Pith. sign in

REVIEW 3 major objections 6 minor 39 references

DiffPattern-Flex: Efficient Layout Pattern Generation via Discrete Diffusion

T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read DiffPattern-Flex decouples topology generation from legalization and reports 100% legal, more diverse chip layout patterns at 8.37x sampling speed.

desk verdict Solid engineering extension of the authors' own DiffPattern; the 100% legality claim is well-supported empirically, but missing code and absence of a completeness proof keep it from being a clean accept. read the letter →

arxiv 2505.04173 v1 pith:4VJVAXYO submitted 2025-05-07 cs.LG cs.CV

classification cs.LGcs.CV
keywords PatternGenerationDesignForManufacturabilityDiffusionModelsLegalizationDiscreteLayoutDeepSquishVLSI
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

DiffPattern-Flex seeks to make VLSI layout pattern generation both diverse and trustworthy by stopping the practice of letting a neural network implicitly learn design-rule compliance. Instead, it separates the problem: a discrete diffusion model generates binary topology tensors, and a white-box legalization system assigns geometric vectors by solving a nonlinear program built from the space, width, and area rules. The paper reports that this yields 100% legal patterns across 100,000 Klayout checks, raises library diversity from 10.815 to 11.713 relative to its predecessor DiffPattern, and cuts sampling and legalization time by 8.37x and 2.48x respectively. If correct, the framework matters because downstream design-for-manufacturability tasks such as lithography simulation and hotspot detection require large pattern libraries that are guaranteed to obey design rules, and existing learning-based generators often return illegal patterns.

What carries the argument

The argument rides on two coupled mechanisms. First, the Deep Squish Pattern representation: a lossless compression that folds a binary topology matrix into a multi-channel tensor, exploiting the observation that diffusion models are more sensitive to input size than to channel count, and keeping every entry binary so the discrete diffusion model can be trained without thresholding. Second, the white-box legalization step: a nonlinear system (Equation 14) that, for each generated topology, collects all space, width, and area constraints into a solvable feasibility problem over the geometric vectors; solving it certifies legality before the pattern exists as a layout. The two are coupled by the fact that topology generation only proposes binary matrices, and the legalizer either finds a legal geometry or the topology is discarded.

What would settle it

Run DiffPattern-Flex on a design-rule deck that adds a check not present in Equation (14)—for example a minimum diagonal corner-to-corner spacing or a polygon-enclosure rule—generate a large batch of patterns, and ask Klayout whether any pattern the legalizer accepted is flagged as a violation. A single such violation refutes the claim of 100% legality as stated.

Watch

Extended reading notes

Core claim

The paper's central claim is that layout pattern generation becomes reliable and efficient when topology synthesis and geometric legalization are decoupled. The discrete diffusion model directly outputs binary topology tensors—represented losslessly as Deep Squish Patterns, which fold $\sqrt{C}\times\sqrt{C}$ patches into $C$ channels to shrink the diffusion input without information loss. For each generated topology, the white-box legalizer solves a nonlinear system (Equation 14) that spells out space, width, and per-polygon area constraints over the geometric vectors $\Delta x$ and $\Delta y$; by construction, any solution is DRC-clean, and the paper reports perfect legality on 100,000 generated patterns as verified by Klayout. Because the legalization step is separated from the generative model, design-rule changes require no retraining, and a single topology can be instantiated as many distinct legal patterns.

Load-bearing premise

The 100% legality guarantee rests entirely on the assumption that the space, width, and area constraints written into Equation (14) are exactly the checks the design-rule checker performs; if a rule is missing from that system, a pattern the method calls legal could still be flagged as illegal.

Editorial extensions

If this is right

  • Changing design rules no longer forces retraining: only the constants and constraint sets in the legalization system are updated, while the topology generator stays untouched.
  • A single generated topology can spawn many distinct legal layouts because the nonlinear system typically has multiple geometric-vector solutions.
  • Sampling can be made about 8.37x faster by reversing the diffusion process 10 steps per neural-network call, with diversity effectively unchanged (11.724 to 11.713), while pushing to 20 steps drops diversity to 10.573.
  • Modeling topology tensors as discrete states yields higher diversity than a continuous diffusion baseline (11.713 vs 11.294) under identical training protocols.
  • Data augmentation is made safe by pre-checking every augmented topology with the legalizer, so the training set never contains illegal patterns.

