Pith. sign in

REVIEW 3 major objections 6 minor 47 references

Overcoming Labelled Data Scarcity for Defect Classification in Scanning Tunneling Microscopy

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A two-stage pipeline combining unsupervised segmentation with few-shot classification can classify atomic-scale defects in scanning tunnelling microscopy images of a new surface using as few as one labelled example per class.

desk verdict Useful, honest applied FSL benchmark for STM, but the Si(001) one-shot result is undermined by train/test augmentation leakage and the abstract oversells the accuracy. read the letter →

arxiv 2506.01678 v2 pith:PMFI2YIO submitted 2025-06-02 cond-mat.mtrl-sci cs.AI

Nikola L. Kolev , Max Trouton , Filippo Federici Canova , Geoff Thornton , David Z. Gao , Neil J. Curson , Taylor J. Z. Stock This is my paper
classification cond-mat.mtrl-scics.AI
keywords scanningtunnelingmicroscopyfew-shotlearningdefectclassificationunsupervisedsegmentationU-NetprototypicalnetworksurfacescienceSTMimageanalysis
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

Scanning tunnelling microscopy produces atomic-resolution images whose routine analysis, finding and labelling defects, normally requires large manually annotated datasets for each new surface. This paper claims that a two-stage pipeline can remove that requirement: an unsupervised module generates coarse defect masks, and a few-shot classifier then sorts the masked defects into user-defined classes using only one or three labelled examples per class. The authors demonstrate the approach on three surfaces, Si(001), Ge(001), and TiO2(110), reporting one-shot classification accuracies up to 93%, 62%, and 70% respectively, and total dice scores of 0.96 to 0.99 for the segmentation stage. The point of the work is that a model trained once can be adapted to a previously unseen substrate with only a handful of clicks, making automated STM analysis practical for new materials and adsorbates.

What carries the argument

The load-bearing object is the two-module pipeline, not a single network. Module 1 produces training labels without human annotation: it runs an ImageNet-pretrained fully convolutional network over the STM image, clusters the resulting per-pixel feature vectors with k-means, and thresholds the clusters into a binary defect/lattice mask. A U-Net is then trained on those masks so that inference runs quickly on a CPU. Module 3 is the few-shot classifier, using prototypical networks, which embed each defect crop in a learned vector space and classify a query crop by its distance to the mean embedding, or prototype, of each class's K support examples; the support set is chosen by the user from a handful of crops. The argument turns on this split: the unsupervised module supplies the where, the few-shot module supplies the what, and only the latter requires new labels.

What would settle it

On a fourth surface with a known, manually counted defect inventory, run the full pipeline and compare the final classified feature map with the manual count. If the U-Net's defect-only dice score lands near the 0.64 reported for Si(001), more than a third of defects would never reach the few-shot classifier, and the one-shot adaptation claim would fail as a complete segmentation despite high classifier accuracy.

Watch

Extended reading notes

Core claim

The central claim is that defect classification in STM can be decomposed so that the expensive part, learning what a defect looks like, is done once on well-studied surfaces, while adapting to a new surface costs almost nothing in labels. On germanium and titania, where neither substrate appeared in training, models classify defects with 61 to 70 percent one-shot accuracy; on a held-out set of silicon adsorbate features, the same approach reaches 93 percent. The paper further claims that the segmentation bottleneck can be automated: a pretrained feature extractor followed by k-means clustering produces binary defect maps for a new substrate, and a standard U-Net distilled from those maps segments full scans on a CPU in seconds with total dice scores between 0.96 and 0.99.

Load-bearing premise

The load-bearing premise is that the unsupervised labelling module finds most true defects on any new substrate, because the few-shot classifier can only sort defects that the first pass did not miss.

