Enhancing Stress-Strain Predictions with Seq2Seq and Cross-Attention based on Small Punch Test

Qixin Liu; Rui Yang; Weijian Han; Zhengni Yang

REVIEW 4 major objections 4 minor 24 references

Enhancing Stress-Strain Predictions with Seq2Seq and Cross-Attention based on Small Punch Test

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a Seq2Seq model with multi-head cross-attention and Gramian Angular Field features predicts true stress-strain curves of high-strength steels from small punch test data, with mean absolute errors between 0.15 and…

desk verdict A credible proof-of-concept on simulated small punch test data, but the paper oversells it as an experimental alternative without real-world validation. read the letter →

arxiv 2506.17680 v1 pith:5LBDZAGW submitted 2025-06-21 cs.LG cond-mat.mtrl-scics.AI

Zhengni Yang , Rui Yang , Weijian Han , Qixin Liu This is my paper · ORCID

classification cs.LGcond-mat.mtrl-scics.AI

keywords smallpunchteststress-straincurvepredictionsequence-to-sequenceLSTMcross-attentionGramianangularfieldhigh-strengthsteeldeeplearning

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

This paper attempts to establish that the relationship between small punch test load-displacement data and true stress-strain curves can be learned as a sequence-to-sequence translation task using an LSTM encoder-decoder enhanced with multi-head cross-attention. The authors generate 5,000 simulated high-strength steel samples with finite element analysis, encode the load-displacement sequence as a Gramian Angular Field image, and combine 1D and 2D convolutional features with the raw sequence to train the model. On a held-out set of 500 simulated samples, they report maximum and minimum mean absolute errors of 5.58 MPa and 0.15 MPa, with R-squared values between 0.986 and 0.999, compared against simpler LSTM, ANN, GRU, and Transformer baselines. If correct, this would let engineers obtain full stress-strain curves from a miniature punch test rather than machining tensile specimens, reducing material volume and cost for characterizing high-strength steels.

What carries the argument

The architecture is an LSTM-based sequence-to-sequence encoder-decoder whose input is a concatenated feature matrix $M = \text{Concat}(D, F^{1D}, F^{2D})$, where $D$ is the raw load-displacement sequence, $F^{1D}$ comes from 1D convolutions with kernel sizes 3, 5, and 7, and $F^{2D}$ comes from a 2D convolution on the Gramian Angular Field image. The GAF maps the sequence into a symmetric matrix by computing angles $\theta_i = \arccos(\text{normalized } x_i)$ and entries $G_{ij} = \cos(\theta_i + \theta_j)$, which preserves both order and magnitude. The decoder uses the encoder's hidden states through multi-head cross-attention, with attention weights $\alpha_{t,i}$ proportional to $H_i^\top O_t$, to generate each predicted stress-strain point, and a fully connected layer produces the final curve.

What would settle it

The claim would be falsified by running the trained model on experimental small punch test load-displacement data from a known high-strength steel and comparing the predicted stress-strain curve with a measured tensile curve: if the mean absolute error exceeds the reported 0.15 to 5.58 MPa range, or systematic bias appears in the strain-hardening region, the simulated-data transfer assumption fails.

Watch

Extended reading notes

Core claim

The central claim is that the mapping from small punch test load-displacement sequences to true stress-strain sequences is learnable with high accuracy when the input is enriched by both temporal and image-like features. In the paper's experiments, the proposed model achieves a maximum mean absolute error of 5.58 MPa and a minimum of 0.15 MPa on 500 held-out simulated samples, with R-squared values spanning 0.986 to 0.999. The paper attributes the improvement primarily to the Gramian Angular Field image representation, which preserves temporal order and magnitude, and to multi-head cross-attention, which aligns each decoder step with encoder hidden states. It positions this architecture as an alternative to conventional experimental techniques for determining true stress-strain relationships.

Load-bearing premise

Everything rests on the assumption that the finite-element simulations of the small punch test and the tensile test faithfully represent real high-strength steel behavior, so that a model trained on simulated pairs will also predict from experimental small punch data.

