Neural architec- ture search without training,

    J. Mellor, J. Turner, A. Storkey, and E. J. Crowley, “Neural architec- ture search without training,” in International conference on machine learning. PMLR, 2021, pp. 7588–7598

    2021

  19. [25]

    Network pruning via transformable architecture search,

    X. Dong and Y . Yang, “Network pruning via transformable architecture search,” Advances in Neural Information Processing Systems , vol. 32, 2019

    2019

  20. [26]

    Fbnetv2: Differentiable neural architecture search for spatial and channel dimensions,

    A. Wan, X. Dai, P. Zhang, Z. He, Y . Tian, S. Xie, B. Wu, M. Yu, T. Xu, K. Chen et al. , “Fbnetv2: Differentiable neural architecture search for spatial and channel dimensions,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 12 ...

    2020

  21. [27]

    Neural architecture search on imagenet in four gpu hours: A theoretically inspired perspective,

    W. Chen, X. Gong, and Z. Wang, “Neural architecture search on imagenet in four gpu hours: A theoretically inspired perspective,” arXiv preprint arXiv:2102.11535, 2021

    2021 arXiv

  22. [28]

    Zico: Zero-shot nas via inverse coefficient of variation on gradients,

    G. Li, Y . Yang, K. Bhardwaj, and R. Marculescu, “Zico: Zero-shot nas via inverse coefficient of variation on gradients,” arXiv preprint arXiv:2301.11300, 2023

    2023 arXiv

  23. [29]

    System-level design and integration of a prototype ar/vr hardware featuring a custom low-power dnn accelerator chip in 7nm technology for codec avatars,

    H. E. Sumbul, T. F. Wu, Y . Li, S. S. Sarwar, W. Koven, E. Murphy- Trotzky, X. Cai, E. Ansari, D. H. Morris, H. Liu, D. Kim, E. Beigne, R. Labs, and Meta, “System-level design and integration of a prototype ar/vr hardware featuring a custom low-power dnn accelerator chip in 7n...

    2022

  24. [31]

    E-rnn: Design optimization for efficient recurrent neural networks in fpgas,

    Z. Li, C. Ding, S. Wang, W. Wen, Y . Zhuo, C. Liu, Q. Qiu, W. Xu, X. Lin, X. Qian et al., “E-rnn: Design optimization for efficient recurrent neural networks in fpgas,” in 2019 IEEE International Symposium on High Performance Computer Architecture (HPCA) . IEEE, 2019, pp. 69–80

    2019

  25. [32]

    E-pur: An energy- efficient processing unit for recurrent neural networks,

    F. Silfa, G. Dot, J.-M. Arnau, and A. Gonz `alez, “E-pur: An energy- efficient processing unit for recurrent neural networks,” in Proceedings of the 27th international conference on parallel architectures and compilation techniques, 2018, pp. 1–12

    2018

  26. [34]

    MOBO: A new software for multi-objective building performance optimization,

    M. Palonen, M. Hamdy, and A. Hasan, “MOBO: A new software for multi-objective building performance optimization,” in Proceedings of the 13th Conference of International Building Performance Simulation Association (BS2013) . Chambery, France: IBPSA, 2013, pp. 2566–2573. [Online...

    2013

  27. [35]

    Parallel bayesian optimization of multiple noisy objectives with expected hypervolume improvement,

    S. Daulton, M. Balandat, and E. Bakshy, “Parallel bayesian optimization of multiple noisy objectives with expected hypervolume improvement,” Advances in Neural Information Processing Systems , vol. 34, pp. 2187– 2200, 2021

    2021

  28. [36]

    Differentiable expected hypervolume improvement for parallel multi-objective bayesian optimization,

    ——, “Differentiable expected hypervolume improvement for parallel multi-objective bayesian optimization,” Advances in Neural Information Processing Systems, vol. 33, pp. 9851–9864, 2020

    2020

  29. [37]

    Parego: A hybrid algorithm with on-line landscape ap- proximation for expensive multiobjective optimization problems,

    J. Knowles, “Parego: A hybrid algorithm with on-line landscape ap- proximation for expensive multiobjective optimization problems,” IEEE transactions on evolutionary computation , vol. 10, no. 1, pp. 50–66, 2006

    2006

  30. [38]

    Distributed computing of pareto- optimal solutions with evolutionary algorithms,

    K. Deb, P. Zope, and A. Jain, “Distributed computing of pareto- optimal solutions with evolutionary algorithms,” in Evolutionary Multi- Criterion Optimization: Second International Conference, EMO 2003, Faro, Portugal, April 8–11, 2003. Proceedings 2 . Springer, 2003, pp. 534–549

    2003

  31. [39]

    From a pareto front to pareto regions: A novel standpoint for multiobjective optimization,

    C. M. Rebello, M. A. Martins, D. D. Santana, A. E. Rodrigues, J. M. Loureiro, A. M. Ribeiro, and I. B. Nogueira, “From a pareto front to pareto regions: A novel standpoint for multiobjective optimization,” Mathematics, vol. 9, no. 24, p. 3152, 2021

    2021

  32. [40]

    Multiobjective optimization using evolutionary algorithms—a comparative case study,

    E. Zitzler and L. Thiele, “Multiobjective optimization using evolutionary algorithms—a comparative case study,” in International conference on parallel problem solving from nature . Springer, 1998, pp. 292–301

    1998

  33. [41]

    Sms-emoa: Multiobjective selection based on dominated hypervolume,

    N. Beume, B. Naujoks, and M. Emmerich, “Sms-emoa: Multiobjective selection based on dominated hypervolume,” European journal of oper- ational research, vol. 181, no. 3, pp. 1653–1669, 2007

    2007

Pith tools

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