Editorial extensions

If this is right

  • Researchers studying a new adsorbate on a known substrate can count and classify defects with one to three labelled examples per class, with accuracies that support surface-chemistry statistics on silicon but are weaker on germanium and titania.
  • A new substrate needs only about five scans and an unsupervised labelling pass to train the segmentation stage, not thousands of manually annotated images.
  • Because the support set can be redefined on the fly, a user can re-label classes during an experiment to account for tip changes and exclude anomalous features, which fixed-class supervised models cannot do.
  • Among the tested algorithms, prototypical and matching networks outperform relation and simple-shot networks, and synthetic inverse defects added to training classes improve accuracy by up to 16 percent, giving a labelling-free route to better embeddings.
  • The unsupervised labelling step transfers to other microscopy modes, as the authors show by applying it to a STEM image and obtaining pixel-level segmentation comparable to a published few-shot patch-based method.

Reading between the lines

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

  • Beyond the paper's own claims, the weak point to watch is the U-Net's defect-only dice score: on Si(001) it is 0.64, meaning roughly a third of true defect pixels never reach the classifier, so the final feature map is incomplete regardless of FSL accuracy.
  • On a genuinely new surface, the likely failure point is this under-detection in the unsupervised first pass rather than the few-shot classifier, so testing the pipeline on a fourth surface with a known defect inventory would be the natural next experiment.
  • The one-shot claim would be stronger if support sets were drawn from the same scan or imaging conditions as the query data; the paper's test sets may share more with training than a true first encounter would.
  • The same architecture could be driven by active learning, where the system asks the user to label only the most uncertain crops, potentially raising germanium and titania accuracies toward the silicon result.
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

3 major / 6 minor

Summary. The paper presents a modular pipeline for automated segmentation and classification of atomic-scale features in STM images. It combines an unsupervised labelling module (using ImageNet-pretrained FCN features with k-means clustering) to produce training labels, a U-Net for coarse pixel-wise segmentation of defects, and a few-shot learning classifier (comparing prototypical, matching, relation, and simple-shot networks) that labels cropped defects using a user-supplied support set. The pipeline is evaluated on three surfaces: Si(001) (including Si(001):H and AsH3-exposed), Ge(001):AsH3, and TiO2(110). The authors report U-Net total dice scores of 0.96-0.99 and 1-shot FSL accuracies of up to 93% (Si), 62% (Ge), and 70% (TiO2).

Significance. The problem addressed - reducing manual labelling for STM image analysis - is practically important and timely. The paper's strengths are its clear modular design, the comparison of four FSL algorithms, the use of disjoint train/test classes for the Ge and TiO2 evaluations, the ablation study with inverse-class augmentation, and the open release of code (GitHub) and data (Zenodo). If the claims were fully supported, the work would provide a useful guide for applying FSL in the SPM community. However, the headline one-shot generalization claim rests on the Si(001) result, which is compromised by augmentation leakage between train and test sets, and the Ge and TiO2 accuracies are too low to support the abstract's 'high accuracy' promise. The approach is promising but the evidence needs to be re-evaluated and the claims tempered.