Editorial extensions

If this is right

Engineers could estimate full true stress-strain curves from small punch tests without machining tensile specimens, reducing material volume and cost for high-strength steel characterization.
The comparison results imply that adding GAF-based image features improves accuracy over a plain LSTM sequence model, lowering the maximum MAE from 17.44 MPa to 5.58 MPa in the tested setting.
The method treats SPT-to-stress-strain mapping as a translation task, opening the same architecture to other miniature mechanical tests where one load history is mapped to a material response curve.
The paper's future-work claims indicate the approach is intended to extend to other materials and testing conditions, with transfer learning as a route to unlabeled experimental data.

Reading between the lines

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

Editorial inference: the fixed elastic constants in the simulated data (Young's modulus 70,000 MPa, Poisson's ratio 0.35) mean the trained model is unlikely to generalize to steels with different elastic properties unless retrained or conditioned on those parameters.
Editorial inference: the reported error magnitudes should be interpreted relative to stress level, since a 5.58 MPa error is roughly 0.2% near 2,500 MPa but about 2.8% near 200 MPa; engineering acceptability will vary along the curve.
Editorial inference: a natural next test, not conducted in the paper, would be to use the model as a fast surrogate for inverse finite-element identification of stress-strain curves and compare its output against the standard inverse SPT procedure on the same experimental records.

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

The core of this paper is a sequence-to-sequence model with GAF image features and multi-scale convolutions that maps simulated small punch test load-displacement curves to true stress-strain curves. As a proof-of-concept on synthetic data, it works reasonably well: the proposed model beats the LSTM, ANN, GRU, and transformer baselines on the held-out split, with max MAE of 5.58 MPa and R^2 mostly above 0.98. That part is credible, and I give the authors credit for doing a decent ablation by comparing against architectures that lack the GAF or attention components.

What is genuinely new here is the specific combination of GAF, 1D/2D convolutions, LSTM Seq2Seq, and cross-attention for this task. I don't see that exact pipeline in the prior SPT literature. The data generation procedure is described at a level that makes the simulation setup understandable, though not fully reproducible without code.

The load-bearing weakness is exactly what the stress-test note says: every one of the 5,000 training and test samples comes from the same LS-DYNA model with fixed Young's modulus and Poisson's ratio. The test split is a random split from the same generator, so Table II measures in-distribution inversion of a known deterministic FEM mapping, not generalization to real materials. The abstract's phrase 'experimental results' is misleading; these are simulation results. Without a single experimental SPT-plus-tensile pair, the claim that this is an 'alternative to traditional experimental techniques' is unsupported.

There are smaller issues that a careful referee should flag. Table II only reports min/max MAE and R^2, not means, standard deviations, or repeated runs. Equation (6) defines a single-head dot-product attention, so the 'multi-head cross-attention' label is not backed by the math. The transformer baseline is cited as [5], which is Gehring et al.'s convolutional Seq2Seq, not a transformer. No code or data are provided, which makes the error statistics unverifiable.

The paper would benefit from either adding real experimental validation or repositioning itself as a simulation-only study with explicitly stated limitations. As it stands, it is a plausible engineering contribution with an overstated practical claim. I would send it to peer review because the methodology is coherent and the results on synthetic data are informative, but I would expect major revision to address the validation gap and the mislabeling.

Referee Report

4 major / 4 minor

Summary. The paper proposes a deep-learning pipeline that maps small punch test (SPT) load-displacement curves to true stress-strain curves for high-strength steels. The input sequence is transformed into a Gramian Angular Field (GAF) image, features are extracted with 1D and 2D convolutional layers, and the concatenated features are fed into an LSTM-based Seq2Seq model with a cross-attention mechanism. The model is trained and evaluated on 5,000 LS-DYNA finite-element simulation pairs (4,500 training, 500 testing) and compared against 1D LSTM, ANN, GRU, and Transformer baselines. The authors report minimum and maximum MAEs of 0.15 MPa and 5.58 MPa, respectively, with R² values between 0.986 and 0.999, and conclude that the method offers a promising alternative to traditional experimental tensile testing.