Reading between the lines

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

  • The same decoupling recipe—neural proposal plus solver certification—is a template for other domains where generated geometry must satisfy hard rules, such as PCB routing, analog layout, or floorplanning, whenever a complete constraint set can be written down.
  • The 100% legality figure is certified only for the three rule types encoded in Equation (14); extending the claim to a full production DRC deck would require the constraint set to be derived automatically from the deck and verified against it.
  • Because diversity degrades sharply at $m=20$, an adaptive fast-sampling schedule that uses large $m$ only for simple topologies is a natural testable extension that could preserve diversity at higher speed.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. This paper proposes DiffPattern-Flex, a layout pattern generation framework that separates topology generation from geometric legalization. Topologies are synthesized by a discrete (two-state) diffusion model operating on a 'Deep Squish Pattern' representation obtained by folding binary topology matrices into multi-channel tensors. Geometry is then assigned by solving a nonlinear constraint system (Eq. (14)) that encodes Space, Width, and Area design rules, which the authors state guarantees 100% DRC-clean patterns. The paper also contributes closed-form step-skip sampling for the discrete diffusion process, an initialization strategy for the legalization solver (Solving-E and Solving-D), and a pre-legalized data augmentation scheme. Experiments on ICCAD 2014 data report a diversity H of 11.713 with 100,000 legal patterns, 8.37x sampling speedup (m=10), and 2.48x legalization speedup, outperforming prior DiffPattern and other baselines on the H metric.

Significance. If the claims hold after the technical corrections below, this is a practically relevant contribution: it shows that a discrete diffusion model over binary topologies combined with a white-box optimization-based legalizer can produce large libraries of design-rule-clean patterns with improved complexity diversity and a substantial speedup over the previous DiffPattern. The Deep Squish Pattern is a useful lossless compression idea for pixel-based layout generation, and the decoupling of topology generation from legalization provides a natural path to adapting to changed design rules without retraining. The paper reports external validation with Klayout on 100,000 patterns and gives closed-form transition formulas for the discrete diffusion posterior, which are standard and internally consistent for the single-step case. The main risks are the correctness of the m-step skip formula and the formal completeness of the legalization constraint set.

major comments (3)
  1. [IV-B, Eq. (17)] Equation (17) does not give the correct closed form for q(x_{k-m}|x_k,x0). The forward transition over m steps is governed by the product R = Q_{k-m+1}...Q_k, so the Bayes posterior should contain x_k R^T (elementwise multiplied with x0 in the cumulative matrix), not x_k Q_k^T. Under the notation of Eq. (10), Q_k is the cumulative product, making Eq. (17) incorrect for any m>1; under the alternative reading where Q_k is a single-step matrix, the formula omits all intermediate transitions. Because the fast-sampling acceleration in Table II is derived from this equation, the derivation and the reported 8.37x speedup need to be re-examined. The authors should either correct Eq. (17) (e.g., using R = Q_{k-m+1}...Q_k, or its commutation-based equivalent when the matrices commute) and rerun the experiments, or clearly separate the notation for single-step and cumulative transition matrices.
  2. [III-D, Table I] The 100% legality claim is only as strong as the completeness of the constraint system in Eq. (14) relative to the checker. The paper defines the design rules as Space, Width, and Area, but it does not provide the exact construction of SetS and SetW, nor the precise Klayout rule deck used in the 100,000-pattern verification, nor the code. Without an explicit mapping between Eq. (14) and the Klayout checks, the universal claim '100% legality' is not verifiable. The authors should either state and prove the equivalence between Eq. (14) and the DRC deck, or release the deck/checker configuration, and should clarify that the claim applies only to this specified rule set. The statement in Section III-D that unsolvable cases are discarded also means the guarantee applies to the subset of topologies that reach a solved solution, not unconditionally to all generated topologies.
  3. [Table I] The learning-based baseline results in Table I are taken from the authors' prior DiffPattern work [18] rather than re-run under the authors' evaluation pipeline. Because the diversity metric H is a finite-sample entropy and the evaluation involves the legalization stage, the comparison may not be fully controlled. The authors should rerun the baselines with the same pipeline (same sample count, same H computation, same design-rule constants), or at minimum state explicitly which settings from [18] were used and whether the legalization method affects the baseline diversity numbers.
minor comments (6)
  1. [Throughout] Typos: 'Klayout' should be 'KLayout'; 'Continual' in Section V.F and Table IV should be 'Continuous'; 'prabability density (log)' in Fig. 12 should be 'probability density (log)'.
  2. [Experimental Setup] The numeric values of Spacemin, Widthmin, Areamin, and Areamax used in the experiments are not reported. These constants are necessary for reproducibility, especially because Eq. (14) depends on them.
  3. [III-D, Eq. (14)] The construction of SetS and SetW is described only informally as 'pattern-dependent'; a precise algorithm or a worked example for deriving these sets from a topology matrix would make the legalization system reproducible and auditable.
  4. [III-C] The topology pre-filter that removes Bow-tie shapes is mentioned but never defined. Please specify what topologies are filtered and how the filter is implemented, and provide the measured fraction of filtered samples that supports the 'less than 0.1%' claim.
  5. [II-B, Eq. (14)] The notation sqrt(C) x sqrt(C) in Eq. (14) should be tied to the actual values: with C=16, the fold is 4x4 and the unfolded topology matrix is 4M x 4M. Stating this explicitly would remove ambiguity for readers applying the representation.
  6. [V.E] The statement that fast sampling is 'roughly accelerated by a factor of m' is approximate; the empirical factor in Table II is 8.37x for m=10, which is reasonable but depends on implementation overhead. Please state the overhead components and make clear that m=20 (13.95x) incurs a diversity drop from 11.713 to 10.573.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the reported legality and diversity numbers are measured on generated patterns with an external checker and an explicit entropy metric, not derived from fitted inputs or self-citations.