major comments (3)
  1. [Appendix 11.1 and Section 3.4] The Si(001) 4-way 1-shot accuracy of 93.3 ± 0.4% reported in Table 3 is not a valid measure of generalization to unseen classes. Appendix 11.1 states that the rare feature classes h1 and h2 were augmented with rotations and reflections during training, testing, and validating, so augmented versions of the same source crops appear in both training and test sets. This violates the protocol stated in Section 3.4 that Ttr, Tte, and V cannot share any classes. The reported accuracy may largely reflect memorization of augmented variants of the same instances, rather than classification of genuinely new classes. Although the authors disclose this limitation in the appendix, the 93% figure is the strongest evidence for the abstract's claim of 'high accuracy' and the conclusion's '1-shot accuracies of up to 93%'. Please provide a clean class-disjoint evaluation on Si, or explicitly exclude the augmented classes from the headline claims.
  2. [Abstract and Sections 3.1-3.3] The abstract's statement that the model 'can be adapted to unseen surfaces with as few as one additional labelled data point' overstates the workflow's data requirements. Module 1 (Section 3.2) still requires a GPU-based unsupervised labelling step, and Module 2 (Section 3.3) trains a U-Net on five scans per new substrate; only Module 3 (FSL) uses a single labelled example per class, and it operates on crops produced by the U-Net. The one-labelled-point claim should be scoped to the classification sub-problem, and the abstract should mention the additional segmentation data required for a new surface.
  3. [Section 5 and Table 2] The conclusion's summary of '1-shot accuracies of up to 93%, 62%, and 70%' is misleading when combined with the abstract's 'high accuracy': once the leakage in the Si result is accounted for, the reliable accuracies are 62% (Ge) and 70% (TiO2), which the paper itself states would not be sufficient for definitive conclusions. Furthermore, the U-Net defect dice score for Si(001) is only 0.64 (Table 2), meaning a substantial fraction of defect pixels is missed; this under-detection limits the completeness of the final classified feature map regardless of FSL performance. Please revise the abstract and conclusion to reflect the actual evidence and to discuss the pipeline-level impact of U-Net recall.
minor comments (6)
  1. [Section 1] The Introduction contains a duplicated word: 'training data sets of this size may may not be practical' should read 'may not be practical'.
  2. [Section 3.3] The phrase 'Course segmentation' should be 'Coarse segmentation' to describe the U-Net's binary map.
  3. [Section 3.4 and Appendix 11.1] Section 3.4 states that the only augmentation used in training the FSL networks is rotation by 0, 90, 180, 270 degrees, whereas Appendix 11.1 mentions rotations and reflections for the rare classes h1 and h2; please clarify which augmentations were actually used.
  4. [Table 3 and Appendix 11.4] The kNN baseline rows list training data identical to the test data (classes 5-8 or 9-13), while the FSL models use disjoint classes, so the comparison is not on equal footing; please explain the rationale or provide a more meaningful baseline.
  5. [Section 5] The sentence 'progress can be continued towards an efficient... with minimal labelled data requirements' is followed by 'hyperparameter exploration,.'; the stray comma before the period should be removed.
  6. [Figure 3 caption] The pathway labels in the caption ('(b),(d),(e) Pathway 1' and '(c),(d),(f) Pathway 2') are confusing because panel (d) appears in both; consider simplifying the caption to clarify the flow.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the FSL evaluations use disjoint held-out classes and external pretrained features; the disclosed h1/h2 augmentation overlap is an evaluation caveat, not a circular derivation.

full rationale

The central FSL results are obtained from genuinely disjoint class splits: the silicon test classes 5-8 are absent from the training classes 1-4 and 9-15, the germanium test classes 9-13 are absent from the silicon/titania training classes, and the titania test classes 14-15 are absent from training (Table 3). The embeddings use standard conv4/ResNet18 backbones or ImageNet-pretrained features, all external benchmarks, and the accuracy is measured against held-out labelled crops rather than against the support set used to define the task. The unsupervised labelling module distils ImageNet-pretrained FCN features into a small U-Net; although the U-Net is trained on labels produced by that module, its dice scores are presented as segmentation performance on the three surfaces (Table 2) and are compared with an external supervised benchmark, not reused as the FSL prediction. The inverse-class augmentation is a training-set construction that adds new artificial classes; it is not used to fabricate the test classes. Self-citations to the authors' prior experimental work ([17], [38], [45]) are confined to sample preparation and are not load-bearing for the machine-learning derivation. The paper itself discloses a genuine evaluation-protocol limitation in Appendix 11.1: features h1 and h2 were augmented 'during training, testing, and validating', so augmented versions of the same crops appear in more than one split; this can inflate the silicon 1-shot 93% figure, but it is a robustness and correctness caveat rather than a case where a predicted quantity equals an input by construction. Overall, the derivation chain is self-contained against external benchmarks, so the circularity burden is minimal; the score of 2 reflects only the presence of minor, non-load-bearing self-citations.

