Few-shot Learning on AMS Circuits and Its Application to Parasitic Capacitance Prediction

Hector Rodriguez Rodriguez; Shan Shen; Wenjian Yu; Yibin Zhang

REVIEW 3 major objections 6 minor 30 references

Few-shot Learning on AMS Circuits and Its Application to Parasitic Capacitance Prediction

T0 review · 3 major / 6 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 A graph model pre-trained on a few AMS circuit netlists can predict parasitic coupling and capacitance on unseen designs, with accuracy gains of at least 20% over existing methods.

desk verdict Solid empirical EDA paper with a genuinely new task formulation, but the headline zero-shot numbers need a cleaner evaluation before they're taken at face value. read the letter →

arxiv 2507.06538 v1 pith:A5KBHECW submitted 2025-07-09 cs.LG cs.SYeess.SY

Shan Shen , Yibin Zhang , Hector Rodriguez Rodriguez , Wenjian Yu This is my paper

classification cs.LGcs.SYeess.SY

keywords few-shotlearningparasiticcapacitanceanalog/mixed-signalcircuitsheterogeneousgraphlinkpredictiontransformersubgraphsamplingpositionalencoding

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

CircuitGPS tries to show that the data scarcity problem in analog/mixed-signal design can be overcome by a few-shot graph learning approach. It converts a circuit netlist into a heterogeneous graph whose nodes are nets, pins, and devices, and treats each parasitic coupling as a link to be predicted. Around every link it extracts a 1-hop enclosing subgraph, encodes positions with a cheap distance-based scheme (DSPD), and pre-trains a hybrid graph Transformer on link prediction before fine-tuning on capacitance regression. The reported result is that on three unseen test circuits this pipeline improves coupling-existence accuracy by at least 20% and cuts capacitance-estimation MAE by at least 0.067 compared with the two existing methods. If correct, designers could reuse one pre-trained model for new circuits instead of re-extracting and re-labeling large datasets.

What carries the argument

The load-bearing mechanism is the small-hop enclosing subgraph extracted around each target link in the heterogeneous circuit graph, together with the double-anchor shortest path distance (DSPD) encoding. The enclosing subgraph is the induced subgraph of nodes within h hops of either anchor node; the paper uses h=1 as a balance between information and scalability. DSPD records, for every node in the subgraph, its shortest-path distances to the two anchor nodes, concatenates those distances with node-type embeddings, and feeds them into a parallel MPNN-plus-Transformer layer; this encoding supplies positional and structural information without the cost of Laplacian eigendecomposition. The subgraph representation is what makes few-shot transfer possible, because the target link is represented by a local, task-relevant structure instead of by its position in a huge unique circuit graph.

What would settle it

Take a circuit where the dominant coupling is between two widely separated schematic nodes that run parallel in the layout (for example, two long signal nets on adjacent metal layers), and test the trained model on it. If the model's coupling-existence accuracy or capacitance MAE degrades sharply compared with layout-aware prediction, the assumption that the 1-hop schematic subgraph suffices is false.

Watch

Extended reading notes

Core claim

The central claim is that coupling existence and coupling capacitance in AMS circuits can be predicted in a zero-shot setting from schematic-derived graphs, using subgraph-level representations rather than entire-circuit models. The paper argues that the key to transferability is to decouple each target link from the whole circuit by sampling its 1-hop enclosing subgraph, then to learn a meta-learner on link prediction; the same pre-trained model is then either tested directly or fine-tuned for edge regression. It further claims that the proposed DSPD encoding outperforms other positional encodings, and that including raw circuit statistics as node features during link-prediction pre-training hurts generalization, so those statistics are re-injected only at the task-specific head. On the three zero-shot test circuits, the reported gains relative to ParaGraph and DLPL-Cap are at least 20% in accuracy and at least 0.067 in MAE reduction.

Load-bearing premise

The paper assumes that a 1-hop slice of the schematic graph around two nodes carries enough information to decide whether they are capacitively coupled and how large that capacitance is, so layout effects that reach beyond that local topology can be ignored.