Significance. If the reported accuracy transferred to real experimental SPT data, the method would be practically valuable for miniature-sample mechanical characterization. The paper also provides a useful architecture comparison with four baselines on a moderately sized simulated dataset. However, the current evidence supports only that the network can invert a known deterministic FEM mapping within its training distribution; it does not support the central claim that the method is an alternative to traditional experimental techniques. The significance is therefore conditional on additional validation or a substantially reframed scope.

major comments (4)

[Sections II-A, III, and Abstract] The central claim that the method offers an alternative to traditional experimental techniques is not supported by the reported evidence. All 5,000 SPT/tensile pairs are generated by a single LS-DYNA simulation model, and the test set in Table II is a random split from the same generator. The reported MAE/R² values therefore measure how well the network inverts a known FEM mapping within its training distribution. No experimental SPT+tensile pair, no cross-solver comparison, and no perturbed-simulation validation is reported. The abstract and Section III use the phrase 'experimental results' for simulation output, which obscures this limitation. The authors should either add real experimental validation or explicitly reframe the claims to a simulation study.
[Section II-A, Table I] The simulation parameters appear inconsistent with the target material class. The paper states that the simulated material is high-strength steel, but uses Young's modulus of 70,000 MPa (70 GPa), which is closer to aluminum than to steel, and yield stresses as low as 20.98 MPa, which is far below typical high-strength steel values. This inconsistency undermines the implicit assumption that the trained model will transfer to real high-strength steels. The authors should correct the parameters, justify them, or restrict the claims to the simulated material range.
[Section II-C, Eq. (6)] The architecture is repeatedly described as 'multi-head cross-attention,' but Eq. (6) defines a single attention-weight distribution with no heads, no query/key/value projection matrices, and no description of how multiple heads would be combined. The contribution claimed in the introduction ('Seq2Seq model with multi-head cross-attention') is therefore not actually demonstrated by the equations or the algorithmic description. The authors should provide the full multi-head formulation, including the number of heads, or relabel the mechanism as single-head cross-attention.
[Section III, Table II and Algorithm 1] The statistical evaluation is incomplete. Table II reports only minimum and maximum MAE and R² values over an unspecified grouping of the 500 test samples, with no means, standard deviations, or number of repeated runs. Figure 5 shows only four representative test samples. In addition, the paper omits several hyperparameters required for reproducibility, such as learning rate, number of epochs, batch size, GAF image size, convolutional channel counts, sequence length, and the exact train/test split procedure. Without these, the claimed superiority over baselines cannot be independently assessed or reproduced.

minor comments (4)

[Abstract] There is a grammatical error: 'enhanced by multi-head cross-attention to improved accuracy' should be 'to improve accuracy.'
[References [18] and [5]] The GAF transformation is credited to reference [18], but that paper concerns stock index forecasting and is not the original source of Gramian Angular Fields (Wang and Oates, 2015). Also, the Transformer baseline is cited to [5], which is a convolutional sequence-to-sequence paper, not the standard Transformer reference; the authors should cite Vaswani et al. (2017) for the Transformer baseline.
[Section III and Conclusion] The conclusion claims 'computational efficiency,' but no runtime, parameter count, or computational-cost comparison is reported anywhere in Section III. This claim should be either substantiated or removed.
[Figure 5(d)] The discussion of Figure 5(d) states that the stress scale was expanded to 2500 MPa, which diminishes the apparent fluctuations of the ANN model; this visual scaling issue should be acknowledged more clearly so that readers do not overinterpret the figure.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: supervised fit on held-out simulated data; the weaknesses are external-validity and implementation-claim mismatches, not derivation circularity.

full rationale

This paper is a supervised machine-learning application rather than a derivation from first principles, so there is no derivation chain in which a predicted quantity is equivalent to an input by construction. The claimed accuracy (MAE 0.15–5.58 MPa, R² 0.986–0.999 in Table II) is measured on a held-out split of 500 samples from the same LS-DYNA generator used for the 4,500 training samples. That is a legitimate in-distribution interpolation test, not a fitted-input-called-prediction step, because the test samples are not used in training. The only self-citations ([20], [21]) are contextual references to prior fault-diagnosis and multi-fidelity SPT work; they do not supply a load-bearing theorem, ansatz, or uniqueness argument, so no self-citation chain forces the result. The main weakness is external validity, not circularity: both the input load-displacement curves and the target stress-strain curves are produced by one FEM simulator with fixed Young's modulus (70,000 MPa) and Poisson's ratio (0.35), and no experimental SPT–tensile pair is tested. The near-perfect errors therefore demonstrate interpolation within one simulation family, not experimentally verified transfer to real steel behavior. Separately, Eq. (6) defines single-head dot-product attention despite the paper's repeated 'multi-head cross-attention' label, which is an implementation/claim mismatch and a correctness concern, but not circularity. No specific reduction of a prediction to its own inputs, to a fitted parameter renamed as a prediction, or to a self-citation can be exhibited, so the circularity score is 0.