full rationale

The derivation chain is self-contained. Topology tensors are generated by a discrete diffusion model trained with a standard variational objective (Eqs. 5-13), and diversity is measured on generated samples using the entropy definition in Eq. (4). No parameter is fitted to the headline diversity or legality numbers and then reported as a prediction. The legalization phase solves the nonlinear system in Eq. (14), and legality is subsequently checked by Klayout, an external tool, so the 100% legal-pattern claim is an empirical verification rather than a restatement of the objective. The augmentation pipeline in Eq. (15) is filtered by the same legalizer; while augmentation is deliberately chosen to increase the diversity metric, that is a training-data design choice, not a circular derivation. The main self-citations — DiffPattern [18] as a baseline and squish-pattern related work [20], [26] — are used for comparison or as prior representation work, and are not invoked to justify the central claim. The paper's own caveats, such as Section III-D's statement that unsolvable cases 'can simply be discarded' and Section V-C's admission that extremely strict design rules 'may fail to find a legal solution,' are completeness limitations of the white-box DRC model, not circularity. The headline legalization guarantee therefore rests on an unproven completeness assumption for Eq. (14), which is a correctness risk, but no step in the paper reduces to its own input by construction.

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

The framework leans on standard diffusion-model mathematics and a lossless squish representation from prior patents. The load-bearing extra assumptions are that the white-box constraint system in Equation (14) completely matches Klayout's DRC checks and that the nonlinear solver can always reach a feasible solution under the tested rules; both are supported only empirically.

free parameters (7)
  • m (fast-sampling step skip) = 10
    Chosen as a trade-off between sampling speed and diversity. Table II shows m=20 drops diversity from 11.713 to 10.573.
  • C (channel folding size) = 16
    Patch size sqrt(C) x sqrt(C) = 4x4 chosen to balance information density and input size; affects model capacity and diversity.
  • beta1, betaK (noise schedule endpoints) = 0.01, 0.5
    Linear noise schedule from Equation (8), chosen to make the forward process converge to a uniform stationary distribution.
  • K (diffusion steps) = 1000
    Set to guarantee convergence of the forward process to the stationary distribution.
  • lambda (loss weight) = 0.001
    Balances the reconstruction term in Equation (9); no sensitivity study is reported.
  • Augmentation probabilities = [0.5, 1.0, 0.5, 0.5]
    Probabilities for random flip, rotation, mirror, and concatenate-crop; tuned with an ablation on the concatenate-crop probability only (Fig. 14).
  • Divide size for legalization = 4
    Number of sub-problems in the divide-and-conquer legalization initialization, fixed without ablation.
assumptions (4)
  • standard math A doubly stochastic transition matrix with a linearly increasing beta schedule makes q(xK|x0) close to [0.5, 0.5].
    Invoked in Section III-C, Equations (6) to (8); convergence is stated rather than proved and follows the D3PM framework.
  • domain assumption The Squish Pattern representation, topology matrix plus geometric vectors, is lossless and sufficient to reconstruct the original layout.
    Taken from patent [20] and used in Section II-B; the paper does not re-derive this property.
  • domain assumption Equation (14) is a complete encoding of the design rules checked by Klayout.
    Section III-D states that all constraints are listed, but no proof is given that the nonlinear system cannot miss a DRC rule.
  • domain assumption The nonlinear solver will find a feasible solution within the time budget for generated topologies under the tested design rules.
    Section III-D says 'never occurred in our experiments (more than 1.0e8 attempts)', which is empirical support but not a guarantee.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DiffPattern-Flex: Efficient Layout Pattern Generation via Discrete Diffusion." pith.science (2026). https://pith.science/paper/4VJVAXYO