Editorial extensions

If this is right

A pre-trained CircuitGPS meta-learner can be applied directly to unseen AMS circuits for coupling detection, so new designs may not need labeled post-layout data for that task.
The same subgraph representation transfers across three prediction tasks, so a single pre-trained model could cover coupling classification, coupling capacitance regression, and node-level ground capacitance estimation.
A cheap positional encoding (DSPD) achieves accuracy at least matching expensive Laplacian encodings, which makes the approach scale to million-node circuit graphs.
Classic message-passing layers alone are competitive with, or better than, Transformer layers on these circuit tasks, which simplifies the model and reduces training cost.
Fine-tuning the full model on a small regression dataset improves capacitance prediction further, meaning the few-shot pipeline has a usable adaptation path beyond zero-shot testing.

Reading between the lines

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

The comparison baselines (ParaGraph, DLPL-Cap) do not use subgraph sampling or positional encodings, so part of the reported margin may come from those components rather than from few-shot pre-training per se; a SEAL-style baseline with the same sampling would isolate the pre-training contribution.
Because all datasets are 28nm SRAM-family circuits, the zero-shot claim is demonstrated within one technology and design family; whether the model transfers across technology nodes or to non-SRAM analog blocks is an open question.
The observation that circuit statistics hurt link prediction but help regression suggests a general recipe: strip noisy node features during pre-training, then re-inject them in the task head for relational regression problems.
A testable extension is to add layout-aware geometry (wire coordinates, metal layers) as auxiliary node/edge features while keeping the schematic subgraph, which could close the gap to layout-driven coupling not visible in 1-hop schematic topology.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Colleague,

The paper has a genuinely new angle: it's the first in this line of work to treat parasitic coupling as a link-prediction problem on the schematic graph, with subgraph sampling and a cheap positional encoding (DSPD) that is a minor variant of DRNL. The empirical gains are large and consistent across three unseen circuits, and the ablations are useful—especially the finding that GatedGCN alone beats fancier transformers. Code and data are public, and the SPICE energy validation is a nice touch.

The soft spots are real but not fatal. The most concrete one is that the authors use DIGITAL_CLK_GEN, one of the three test circuits, to choose the positional encoding (Table II) and the layer architecture (Tables III and VII). That is outcome-dependent model selection on a test set, and it inflates the numbers for that circuit. The other two test circuits are untouched, so the generalization claim still has support, but the headline “at least 20%” is partly selected on the test set. A re-run with architecture selection on training validation only would fix this.

The stress-test about negative links is worth taking seriously. Generating negatives by permuting endpoints of positive links creates pairs from unrelated device instances, which are easier to separate than realistic non-coupled pairs that are physically close. The model may be learning “different device neighborhood” rather than the physics of coupling. Because the baselines see the same negatives, the relative comparison is fair, but the absolute accuracy on real layouts could be lower. This isn't a fatal flaw, but it means the zero-shot claim should be qualified.

The capacitance range filter (10^-21 to 10^-15 F) trims the regression target, so the reported MAE reduction is for a restricted set. And the “few-shot” label is loose—the zero-shot results are plain transfer learning with no support-set adaptation.

Overall: the core method is plausible, the gains are large, and the limitations are the usual EDA-paper sins rather than a broken central argument. The paper deserves serious refereeing. I'd ask for a revised version that re-selects hyperparameters without touching the test set and adds a harder negative-sampling experiment.

Referee Report

3 major / 6 minor

Summary. The paper proposes CircuitGPS, a few-shot learning framework for parasitic effect prediction in AMS circuits. Circuit netlists are converted into heterogeneous graphs, coupling capacitances are treated as links, and a small-hop enclosing subgraph sampling converts each link into a graph-level instance. The model is a hybrid graph Transformer (built on GraphGPS) with a new positional encoding called DSPD (double-anchor shortest path distance). The approach is pre-trained on link prediction and then fine-tuned for edge (capacitance) regression and also applied to node regression. Experiments on three AMS test circuits report large improvements over ParaGraph and DLPL-Cap, with the abstract claiming at least 20% higher coupling-existence accuracy and at least 0.067 lower capacitance MAE, all asserted to be zero-shot.