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

The central claim rests on the fidelity of FEM-generated training data, the adequacy of GAF as a feature encoding, and the representativeness of the simulated parameter ranges. No new physical entities are introduced.

free parameters (4)

Learned weights of the full network = not disclosed
All prediction performance comes from these weights fitted to the 4,500 simulated training samples.
LSTM hidden size and number of layers = 128 and 5
Hand-chosen architecture hyperparameters; no ablation study is reported.
1D convolution kernel sizes = 3, 5, 7
Chosen for multi-scale temporal feature extraction; no comparison to other kernel sets.
Optimizer, learning rate, loss function, sequence length, GAF image size, number of attention heads = not reported
Training details that would be needed to reproduce the results are absent from Section II and Algorithm 1.

assumptions (4)

domain assumption LS-DYNA FEM simulations faithfully represent real small punch test and tensile test behavior
Section II-A generates all 5,000 samples by simulation; no experimental validation is provided, so the transfer to real tests is assumed.
domain assumption GAF transformation preserves enough temporal and magnitude information for stress-strain recovery
Section II-B adopts GAF without ablation against other sequence encodings, so its sufficiency is assumed.
domain assumption The 4,500/500 train/test split is i.i.d. and representative of high-strength steel parameter ranges
Section III describes the ranges but no random seed, no check for duplicate parameter combinations, and no distribution shift analysis.
domain assumption The simulated material model (yield stress 20.98-1907.53 MPa, hardening exponent 0.068-0.4046) covers the target class of high-strength steels
Section II-A uses a single Young's modulus and Poisson's ratio for all samples, so variation across actual steel grades is not modeled.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Enhancing Stress-Strain Predictions with Seq2Seq and Cross-Attention based on Small Punch Test." pith.science (2026). https://pith.science/paper/5LBDZAGW

@misc{pith2026250617680,
  author       = {Pith},
  title        = {Pith review of: Enhancing Stress-Strain Predictions with Seq2Seq and Cross-Attention based on Small Punch Test},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5LBDZAGW}},
  note         = {Machine review of arXiv:2506.17680}
}

read the original abstract

This paper introduces a novel deep-learning approach to predict true stress-strain curves of high-strength steels from small punch test (SPT) load-displacement data. The proposed approach uses Gramian Angular Field (GAF) to transform load-displacement sequences into images, capturing spatial-temporal features and employs a Sequence-to-Sequence (Seq2Seq) model with an LSTM-based encoder-decoder architecture, enhanced by multi-head cross-attention to improved accuracy. Experimental results demonstrate that the proposed approach achieves superior prediction accuracy, with minimum and maximum mean absolute errors of 0.15 MPa and 5.58 MPa, respectively. The proposed method offers a promising alternative to traditional experimental techniques in materials science, enhancing the accuracy and efficiency of true stress-strain relationship predictions.

Figures

Figures reproduced from arXiv: 2506.17680 by the authors.

**Figure 1.** Structure of Proposed Approach A. Data Generation This study generated training and testing data for highstrength steel using LS-DYNA software to simulate loaddisplacement and stress-strain curves. The finite element method (FEM) modeled the SPT and tensile tests, accurately (a) Progressive Stages of Punch Impact on the Simulation Model (b) Progressive Stages of Tensile Test on the Simulation Model [PITH_FULL_IMA… view at source ↗

**Figure 3.** Load-Displacement Curves and Stress-Strain Curves of Training [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 5.** Figure 5: Results Comparison of Testing Samples and Transformer-based models, which lack a cross-attention mechanism, the maximum MAE values are 51.72 MPa and 450.77 MPa, while the minimum MAE values are 1.86 MPa and 6.25 MPa, respectively. Notably, the Transformer-based model p…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 24 canonical work pages