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

The main free parameters are user-chosen clustering granularities and training durations; none are fitted to the test sets. The core domain assumptions are that ImageNet features transfer to STM images and that the training classes span the space of unseen defects, both of which are only partially validated.

free parameters (4)
  • k_means_k_high_res = 7
    User-determined number of clusters in the high-resolution k-means segmentation (Figure 3e); controls the granularity of the defect/lattice separation.
  • k_means_k_low_res = 5
    User-determined number of clusters in the low-resolution k-means segmentation (Figure 3f); used to identify phase domains.
  • unet_crop_size = 64x64 pixels
    Crop size for U-Net training and inference, chosen to balance context and resolution; a design choice that affects boundary precision.
  • fsl_training_episodes = 2000 (proto/matching), 5000 (relation), 20 epochs (simple shot)
    Training durations chosen so that test accuracy converges; not tuned per dataset, but the relation network's longer training reflects its larger capacity.
assumptions (4)
  • domain assumption ImageNet-pretrained FCN-ResNet101 features separate STM defect pixels from lattice pixels via k-means.
    Module 1 relies on feature transfer from natural images to STM; validated only on the three surfaces studied.
  • domain assumption The user-defined FSL support set correctly labels K examples per class.
    The workflow depends on a human user to define classes and choose representative crops in step (iv); mislabelling propagates to all queries.
  • domain assumption Training classes from Si, Ge, and TiO2 are representative enough to embed unseen surface features.
    FSL generalization to unseen surfaces relies on the latent space learned from the training classes; the low Ge accuracy suggests this assumption is fragile.
  • domain assumption Ground-truth labels used for evaluation are correct.
    No independent verification of the human labels is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Overcoming Labelled Data Scarcity for Defect Classification in Scanning Tunneling Microscopy." pith.science (2026). https://pith.science/paper/PMFI2YIO

@misc{pith2026250601678,
  author       = {Pith},
  title        = {Pith review of: Overcoming Labelled Data Scarcity for Defect Classification in Scanning Tunneling Microscopy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PMFI2YIO}},
  note         = {Machine review of arXiv:2506.01678}
}
abstract

Scanning tunnelling microscopy (STM) is a powerful technique for imaging surfaces with atomic resolution, providing insight into physical and chemical processes at the level of single atoms and molecules. A regular task of STM image analysis is the identification and labelling of features of interest against a uniform background. Performing this manually is a labour-intensive task, requiring significant human effort. To reduce this burden, we propose an automated approach to the segmentation of STM images that uses both few-shot learning and unsupervised learning. Our technique offers greater flexibility compared to previous supervised methods; it removes the requirement for large manually annotated datasets and is thus easier to adapt to an unseen surface while still maintaining a high accuracy. We demonstrate the effectiveness of our approach by using it to recognise atomic features on three distinct surfaces: Si(001), Ge(001), and TiO$_2$(110), including adsorbed AsH$_3$ molecules on the silicon and germanium surfaces. Our model exhibits strong generalisation capabilities, and following initial training, can be adapted to unseen surfaces with as few as one additional labelled data point. This work is a significant step towards efficient and material-agnostic, automatic segmentation of STM images.

Figures

Figures reproduced from arXiv: 2506.01678 by the authors.

Figure 1
Figure 1. (a) Filled and (b) empty state STM images of the same area of a Si(001):H surface. Imaging [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. STM Image Segmentation and Feature Classification Workflow: The workflow is illustrated using [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Overview of the unsupervised pipeline for STM image segmentation. (a) Original filled-state [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Illustration of a metric based approach to few-shot learning. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Final image segmentation plus feature classification on the three substrates: (a) Si(001) (b) Ge(001):AsH3 (c) TiO2(110). The colour coded squares around the features are the ones that were given as examples to the few-shot learning algorithm. In each case we use the b…
Figure 6
Figure 6. Figure 6: Crops of the different features used to train and evaluate our networks. [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]
Figure 7
Figure 7. Figure 7: Less computationally heavy procedure for unsupervised labelling. The procedure from (a) to (c) [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 8
Figure 8. Figure 8: (a) The relation module from the original relation network paper with a neural network metric [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]
Figure 9
Figure 9. Figure 9: Comparison of the STEM data segmentation from [ [PITH_FULL_IMAGE:figures/full_fig_p021_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 45 canonical work pages