Significance. If the reported results hold, the work offers a practical method for predicting layout-derived parasitic effects from schematic netlists with limited labeled data, which is a real bottleneck in AMS design. The paper also contributes DSPD as a computationally cheap positional encoding and provides ablation results that clarify the relative strengths of MPNNs and Transformers on circuit graphs. The public code repository, configuration files, and the use of multiple distinct AMS designs are strengths for reproducibility. However, the central empirical claims depend on two evaluation choices that currently weaken their validity: the synthetic negative-link generation and the use of one test circuit for both model selection and final evaluation. These need to be addressed before the contributions can be fully accepted.

major comments (3)

[Section III-B (Negative link generation)] The negative links are formed by swapping the endpoints of existing positive links, e.g., from positive links (i,j) and (m,n) the negatives are (i,n) and (m,j). This creates a synthetic negative distribution that is not representative of genuinely uncoupled net/pin pairs in a real layout, where many uncoupled pairs share similar local neighborhoods and are spatially proximate. Since the model is both trained and evaluated on this same synthetic negative distribution, the large accuracy and AUC improvements in Table V may reflect the model learning to discriminate real couplings from arbitrary endpoint swaps rather than learning the physical rules of coupling existence. This directly affects the abstract's claim of improving coupling-existence accuracy by at least 20% and the zero-shot generalization claim. The authors should re-run the link prediction experiments with a negative set drawn from actual non-coupled pairs that are realistic (e.g., pairs of nets/pins in the same circuit region but absent from the post-layout SPF coupling list) and report metrics on that set, ideally alongside the synthetic-negative results.
[Section IV-B and Tables II, III, VII] The DSPD positional encoding and the GatedGCN layer type are selected by ablation studies that use SSRAM as training data and DIGITAL_CLK_GEN as the test/validation circuit (stated before Table II and in Table III's caption). The same DIGITAL_CLK_GEN circuit is then reported as one of the three zero-shot test datasets in Tables V, VI, and VIII. This means the architectural configuration was chosen using the performance on a benchmark that later serves as a test case, so the results for DIGITAL_CLK_GEN are not a blind zero-shot evaluation. The statement 'all results in this work are reported from zero-shot learning' is therefore misleading for at least one of the headline test circuits. The authors should either select hyperparameters and layer choices using only training-circuit validation, or clearly separate the configuration-selection results from the final zero-shot results, and re-evaluate the aggregate claims without treating DIGITAL_CLK_GEN as an unbiased test case.
[Section IV-C (Edge Regression Task)] The capacitance regression is restricted to links with yC in [1e-21 F, 1e-15 F] (the text has a typo '1−21F'), while negative links from pre-training are assigned zero capacitance and are kept. This filtering changes the regression task into a hybrid of zero/nonzero classification and narrow-range regression. The reported MAE reductions in Table VI (e.g., 0.153 to 0.083) are therefore only valid within this filtered distribution, and the paper does not state what fraction of positive links are removed by the filter or how the zero-capacitance negatives are counted in the regression error. The authors should report the retained-link fraction, the capacitance distribution after filtering, and ideally show sensitivity to the filter bounds or separate the evaluation of positive and negative links to demonstrate that the claimed MAE improvement is not an artifact of the filtering choice.

minor comments (6)

[Abstract and Section IV-B] The phrase 'improves the accuracy of coupling existence by at least 20%' should clarify whether this is an absolute percentage-point increase or a relative improvement; Table V shows absolute increases of about 0.20–0.26, while relative increases are much larger. The same applies to 'reduces the MAE by at least 0.067', which seems to refer to a specific comparison and not all rows of Table VI.
[Table IV] In the dataset statistics table, the columns 'N' and 'N E' are not defined in the caption or in the surrounding text; the authors should spell out that these are the numbers of nodes and edges, respectively, and explain the notation for N/G1m,n and NE/G1m,n.
[Section IV-C] The expression '1−21F ≤ yC ≤ 1−15F' appears to be a formatting error; it should be '1e-21 F ≤ yC ≤ 1e-15 F'. Also, 'we only kept the nets' should probably be 'links' if it refers to the target coupling links.
[Throughout] The name 'DIGIT AL CLK GEN' appears with an unexpected space in multiple places; it should be 'DIGITAL_CLK_GEN' to match the dataset naming convention used in the figures and tables.
[Section III-B] The paper states that 'the 1-hop sampling achieves the best balance between performance and efficiency' but provides no ablation over the hop count h. Given that the choice of h is a key hyperparameter tied to the SEAL theory, a small ablation (e.g., h=1 vs h=2) on the training circuits would strengthen the justification.
[Table VII] In the layer-configuration ablation, the entry 'GatedGCN None' means an MPNN without an attention component; adding a brief explanation of the 'None' notation in the caption would improve readability.