@misc{pith2026250504173,
  author       = {Pith},
  title        = {Pith review of: DiffPattern-Flex: Efficient Layout Pattern Generation via Discrete Diffusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4VJVAXYO}},
  note         = {Machine review of arXiv:2505.04173}
}
read the original abstract

Recent advancements in layout pattern generation have been dominated by deep generative models. However, relying solely on neural networks for legality guarantees raises concerns in many practical applications. In this paper, we present \tool{DiffPattern}-Flex, a novel approach designed to generate reliable layout patterns efficiently. \tool{DiffPattern}-Flex incorporates a new method for generating diverse topologies using a discrete diffusion model while maintaining a lossless and compute-efficient layout representation. To ensure legal pattern generation, we employ {an} optimization-based, white-box pattern assessment process based on specific design rules. Furthermore, fast sampling and efficient legalization technologies are employed to accelerate the generation process. Experimental results across various benchmarks demonstrate that \tool{DiffPattern}-Flex significantly outperforms existing methods and excels at producing reliable layout patterns.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 31 canonical work pages

  1. [18]

    Diff- Pattern: Layout pattern generation via discrete diffusion,

    Z. Wang, Y . Shen, W. Zhao, Y . Bai, G. Chen, F. Farnia, and B. Yu, “Diff- Pattern: Layout pattern generation via discrete diffusion,” in ACM/IEEE Design Automation Conference (DAC) . IEEE, 2023, pp. 1–6

  2. [1]

    X-check: Gpu-accelerated design rule checking via parallel sweepline algorithms,

    Z. He, Y . Ma, and B. Yu, “X-check: Gpu-accelerated design rule checking via parallel sweepline algorithms,” in IEEE/ACM International Conference on Computer-Aided Design (ICCAD) , 2022, pp. 1–9

  3. [2]

    A formal approach to design-rule checking,

    H. Modarres and R. J. Lomax, “A formal approach to design-rule checking,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (TCAD) , vol. 6, no. 4, pp. 561–573, 1987

  4. [3]

    Pdrc: Package design rule checking via gpu-accelerated geometric intersection algo- rithms for non-manhattan geometry,

    J. Jiang, L. Zou, W. Zhao, Z. He, T. Chen, and B. Yu, “Pdrc: Package design rule checking via gpu-accelerated geometric intersection algo- rithms for non-manhattan geometry,” in ACM/IEEE Design Automation Conference (DAC), 2024

  5. [4]

    MOSAIC: Mask optimizing solution with process window aware inverse correction,

    J.-R. Gao, X. Xu, B. Yu, and D. Z. Pan, “MOSAIC: Mask optimizing solution with process window aware inverse correction,” in ACM/IEEE Design Automation Conference (DAC) . IEEE, 2014, pp. 1–6. 11

  6. [5]

    Automated optical proximity correction: a rules-based approach,

    O. W. Otto, J. G. Garofalo, K. Low, C.-M. Yuan, R. C. Henderson, C. Pierrat, R. L. Kostelak, S. Vaidya, and P. Vasudev, “Automated optical proximity correction: a rules-based approach,” in Optical/Laser Microlithography VII, vol. 2197. SPIE, 1994, pp. 278–293

  7. [6]

    Model-based opc extension in openilt,

    S. Zheng, G. Xiao, G. Yan, M. Dong, Y . Li, H. Chen, Y . Ma, B. Yu, and M. Wong, “Model-based opc extension in openilt,” in International Symposium of Electronics Design Automation (ISEDA) . IEEE, 2024, pp. 568–573

  8. [7]