[5]

Convolutional sequence to sequence learning

Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N Dauphin. Convolutional sequence to sequence learning. InInternational conference on machine learning, pages 1243–1252. PMLR, 2017

work page 2017
[1]

Y . Cao, Y . Zu, Y . Zhen, F. Li, and G. Wu. Determination of the true stress-strain relations of high-grade pipeline steels based on small punch test correlation method. International Journal of Pressure Vessels and Piping, 199:104739, 2022

work page 2022
[2]

S. Chen, R. Yang, M. Zhong, X. Xi, and C. Liu. A random forest and model-based hybrid method of fault diagnosis for satellite attitude control systems.IEEE Transactions on Instrumentation and Measurement, 72: 1–13, 2023

work page 2023
[3]

Cornaggia, G

A. Cornaggia, G. Cocchetti, G. Maier, and V . Buljak. Inverse structural analyses on small punch tests, with model reduction and stochastic approach. In2018 IEEE International Conference on Environment and Electrical Engineering and 2018 IEEE Industrial and Commercial Power Systems Europe (EEEIC / I&CPS Europe), pages 1–5. IEEE, 2018

work page 2018
[4]

H. Deng, T. Zhang, Y . Dai, J. Shi, Y . Zhong, and H. Li. Deep non-rigid structure-from-motion: A sequence-to- sequence translation perspective.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12): 10814–10828, 2024

work page 2024
[6]

Georgiev, C

G. Georgiev, C. Dyer, and A. W. Black. Seq2seq models for end-to-end speech recognition. In2017 IEEE Auto- matic Speech Recognition and Understanding Workshop (ASRU), pages 317–323. IEEE, 2017

work page 2017
[7]

Mozhdeh Gheini, Xiang Ren, and Jonathan May. Cross- attention is all you need: Adapting pretrained trans- formers for machine translation.2021 Conference on Empirical Methods in Natural Language Processing, 11: 1754–1765, 2021

work page 2021
[8]

Hatakeyama-Sato

K. Hatakeyama-Sato. Machine learning for material science.The Brain & Neural Networks, 28(1):20–47, 2021

work page 2021

Show all 24 references

[9]

Jan ˇca, J

A. Jan ˇca, J. Siegl, and P. Hau ˇsild. Small punch test evaluation methods for material characterisation.Journal of Nuclear Materials, 481:201–213, 2016

2016
[10]

Kamaya and G

M. Kamaya and G. Lucas. Small punch testing for struc- tural materials: A review.Journal of Nuclear Materials, 510:255–269, 2018

2018
[11]

I. A. Klevtsov, D. N. Neshumaev, and A. V . Dedov. A method of using miniature samples for determin- ing mechanical properties of metal of power-generating equipment at thermal power stations in estonia.Thermal Engineering, 56:426–431, 2009

2009
[12]

K. Li, J. Peng, and C. Zhou. Construction of whole stress-strain curve by small punch test and inverse finite element.Results in Physics, 11:440–448, 2018

2018
[13]

Explainable sequence-to- sequence gru neural network for pollution forecasting

Sara Mirzavand Borujeni, Leila Arras, Vignesh Srini- vasan, and Wojciech Samek. Explainable sequence-to- sequence gru neural network for pollution forecasting. Scientific Reports, 13(1):9940, 2023

2023
[14]

M. Song, X. Li, Y . Cao, and S. Ma. Construction of true stress-strain curve of metallic material by artificial neural network and small punch test.Journal of Physics: Conference Series, 1676(1):012130, 2020

2020
[15]

Sterjovski, D

Z. Sterjovski, D. Nolan, K. R. Carpenter, D. P. Dunne, and J. Norrish. Artificial neural networks for modelling the mechanical properties of steels in various applica- tions.Journal of Materials Processing Technology, 170 (3):536–544, 2005

2005
[16]

Sutskever, O

I. Sutskever, O. Vinyals, and Q. V . Le. Sequence to sequence learning with neural networks.Advances in Neural Information Processing Systems, 3(7):3104–3112, 2014

2014
[17]