Circularity Check

1 steps flagged · score 4.0 of 10

Partial circularity: the DSPD and GatedGCN configuration are selected using the DIGITAL_CLK_GEN test circuit, and the same circuit's results are then reported as zero-shot predictions.

fitted input called prediction [Section III-C (Table II) and Section IV-C (Tables V, VI, VII)]
"We take the AMS design SSRAM [23] as a small training dataset and the circuit DIGIT AL CLK GEN as a test dataset to evaluate CircuitGPS in a zero-shot learning setting. The proposed DSPD demonstrates the highest accuracy while ensuring computational efficiency. ... We also conducted an ablation study of different GPS layer configurations for edge regression on the small training set SSRAM and the test dataset DIGIT AL CLK GEN."

The positional encoding (DSPD) and the MPNN/attention layer configuration are chosen by comparing their link-prediction and regression metrics on the DIGITAL_CLK_GEN test circuit (Tables II, III, VII). The same circuit then appears as a headline zero-shot result in Tables V and VI. The reported accuracy and MAE for DIGITAL_CLK_GEN are therefore in-sample selection results: the model family was selected to optimize the very metric later presented as an out-of-sample prediction. This is a fitted hyperparameter (architecture/PE choice) renamed as a zero-shot prediction. The other two test circuits, TIMING_CONTROL and ARRAY_128_32, were not used for this selection, so the circularity is partial rather than total.

full rationale

Apart from the configuration-selection issue on DIGITAL_CLK_GEN, the core derivation is self-contained. CircuitGPS is pre-trained on SSRAM, ULTRA8T, and SANDWICH-RAM and evaluated on three circuits withheld from training, so the main accuracy/MAE improvements are not derived from the quantities they predict. Self-citations (Refs. [19], [23], [29]) appear as a baseline method, dataset sources, and a class-imbalance observation; none is load-bearing in the sense of forcing the reported numerical results. The synthetic negative-link generation (Section III-B) is a potential external-validity concern about whether 'coupling absence' is representative of physically non-coupled pairs, but it does not make a prediction equal to its input by construction; it is a benchmark-construction issue rather than a circularity. The one genuine circular pattern is the use of the DIGITAL_CLK_GEN test circuit to select DSPD and the GatedGCN layer before reporting zero-shot results on that same circuit. Because the other two test circuits are independent, the central claim retains independent support.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced; DSPD is a methodological encoding, not a postulated object. The central result rests on modeling choices about subgraph size, negative-link validity, and the sufficiency of schematic-level features.

free parameters (3)

Subgraph sampling hop count h = 1
Chosen as the best balance of performance and efficiency (Section III-B); the central model receives only 1-hop neighborhoods, so this choice shapes all results.
Capacitance regression range = 1e-21 F to 1e-15 F
Hand-selected inclusion window for regression targets (Section IV-C); extreme capacitances are excluded before MAE and R2 are computed.
Link-type balancing sample count = |En2n| instances per link type
To mitigate class imbalance, the authors randomly sample |En2n| instances from each link type (Section III-B), which changes the training distribution.

assumptions (4)

domain assumption Post-layout SPF netlists provide complete and correct ground-truth coupling links and capacitance values.
Section IV-A states ground truth is collected from SPF files; if extraction is incomplete, labels are wrong.
domain assumption The schematic graph plus device statistics (transistor width, length, fingers, etc.) carries the information needed to predict parasitic coupling.
The prediction model never sees layout coordinates; the paper assumes schematic and local statistics suffice.
standard math The gamma-decaying theory of SEAL justifies using a small (1-hop) enclosing subgraph for link prediction in circuit graphs.
Invoked in Section III-B to justify the 1-hop sampling choice.
domain assumption Negative links formed by permuting the endpoints of positive links are valid examples of absent coupling.
Section III-B defines negative link generation; false negatives would bias pretraining.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Few-shot Learning on AMS Circuits and Its Application to Parasitic Capacitance Prediction." pith.science (2026). https://pith.science/paper/A5KBHECW