    An efficient layout decomposition approach for triple patterning lithography,

    J. Kuang and E. F. Young, “An efficient layout decomposition approach for triple patterning lithography,” in ACM/IEEE Design Automation Conference (DAC). IEEE, 2013, pp. 1–6

Show all 39 references
  1. [8]

    Layout decomposition for triple patterning lithography,

    B. Yu, K. Yuan, D. Ding, and D. Z. Pan, “Layout decomposition for triple patterning lithography,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (TCAD) , vol. 34, no. 3, pp. 433–446, 2015

  2. [9]

    Ultra-fast source mask optimization via conditional discrete diffusion,

    G. Chen, Z. Wang, B. Yu, D. Z. Pan, and M. D. Wong, “Ultra-fast source mask optimization via conditional discrete diffusion,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (TCAD) , 2024

  3. [10]

    Faster region-based hotspot detection,

    R. Chen, W. Zhong, H. Yang, H. Geng, X. Zeng, and B. Yu, “Faster region-based hotspot detection,” in ACM/IEEE Design Automation Con- ference (DAC), 2019, pp. 1–6

  4. [11]

    Layout hotspot detection with feature tensor generation and deep biased learning,

    H. Yang, J. Su, Y . Zou, B. Yu, and E. F. Young, “Layout hotspot detection with feature tensor generation and deep biased learning,” in ACM/IEEE Design Automation Conference (DAC) , 2017, pp. 1–6

  5. [12]

    Hotspot detection via attention-based deep layout metric learning,

    H. Geng, H. Yang, L. Zhang, F. Yang, X. Zeng, and B. Yu, “Hotspot detection via attention-based deep layout metric learning,” IEEE Trans- actions on Computer-Aided Design of Integrated Circuits and Systems (TCAD), vol. 41, no. 8, pp. 2685–2698, 2022

  6. [13]

    Enhanced hotspot de- tection through synthetic pattern generation and design of experiments,

    G. R. Reddy, C. Xanthopoulos, and Y . Makris, “Enhanced hotspot de- tection through synthetic pattern generation and design of experiments,” in IEEE VLSI Test Symposium (VTS) . IEEE, 2018, pp. 1–6

  7. [14]

    LithoROC: lithography hotspot detection with explicit ROC optimization,

    W. Ye, Y . Lin, M. Li, Q. Liu, and D. Z. Pan, “LithoROC: lithography hotspot detection with explicit ROC optimization,” in ACM/IEEE Design Automation Conference (DAC), 2019, pp. 292–298

  8. [15]

    DeePattern: Lay- out pattern generation with transforming convolutional auto-encoder,

    H. Yang, P. Pathak, F. Gennari, Y .-C. Lai, and B. Yu, “DeePattern: Lay- out pattern generation with transforming convolutional auto-encoder,” in ACM/IEEE Design Automation Conference (DAC) , 2019, pp. 1–6

  9. [16]

    Layout pattern generation and le- galization with generative learning models,

    X. Zhang, J. Shiely, and E. F. Young, “Layout pattern generation and le- galization with generative learning models,” in IEEE/ACM International Conference on Computer-Aided Design (ICCAD). IEEE, 2020, pp. 1–9

  10. [17]

    Lay- ouTransformer: Generating Layout Patterns with Transformer via Se- quential Pattern Modeling,

    L. Wen, Y . Zhu, L. Ye, G. Chen, B. Yu, J. Liu, and C. Xu, “Lay- ouTransformer: Generating Layout Patterns with Transformer via Se- quential Pattern Modeling,” in IEEE/ACM International Conference on Computer-Aided Design (ICCAD) . IEEE, 2022

  11. [19]

    ChatPattern: Layout Pattern Customization via Natural Language,

    Z. Wang, Y . Shen, X. Yao, W. Zhao, Y . Bai, F. Farnia, and B. Yu, “ChatPattern: Layout Pattern Customization via Natural Language,” in ACM/IEEE Design Automation Conference (DAC) , 2024

  12. [20]

    Topology design using squish patterns,

    F. E. Gennari and Y .-C. Lai, “Topology design using squish patterns,” Sep. 9 2014, uS Patent 8,832,621

  13. [21]

    Denoising diffusion probabilistic mod- els,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic mod- els,” Annual Conference on Neural Information Processing Systems (NeurIPS), vol. 33, pp. 6840–6851, 2020

  14. [22]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” in International Conference on Learning Representations (ICLR) , 2020