  1. [1]

    Matthews, Le Wang, Yingge Du, and Steven R

    Sarah Akers, Elizabeth Kautz, Andrea Trevino- Gavito, Matthew Olszta, Bethany E. Matthews, Le Wang, Yingge Du, and Steven R. Spurgeon. Rapid and flexible segmentation of electron mi- croscopy data using few-shot machine learning. npj Computational Materials, 7(1):187, 2021

    work page 2021

  2. [2]

    Segment anything for microscopy

    Anwai Archit, Sushmita Nair, Nabeel Khalid, Paul Hilt, Vikas Rajashekar, Marei Freitag, Sag- nik Gupta, Andreas Dengel, Sheraz Ahmed, and Constantin Pape. Segment anything for microscopy. bioRxiv, 2023

    work page 2023

  3. [3]

    Direct visualization of defect- mediated dissociation of water on tio2 (110)

    Oier Bikondoa, Chi L Pang, Roslinda Ithnin, Christopher A Muryn, Hiroshi Onishi, and Ge- off Thornton. Direct visualization of defect- mediated dissociation of water on tio2 (110). Nature materials, 5(3):189–192, 2006

    work page 2006

  4. [4]

    Binnig, H

    G. Binnig, H. Rohrer, Ch. Gerber, and E. Weibel. 7 × 7 reconstruction on si(111) re- solved in real space. Phys. Rev. Lett., 50:120– 123, Jan 1983

    work page 1983

  5. [5]

    D. R. Bowler, J. H. G. Owen, K. Miki, and G. A. D. Briggs. Diffusion of paired hydrogen on si(001). Phys. Rev. B, 57:8790–8793, Apr 1998. 11

    work page 1998

  6. [6]

    S. K. Broome, M. A.and Gorman, M. G. House, S. J. Hile, J. G. Keizer, D. Keith, Hill C. D., T. F. Watson, W. J. Baker, L. C. L. Hollenberg, and M. Y . Simmons. Two-electron spin correlations in precision placed donors in silicon. Nature Communications, 9, March 2018

    work page 2018

  7. [7]

    Choudhary, Camp C

    K. Choudhary, Camp C. Kevin F., Kalinin Sergei V ., R. Vasudevan, M. Zi- atdinov, and F. Tavazza. Computational scanning tunneling microscope image database. Scientific Data, 2021

    work page 2021

  8. [8]

    Atomic defect classi- fication of the h–si(100) surface through multi- mode scanning probe microscopy

    Jeremiah Croshaw, Thomas Dienel, Taleana Huff, and Robert Wolkow. Atomic defect classi- fication of the h–si(100) surface through multi- mode scanning probe microscopy. Beilstein Journal of Nanotechnology, 11:1346–1360, sep 2020

    work page 2020

Show all 47 references
  1. [9]

    The surface science of titanium dioxide

    Ulrike Diebold. The surface science of titanium dioxide. Surface Science Reports , 48(5):53– 229, 2003

    2003

  2. [10]

    Gordon, P

    O. Gordon, P. D’Hondt, L. Knijff, S. E. Freeney, F. Junqueira, P. Moriarty, and I. Swart. Scanning tunneling state recognition with multi-class neu- ral network ensembles. Review of Scientific Instruments, 90(10):103704, 10 2019

    2019

  3. [11]

    Embedding human heuristics in machine-learning-enabled probe microscopy

    Oliver M Gordon, Filipe LQ Junqueira, and Philip J Moriarty. Embedding human heuristics in machine-learning-enabled probe microscopy. Machine Learning: Science and Technology , 1(1):015001, 2020