@misc{pith2026250706538,
  author       = {Pith},
  title        = {Pith review of: Few-shot Learning on AMS Circuits and Its Application to Parasitic Capacitance Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A5KBHECW}},
  note         = {Machine review of arXiv:2507.06538}
}

read the original abstract

Graph representation learning is a powerful method to extract features from graph-structured data, such as analog/mixed-signal (AMS) circuits. However, training deep learning models for AMS designs is severely limited by the scarcity of integrated circuit design data. In this work, we present CircuitGPS, a few-shot learning method for parasitic effect prediction in AMS circuits. The circuit netlist is represented as a heterogeneous graph, with the coupling capacitance modeled as a link. CircuitGPS is pre-trained on link prediction and fine-tuned on edge regression. The proposed method starts with a small-hop sampling technique that converts a link or a node into a subgraph. Then, the subgraph embeddings are learned with a hybrid graph Transformer. Additionally, CircuitGPS integrates a low-cost positional encoding that summarizes the positional and structural information of the sampled subgraph. CircuitGPS improves the accuracy of coupling existence by at least 20\% and reduces the MAE of capacitance estimation by at least 0.067 compared to existing methods. Our method demonstrates strong inherent scalability, enabling direct application to diverse AMS circuit designs through zero-shot learning. Furthermore, the ablation studies provide valuable insights into graph models for representation learning.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 20 canonical work pages

[1]

Generalizing from a few examples: A survey on few-shot learning,

Y . Wang, Q. Yao, J. T. Kwok, and L. M. Ni, “Generalizing from a few examples: A survey on few-shot learning,” ACM computing surveys (csur), vol. 53, no. 3, pp. 1–34, 2020

2020
[2]

Pretraining graph neural networks for few-shot analog circuit modeling and design,

K. Hakhamaneshi, M. Nassar, M. Phielipp, P. Abbeel, and V . Sto- janovic, “Pretraining graph neural networks for few-shot analog circuit modeling and design,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 42, no. 7, pp. 2163–2173, 2022

work page 2022
[3]

Yu and X

W. Yu and X. Wang, Advanced field-solver techniques for RC extraction of integrated circuits. Springer, 2014

work page 2014
[4]

Variational capacitance extraction of on- chip interconnects based on continuous surface model,

W. Yu, C. Hu, and W. Zhang, “Variational capacitance extraction of on- chip interconnects based on continuous surface model,” in Proceedings of the 46th Annual Design Automation Conference, 2009, pp. 758–763

work page 2009
[5]

Link prediction based on graph neural networks,

M. Zhang and Y . Chen, “Link prediction based on graph neural networks,” Advances in neural information processing systems, vol. 31, 2018

2018
[6]

Recipe for a General, Powerful, Scalable Graph Trans- former,

L. Ramp ´aˇsek, M. Galkin, V . P. Dwivedi, A. T. Luu, G. Wolf, and D. Beaini, “Recipe for a General, Powerful, Scalable Graph Trans- former,” Advances in Neural Information Processing Systems , vol. 35, 2022

work page 2022
[7]

Optimization as a model for few-shot learning,

S. Ravi and H. Larochelle, “Optimization as a model for few-shot learning,” in International conference on learning representations , 2017

work page 2017
[8]

Weisfeiler-lehman neural machine for link prediction,

M. Zhang and Y . Chen, “Weisfeiler-lehman neural machine for link prediction,” in Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining , 2017, pp. 575– 583

work page 2017

Show all 30 references

[9]

Link prediction in complex networks: A survey,

L. L ¨u and T. Zhou, “Link prediction in complex networks: A survey,” Physica A: statistical mechanics and its applications , vol. 390, no. 6, pp. 1150–1170, 2011