  15. [23]

    Diffusion models: A comprehensive survey of methods and applications,

    L. Yang, Z. Zhang, Y . Song, S. Hong, R. Xu, Y . Zhao, Y . Shao, W. Zhang, B. Cui, and M.-H. Yang, “Diffusion models: A comprehensive survey of methods and applications,” arXiv preprint arXiv:2209.00796 , 2022

  16. [24]

    J. R. Norris, Markov chains. Cambridge university press, 1998, no. 2

  17. [25]

    Tutorial on variational autoencoders,

    C. Doersch, “Tutorial on variational autoencoders,” arXiv preprint arXiv:1606.05908, 2016

  18. [26]

    Detecting multi-layer layout hotspots with adaptive squish patterns,

    H. Yang, P. Pathak, F. Gennari, Y .-C. Lai, and B. Yu, “Detecting multi-layer layout hotspots with adaptive squish patterns,” in ACM/IEEE Design Automation Conference (DAC) , 2019, pp. 299–304

  19. [27]

    Structured denoising diffusion models in discrete state-spaces,

    J. Austin, D. D. Johnson, J. Ho, D. Tarlow, and R. van den Berg, “Structured denoising diffusion models in discrete state-spaces,” An- nual Conference on Neural Information Processing Systems (NeurIPS) , vol. 34, pp. 17 981–17 993, 2021

  20. [28]

    The art of data augmentation,

    D. A. Van Dyk and X.-L. Meng, “The art of data augmentation,” Journal of Computational and Graphical Statistics , vol. 10, no. 1, pp. 1–50, 2001

  21. [29]

    A survey on image data augmen- tation for deep learning,

    C. Shorten and T. M. Khoshgoftaar, “A survey on image data augmen- tation for deep learning,” Journal of big data , vol. 6, no. 1, pp. 1–48, 2019

  22. [30]

    Learn- ing with noisy labels,

    N. Natarajan, I. S. Dhillon, P. K. Ravikumar, and A. Tewari, “Learn- ing with noisy labels,” in Annual Conference on Neural Information Processing Systems (NeurIPS) , vol. 26, 2013

  23. [31]

    Truncate-split-contrast: a framework for learning from mislabeled videos,

    Z. Wang, J. Weng, C. Yuan, and J. Wang, “Truncate-split-contrast: a framework for learning from mislabeled videos,” in AAAI Conference on Artificial Intelligence (AAAI) , vol. 37, no. 3, 2023, pp. 2751–2758

  24. [32]

    A software package for sequential quadratic programming,

    D. Kraft, “A software package for sequential quadratic programming,” Forschungsbericht- Deutsche Forschungs- und Versuchsanstalt fur Luft- und Raumfahrt, 1988

  25. [33]

    A computationally efficient feasible sequential quadratic programming algorithm,

    C. T. Lawrence and A. L. Tits, “A computationally efficient feasible sequential quadratic programming algorithm,” Siam Journal on optimiza- tion, vol. 11, no. 4, pp. 1092–1118, 2001

  26. [34]

    Sequential quadratic programming,

    P. T. Boggs and J. W. Tolle, “Sequential quadratic programming,” Acta numerica, vol. 4, pp. 1–51, 1995

  27. [35]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in International Conference on Medical Image Computing and Computer-Assisted Intervention (MIC- CAI). Springer, 2015, pp. 234–241

  28. [36]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Annual Con- ference on Neural Information Processing Systems (NeurIPS) , vol. 30, 2017

  29. [37]

    A note on the evaluation of generative models,

    L. Theis, A. v. d. Oord, and M. Bethge, “A note on the evaluation of generative models,” arXiv preprint arXiv:1511.01844 , 2015

  30. [38]

    On the evaluation of generative models in distributed learning tasks,

    Z. Wang, F. Farnia, Z. Lin, Y . Shen, and B. Yu, “On the evaluation of generative models in distributed learning tasks,” arXiv preprint arXiv:2310.11714, 2023. Zixiao Wang received his B.Eng. degree in Automa- tion from Tsinghua University (THU) in 2019 and an MSc.degree in Co...

  31. [2021]

    and Ph.D

    He received his M.Sc. and Ph.D. degrees in Electrical Engineering from Stanford University where he was a graduate research assistant at the Information Systems Laboratory advised by David Tse. He also received his B.Sc. degree in Electrical Engineering and Mathematics from Sh...

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.