    2020

  4. [12]

    Nanoscale imaging of mobile carriers and trapped charges in delta doped silicon p–n junc- tions

    Georg Gramse, Alexander Kölker, Tomáš Šk- ereˇn, Taylor JZ Stock, Gabriel Aeppli, Ferry Kienberger, Andreas Fuhrer, and Neil J Cur- son. Nanoscale imaging of mobile carriers and trapped charges in delta doped silicon p–n junc- tions. Nature Electronics, 3(9):531–538, 2020

    2020

  5. [13]

    Torch- bench: Benchmarking pytorch with high api surface coverage

    Yueming Hao, Xu Zhao, Bin Bao, David Berard, Will Constable, Adnan Aziz, and Xu Liu. Torch- bench: Benchmarking pytorch with high api surface coverage. arXiv e-prints, pages arXiv– 2304, 2023

    2023

  6. [15]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CoRR, abs/1512.03385, 2015

    2015 arXiv

  7. [16]

    Metaau- dio: A few-shot audio classification benchmark

    Calum Heggan, Sam Budgett, Timothy Hospedales, and Mehrdad Yaghoobi. Metaau- dio: A few-shot audio classification benchmark. In International Conference on Artificial Neural Networks , pages 219–230. Springer, 2022

    2022

  8. [17]

    Emily V . S. Hofmann, Taylor J. Z. Stock, Oliver Warschkow, Rebecca Conybeare, Neil J. Cur- son, and Steven R. Schofield. Room temper- ature incorporation of arsenic atoms into the germanium (001) surface. Angewandte Chemie International Edition, 62(7):e202213982, 2023

    2023

  9. [18]

    Carlo Innamorati, Tobias Ritschel, Tim Weyrich, and Niloy J. Mitra. Learning on the edge: Explicit boundary handling in cnns. CoRR, abs/1805.03106, 2018

    2018 arXiv

  10. [19]

    Davenport, Zhe- hao Ge, Eberth A

    Frédéric Joucken, John L. Davenport, Zhe- hao Ge, Eberth A. Quezada-Lopez, Takashi Taniguchi, Kenji Watanabe, Jairo Velasco, Jérôme Lagoute, and Robert A. Kaindl. De- noising scanning tunneling microscopy images of graphene with supervised machine learning. Phys. Rev. Mater., ...

    2022

  11. [20]

    Efficient few-shot machine learning for classification of ebsd patterns

    Kevin Kaufmann, Hobson Lane, Xiao Liu, and Kenneth S Vecchio. Efficient few-shot machine learning for classification of ebsd patterns. Sci- entific reports, 11(1):8172, 2021

    2021

  12. [21]

    Huang, and Bryan K

    Abid Khan, Chia-Hao Lee, Pinshane Y . Huang, and Bryan K. Clark. Leveraging generative adversarial networks to create realistic scanning transmission electron microscopy images. npj Computational Materials, 9(1), may 2023

    2023

  13. [22]

    Krull, P

    A. Krull, P. Hirsch, C. Rother, A. Schiffrin, and Krull C. Artificial-intelligence-driven scanning probe microscopy. Nature, Mar 2020

    2020

  14. [23]

    Prototypical network for radar image recogni- tion with few samples

    Diyang Liu, Xunzhang Gao, and Qinmu Shen. Prototypical network for radar image recogni- tion with few samples. Journal of Physics: Con- ference Series, 1634(1):012116, sep 2020. 12

    2020

  15. [24]

    Fully convolutional networks for se- mantic segmentation

    Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for se- mantic segmentation. CoRR, abs/1411.4038, 2014

    2014 arXiv

  16. [25]

    Nanoscale patterning and oxida- tion of h-passivated si (100)-2× 1 surfaces with an ultrahigh vacuum scanning tunneling micro- scope