2011
[10]

Ammus: A survey of transformer-based pretrained models in natural language processing,

K. S. Kalyan, A. Rajasekharan, and S. Sangeetha, “Ammus: A survey of transformer-based pretrained models in natural language processing,” arXiv:2108.05542, 2021

2021 arXiv
[11]

A survey on vision transformer,

K. Han, Y . Wang, H. Chen, X. Chen, J. Guo, Z. Liu, Y . Tang, A. Xiao, C. Xu, Y . Xu et al., “A survey on vision transformer,” IEEE Transactions on Pattern Analysis and Machine Intelligence , 2022

2022
[12]

A generalization of transformer networks to graphs,

V . P. Dwivedi and X. Bresson, “A generalization of transformer networks to graphs,” arXiv:2012.09699, 2020

2012 arXiv
[13]

Rethinking graph transformers with spectral attention,

D. Kreuzer, D. Beaini, W. L. Hamilton, V . L ´etourneau, and P. Tossou, “Rethinking graph transformers with spectral attention,” in Advances in Neural Information Processing Systems , 2021

2021
[14]

Do transformers really perform badly for graph representation?

C. Ying, T. Cai, S. Luo, S. Zheng, G. Ke, D. He, Y . Shen, and T.-Y . Liu, “Do transformers really perform badly for graph representation?” in Advances in Neural Information Processing Systems , 2021

2021
[15]

Graph neural networks with learnable structural and positional repre- sentations,

V . P. Dwivedi, A. T. Luu, T. Laurent, Y . Bengio, and X. Bresson, “Graph neural networks with learnable structural and positional repre- sentations,” in International Conference on Learning Representations , 2022

2022
[16]

Directional graph networks,

D. Beaini, S. Passaro, V . L ´etourneau, W. Hamilton, G. Corso, and P. Li `o, “Directional graph networks,” in International Conference on Machine Learning. PMLR, 2021, pp. 748–758

2021
[17]

Understanding over-squashing and bottlenecks on graphs via curvature,

J. Topping, F. Di Giovanni, B. P. Chamberlain, X. Dong, and M. M. Bronstein, “Understanding over-squashing and bottlenecks on graphs via curvature,” arXiv:2111.14522, 2021

2021 arXiv
[18]

ParaGraph: Layout parasitics and device parameter prediction using graph neural networks,

H. Ren, G. F. Kokai, W. J. Turner, and T.-S. Ku, “ParaGraph: Layout parasitics and device parameter prediction using graph neural networks,” in Proc. DAC, 2020, pp. 1–6

2020
[19]

Deep-learning- based pre-layout parasitic capacitance prediction on sram designs,

S. Shen, D. Yang, Y . Xie, C. Pei, B. Yu, and W. Yu, “Deep-learning- based pre-layout parasitic capacitance prediction on sram designs,” in Proceedings of the Great Lakes Symposium on VLSI 2024 , 2024, pp. 440–445

2024
[20]

Rethinking attention with performers,

K. M. Choromanski, V . Likhosherstov, D. Dohan, X. Song, A. Gane, T. Sarlos, P. Hawkins, J. Q. Davis, A. Mohiuddin, L. Kaiser et al. , “Rethinking attention with performers,” in International Conference on Learning Representations , 2020

2020
[21]

Benchmarking graph neural networks,

V . P. Dwivedi, C. K. Joshi, T. Laurent, Y . Bengio, and X. Bresson, “Benchmarking graph neural networks,” arXiv:2003.00982, 2020

2003 arXiv
[22]

Com- paring graph transformers via positional encodings,

M. Black, Z. Wan, G. Mishne, A. Nayyeri, and Y . Wang, “Com- paring graph transformers via positional encodings,” arXiv preprint arXiv:2402.14202, 2024

2024 arXiv
[23]

TS cache: A fast cache with timing-speculation mechanism under low supply voltages,

S. Shen, T. Shao, X. Shang, Y . Guo, M. Ling, J. Yang, and L. Shi, “TS cache: A fast cache with timing-speculation mechanism under low supply voltages,” IEEE Transactions on Very Large Scale Integration (VLSI) Systems, vol. 28, no. 1, pp. 252–262, 2019