H. Wang, C. Zhang, B. Zhou, S. Xue, P. Jia, and X. Zhu. Prediction of triaxial mechanical properties of rocks based on mesoscopic finite element numerical simulation and multi-objective machine learning.Journal of King Saud University-Science, 35(7):102846, 2023

2023
[18]

Forecasting stock market in- dexes using principle component analysis and stochastic time effective neural networks.Neurocomputing, 156: 68–78, 2015

Jie Wang and Jun Wang. Forecasting stock market in- dexes using principle component analysis and stochastic time effective neural networks.Neurocomputing, 156: 68–78, 2015

2015
[19]

Y . Xue, R. Yang, X. Chen, Z. Tian, and Z. Wang. A novel local binary temporal convolutional neural network for bearing fault diagnosis.IEEE Transactions on Instru- mentation and Measurement, 72:1–13, 2023

2023
[20]

Z. Yang, R. Yang, and M. Huang. Rolling bearing incip- ient fault diagnosis method based on improved transfer learning with hybrid feature extraction.Sensors, 21(23): 7894, 2021

2021
[21]

Z. Yang, J. Zou, L. Huang, R. Yang, J. Zhang, C. Tong, J. Kong, Z. Zhan, and Q. Liu. Machine learning-based extraction of mechanical properties from multi-fidelity small punch test data.Advances in Manufacturing, 2195 (3597):1–14, 2025

2025
[22]

Zhan and H

Z. Zhan and H. Li. Machine learning based fatigue life prediction with effects of additive manufacturing process parameters for printed ss 316l.International Journal of Fatigue, 142:105941, 2021

2021
[23]

Zhang and A

D. Zhang and A. Subic. Advanced materials and their applications in small punch testing: A review.Materials & Design, 171:107734, 2019

2019
[24]

Zhang, R

Y . Zhang, R. Yang, Y . Yue, E. G. Lim, and Z. Wang. An overview of algorithms for contactless cardiac feature extraction from radar signals: Advances and challenges. IEEE Transactions on Instrumentation and Measurement, 72:1–20, 2023

2023

Pith tools

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

[1] [5]

Convolutional sequence to sequence learning

Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N Dauphin. Convolutional sequence to sequence learning. InInternational conference on machine learning, pages 1243–1252. PMLR, 2017

work page 2017

[2] [1]

Y . Cao, Y . Zu, Y . Zhen, F. Li, and G. Wu. Determination of the true stress-strain relations of high-grade pipeline steels based on small punch test correlation method. International Journal of Pressure Vessels and Piping, 199:104739, 2022

work page 2022

[3] [2]

S. Chen, R. Yang, M. Zhong, X. Xi, and C. Liu. A random forest and model-based hybrid method of fault diagnosis for satellite attitude control systems.IEEE Transactions on Instrumentation and Measurement, 72: 1–13, 2023

work page 2023

[4] [3]

Cornaggia, G

A. Cornaggia, G. Cocchetti, G. Maier, and V . Buljak. Inverse structural analyses on small punch tests, with model reduction and stochastic approach. In2018 IEEE International Conference on Environment and Electrical Engineering and 2018 IEEE Industrial and Commercial Power Systems Europe (EEEIC / I&CPS Europe), pages 1–5. IEEE, 2018

work page 2018

[5] [4]

H. Deng, T. Zhang, Y . Dai, J. Shi, Y . Zhong, and H. Li. Deep non-rigid structure-from-motion: A sequence-to- sequence translation perspective.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12): 10814–10828, 2024

work page 2024

[6] [6]

Georgiev, C

G. Georgiev, C. Dyer, and A. W. Black. Seq2seq models for end-to-end speech recognition. In2017 IEEE Auto- matic Speech Recognition and Understanding Workshop (ASRU), pages 317–323. IEEE, 2017

work page 2017

[7] [7]

Mozhdeh Gheini, Xiang Ren, and Jonathan May. Cross- attention is all you need: Adapting pretrained trans- formers for machine translation.2021 Conference on Empirical Methods in Natural Language Processing, 11: 1754–1765, 2021

work page 2021

[8] [8]

Hatakeyama-Sato

K. Hatakeyama-Sato. Machine learning for material science.The Brain & Neural Networks, 28(1):20–47, 2021

work page 2021