    JW Lyding, T-C Shen, JS Hubacek, JR Tucker, and GC Abeln. Nanoscale patterning and oxida- tion of h-passivated si (100)-2× 1 surfaces with an ultrahigh vacuum scanning tunneling micro- scope. Applied physics letters , 64(15):2010– 2012, 1994

    2010

  17. [26]

    Proto-clip: Vision-language prototypical network for few-shot learning

    Kamalesh Palanisamy, Yu-Wei Chao, Xinya Du, Yu Xiang, et al. Proto-clip: Vision-language prototypical network for few-shot learning. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2594–2601. IEEE, 2024

    2024

  18. [27]

    Structure of clean and adsorbate- covered single-crystal rutile tio2 surfaces

    Chi Lun Pang, Robert Lindsay, and Geoff Thornton. Structure of clean and adsorbate- covered single-crystal rutile tio2 surfaces. Chemical reviews, 113(6):3887–3948, 2013

    2013

  19. [28]

    Intro- ducing hann windows for reducing edge-effects in patch-based image segmentation

    Nicolas Pielawski and Carolina Wählby. Intro- ducing hann windows for reducing edge-effects in patch-based image segmentation. PLOS ONE, 15(3):e0229839, March 2020

    2020

  20. [29]

    Deep learning-guided surface characterization for autonomous hydro- gen lithography

    Mohammad Rashidi, Jeremiah Croshaw, Kieran Mastel, Marcus Tamura, Hedieh Hosseinzadeh, and Robert A Wolkow. Deep learning-guided surface characterization for autonomous hydro- gen lithography. Machine Learning: Science and Technology, 1(2):025001, mar 2020

    2020

  21. [30]

    Au- tonomous scanning probe microscopy in situ tip conditioning through machine learning

    Mohammad Rashidi and Robert A Wolkow. Au- tonomous scanning probe microscopy in situ tip conditioning through machine learning. ACS nano, 12(6):5185–5189, 2018

    2018

  22. [31]

    Dissociative adsorption of diethyl ether on si(001) studied by means of scanning tunneling microscopy and photoelec- tron spectroscopy

    Marcel Reutzel, Gerson Mette, Peter Stromberger, Ulrich Koert, Michael Dürr, and Ulrich Höfer. Dissociative adsorption of diethyl ether on si(001) studied by means of scanning tunneling microscopy and photoelec- tron spectroscopy. The Journal of Physical Chemistry C, 119, March 2015

    2015

  23. [32]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medi- cal image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, par...

    2015

  24. [33]

    True first- order surface phase transition without nanoscale phase separation

    Hyungjoon Shim and Geunseop Lee. True first- order surface phase transition without nanoscale phase separation. ACS Nano, 17, Jul 2023

    2023

  25. [34]

    Bipolar device fabrication using a scan- ning tunnelling microscope

    Tomáš Škere ˇn, Sigrun A Köster, Bastien Douhard, Claudia Fleischmann, and Andreas Fuhrer. Bipolar device fabrication using a scan- ning tunnelling microscope. Nature Electronics, 3(9):524–530, 2020

    2020

  26. [35]

    Prototypical networks for few-shot learning.Ad- vances in neural information processing sys- tems, 30, 2017

    Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning.Ad- vances in neural information processing sys- tems, 30, 2017

    2017

  27. [36]

    Atom- istic origin of metal versus charge-density-wave phase separation in indium atomic wires on si(111)

    Sun Kyu Song and Han Woong Yeom. Atom- istic origin of metal versus charge-density-wave phase separation in indium atomic wires on si(111). Phys. Rev. B, 104:035420, Jul 2021

    2021

  28. [37]

    A com- prehensive survey of few-shot learning: Evolu- tion, applications, challenges, and opportunities

    Yisheng Song, Ting Wang, Puyu Cai, Subrota K Mondal, and Jyoti Prakash Sahoo. A com- prehensive survey of few-shot learning: Evolu- tion, applications, challenges, and opportunities. ACM Computing Surveys, 55(13s), 2023

    2023

  29. [38]