2019
[24]

Structural information enhanced graph representation for link prediction,

L. Shi, B. Hu, D. Zhao, J. He, Z. Zhang, and J. Zhou, “Structural information enhanced graph representation for link prediction,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, no. 13, 2024, pp. 14 964–14 972

2024
[25]

Residual gated graph convnets,

X. Bresson and T. Laurent, “Residual gated graph convnets,” arXiv preprint arXiv:1711.07553, 2017

2017 arXiv
[26]

Classic gnns are strong base- lines: Reassessing gnns for node classification,

Y . Luo, L. Shi, and X.-M. Wu, “Classic gnns are strong base- lines: Reassessing gnns for node classification,” arXiv preprint arXiv:2406.08993, 2024

2024 arXiv
[27]

Fast graph representation learning with PyTorch Geometric,

M. Fey and J. E. Lenssen, “Fast graph representation learning with PyTorch Geometric,” in ICLR Workshop on Representation Learning on Graphs and Manifolds , 2019

2019
[28]

Design space for graph neural networks,

J. You, R. Ying, and J. Leskovec, “Design space for graph neural networks,” in Advances in Neural Information Processing Systems , 2020

2020
[29]

Ultra8t: A sub-threshold 8t sram with leakage detection,

S. Shen, H. Xu, Y . Zhou, M. Ling, and W. Yu, “Ultra8t: A sub-threshold 8t sram with leakage detection,” Integration, vol. 98, p. 102233, 2024

2024
[30]

24.4 sandwich-RAM: An energy-efficient in-memory BWN architecture with pulse-width modulation,

J. Yang, Y . Kong, Z. Wang, Y . Liu, B. Wang, S. Yin, and L. Shi, “24.4 sandwich-RAM: An energy-efficient in-memory BWN architecture with pulse-width modulation,” in Proc. Int. Solid-State Circuits Conf. (ISSCC), 2019, pp. 394–396

2019

Pith tools

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

[1] [1]

Generalizing from a few examples: A survey on few-shot learning,

Y . Wang, Q. Yao, J. T. Kwok, and L. M. Ni, “Generalizing from a few examples: A survey on few-shot learning,” ACM computing surveys (csur), vol. 53, no. 3, pp. 1–34, 2020

2020

[2] [2]

Pretraining graph neural networks for few-shot analog circuit modeling and design,

K. Hakhamaneshi, M. Nassar, M. Phielipp, P. Abbeel, and V . Sto- janovic, “Pretraining graph neural networks for few-shot analog circuit modeling and design,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 42, no. 7, pp. 2163–2173, 2022

work page 2022

[3] [3]

Yu and X

W. Yu and X. Wang, Advanced field-solver techniques for RC extraction of integrated circuits. Springer, 2014

work page 2014

[4] [4]

Variational capacitance extraction of on- chip interconnects based on continuous surface model,

W. Yu, C. Hu, and W. Zhang, “Variational capacitance extraction of on- chip interconnects based on continuous surface model,” in Proceedings of the 46th Annual Design Automation Conference, 2009, pp. 758–763

work page 2009

[5] [5]

Link prediction based on graph neural networks,

M. Zhang and Y . Chen, “Link prediction based on graph neural networks,” Advances in neural information processing systems, vol. 31, 2018

2018

[6] [6]

Recipe for a General, Powerful, Scalable Graph Trans- former,

L. Ramp ´aˇsek, M. Galkin, V . P. Dwivedi, A. T. Luu, G. Wolf, and D. Beaini, “Recipe for a General, Powerful, Scalable Graph Trans- former,” Advances in Neural Information Processing Systems , vol. 35, 2022

work page 2022

[7] [7]

Optimization as a model for few-shot learning,

S. Ravi and H. Larochelle, “Optimization as a model for few-shot learning,” in International conference on learning representations , 2017

work page 2017

[8] [8]

Weisfeiler-lehman neural machine for link prediction,

M. Zhang and Y . Chen, “Weisfeiler-lehman neural machine for link prediction,” in Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining , 2017, pp. 575– 583

work page 2017