    Single-atom con- trol of arsenic incorporation in silicon for high- yield artificial lattice fabrication

    Taylor JZ Stock, Oliver Warschkow, Proco- pios C Constantinou, David R Bowler, Steven R Schofield, and Neil J Curson. Single-atom con- trol of arsenic incorporation in silicon for high- yield artificial lattice fabrication. Advanced Ma- terials, 36(24):2312282, 2024

    2024

  30. [39]

    Microstructure segmentation with deep learning encoders pre-trained on a large mi- croscopy dataset

    Joshua Stuckner, Bryan Harder, and Timothy M Smith. Microstructure segmentation with deep learning encoders pre-trained on a large mi- croscopy dataset. npj Computational Materials, 8(1):200, 2022

    2022

  31. [40]

    Learning to compare: Relation network for few-shot learning

    Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1199–1208, 2018

    2018

  32. [41]

    Matching networks 13 for one shot learning

    Oriol Vinyals, Charles Blundell, Timothy Lilli- crap, Daan Wierstra, et al. Matching networks 13 for one shot learning. Advances in neural infor- mation processing systems, 29, 2016

    2016

  33. [42]

    Kashid, Pradeep Namboodiri, Andrew Schmucker, Scott W.and Murphy, M

    Xiqiao Wang, Jonathan Wyrick, Ranjit V . Kashid, Pradeep Namboodiri, Andrew Schmucker, Scott W.and Murphy, M. D. Stewart, and Richard M. Silver. Atomic-scale control of tunneling in donor-based devices. Communications Physics, 3, May 2020

    2020

  34. [43]

    Sim- pleshot: Revisiting nearest-neighbor classifica- tion for few-shot learning

    Yan Wang, Wei-Lun Chao, Kilian Q Wein- berger, and Laurens van der Maaten. Sim- pleshot: Revisiting nearest-neighbor classifica- tion for few-shot learning. arXiv e-prints, pages arXiv–1911, 2019

    1911

  35. [44]

    Warschkow, N

    O. Warschkow, N. J. Curson, S. R. Schofield, N. A. Marks, H. F. Wilson, M. W. Radny, P. V . Smith, T. C. G. Reusch, D. R. McKenzie, and M. Y . Simmons. Reaction paths of phosphine dissociation on silicon (001). The Journal of Chemical Physics, 144(1):014705, 01 2016

    2016

  36. [45]

    Direct visual- ization of a gold nanoparticle electron trapping effect

    Oscar Bentley Jerdmyr Williams, Khabiboulakh Katsiev, Byeongjin Baek, George Harrison, Ge- off Thornton, and Hicham Idriss. Direct visual- ization of a gold nanoparticle electron trapping effect. Journal of the American Chemical Soci- ety, 144(2):1034–1044, 2022

    2022

  37. [46]

    Classifying galaxy morphologies with few-shot learning

    Zhirui Zhang, Zhiqiang Zou, Nan Li, and Yanli Chen. Classifying galaxy morphologies with few-shot learning. Research in Astronomy and Astrophysics, 22(5):055002, 2022

    2022

  38. [47]

    Ro- bust multi-scale multi-feature deep learning for atomic and defect identification in scanning tun- neling microscopy on h-si (100) 2x1 surface

    Maxim Ziatdinov, Udi Fuchs, James HG Owen, John N Randall, and Sergei V Kalinin. Ro- bust multi-scale multi-feature deep learning for atomic and defect identification in scanning tun- neling microscopy on h-si (100) 2x1 surface. arXiv e-prints, pages arXiv–2002, 2020

    2002

  39. [48]

    Maxim Ziatdinov, Ayana Ghosh, Chun Yin Wong, and Sergei V . Kalinin. Atomai frame- work for deep learning analysis of image and spectroscopy data in electron and scanning probe microscopy. Nature Machine Intelligence, 4(12):1101–1112, December 2022. 14 11 Appendix 11.1 Dataset...

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.