Pith. sign in

REVIEW 3 major objections 4 minor 38 references

Automated Grading of Students' Handwritten Graphs: A Comparison of Meta-Learning and Vision-Large Language Models

T0 review · 3 major / 4 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 Specialized few-shot models beat vision-language models on binary grading of handwritten graphs, but the large models win when grading has three classes.

desk verdict A novel dataset and a genuinely new task, with a plausible but not statistically established comparison; the meta-learning split needs to be clarified before the 2-way claim can be trusted. read the letter →

arxiv 2507.03056 v1 pith:RMGJYHKE submitted 2025-07-03 cs.LG cs.AI

Behnam Parsaeifard , Martin Hlosta , Per Bergamin This is my paper · ORCID
classification cs.LGcs.AI
keywords DeepLearningMeta-LearningVision-LanguageModelsAutogradingHandwrittenGraphsFew-ShotMultimodalEducationalAssessment
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

The paper tries to establish that automated grading of students' handwritten graphs is possible with very few labeled examples, and that the better model family depends on the number of grade classes. On a dataset of 1,174 real graph images from a distance university's economics courses, the best meta-learning models outperform vision-language models in 2-way classification, while the best vision-language models outperform the meta-learning models in 3-way classification. The practical point is that an institution choosing an autograding approach would pick a small trained model for a binary rubric check but may get more accuracy from a general-purpose vision-language model once grading has more categories, at the cost of the reliability and privacy concerns the paper flags.

What carries the argument

The load-bearing setup is N-way K-shot episodic meta-learning with a multimodal embedding function that concatenates a ResNet-18 graph embedding with a German BERT text embedding. Grades are derived from a binary rubric vector through the identity $\text{Grade} = \sum_{i=0}^{n-1} C_i\,2^i$, which turns each graph into a class label and makes grading a few-shot classification problem. Five meta-learning algorithms (Matching, Prototypical, Relation, FOMAML, and ProtoMAML) are compared against six API-based vision-language models used with few-shot in-context learning; both families receive the same support-and-query sampling so the comparison rests on the same task construction.

What would settle it

Re-annotate the same 1,174 graphs with two or more independent teachers using the same rubrics, then measure inter-annotator agreement and retrain or re-prompt all models on each annotator's labels; if agreement is low or the best model family changes across annotators, the central crossover claim is not robust.

Watch

Extended reading notes

Core claim

The central discovery is a crossover in model-family performance. In 2-way classification the best meta-learning model, the Prototypical Network, reaches 56.89% mean accuracy at 4-shot, ahead of the best vision-language model, gpt-4.1, at 53.09%. In 3-way classification the best vision-language model, o4-mini, reaches 50.00% at 4-shot, ahead of the best meta-learning model, FOMAML, at 41.30%. The paper presents this as evidence that specialized few-shot training is the better choice when the rubric is a single binary criterion, whereas general-purpose vision-language models become competitive or superior once more grade classes are involved; it does not claim either family is ready for unsupervised deployment.

Load-bearing premise

The whole comparison rests on the single teacher's annotations of all 1,174 graphs being a trustworthy ground truth; if those labels are noisy or subjective, the reported accuracies and the meta-learning-versus-VLLM ranking become artifacts of one person's judgment.

Editorial extensions

If this is right

  • On this dataset, an institution grading a single-criterion rubric would get higher accuracy from a trained Prototypical Network (up to 56.89%) than from the best vision-language model (up to 53.09%).
  • For rubrics producing three grade classes, the best vision-language model (o4-mini, 50.00% at 4-shot) beats the best meta-learning model (FOMAML, 41.30%), so the model-family recommendation reverses with task complexity.
  • Combining the graph image with OCR-extracted text generally improves accuracy over using either modality alone, so autograding pipelines should keep the text rather than discard it.
  • The vision-language results carry the paper's own practical caveats: high cost, privacy concerns, and inconsistent outputs when several graphs are placed in one prompt.
  • The comparison is a few-shot feasibility result over 200 episodes on a small annotated dataset, not a production-grade benchmark.

Reading between the lines

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

  • A natural extension is 4-way classification: if the crossover pattern is real, adding a fourth observed grade class should increase the vision-language model's advantage over trained meta-learners, which the current dataset cannot test because most assignments lack four well-populated grades.
  • Re-annotating the dataset with multiple expert teachers would show whether the reported rankings measure grading ability or one annotator's idiosyncrasies; the paper itself identifies single-annotator labeling as its main limitation.
  • Because each grade is a binary-weighted sum of rubric criteria, tasks with the same class count can differ in how many criteria they involve; future benchmarks could separate number of classes from number of rubric criteria as factors driving model choice.
  • The reported instability of vision-language models with multiple graphs per prompt suggests that batched grading may need per-graph verification prompts or an agreement check rather than a single multi-image call.
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 / 4 minor

Summary. The paper targets automated grading of handwritten graphs in university-level economics courses, a task with little prior work. The authors assemble a real-world dataset of 1,174 annotated images, encode grades via binary grading criteria, and compare five meta-learning models (Matching Network, Prototypical Network, Relation Network, FOMAML, ProtoFOMAML) with six VLLMs (GPT-4.1 variants, GPT-4o variants, o4-mini) under 2-way and 3-way few-shot settings. They report that the best meta-learning models outperform the best VLLMs in 2-way tasks, while the best VLLMs slightly outperform meta-learning in 3-way tasks. The paper also describes a preprocessing pipeline for graph extraction and an ablation study of graph-only versus text-only embeddings.

Significance. If the central comparative claim is reliable, the paper provides a useful empirical data point for model-family selection in low-data visual grading tasks. Its strengths include a real-world educational dataset, a clearly defined task with an explicit grading rubric, multiple meta-learning baselines, an ablation study, and a direct comparison with several VLLMs. However, the support for the headline claim is currently thin: the meta-learning evaluation does not describe a held-out split, the VLLM results are single-run point estimates without variance, and the ground-truth labels come from one annotator. These issues need to be resolved before the claim that meta-learning wins in 2-way and VLLMs win in 3-way can be accepted.

major comments (3)
  1. [Section III-E and Section IV-B] The meta-learning evaluation is not described with a train/eval split at the assignment or class level. Section III-E states that in each episode the authors 'sample classes from the same module assignment,' but it does not state whether the assignments or grade classes used for meta-training are disjoint from those used for meta-test episodes. With only 17 assignments and partially overlapping grade categories, a model could memorize assignment-specific visual features and inflate its accuracy. The reported 2-way 4-shot gap between the best meta-learning model (56.89% ± 3.61) and the best VLLM (53.09%) is smaller than one standard deviation, so even modest leakage could erase or reverse the claimed advantage. Please specify the exact data split (e.g., leave-one-assignment-out or leave-one-class-out) and rerun the evaluation if no split was used.
  2. [Table V and Section IV-D] The VLLM accuracies are single-run point estimates with no variance, confidence intervals, or repeated sampling, despite the use of temperature 0.1 with 'slight variability in the outputs.' The central comparative claim rests on small differences, such as the 3-way 1-shot gap of 43.75% (gpt-4o) versus 42.69% (Matching Network), and on the 3-way 4-shot result of 50.00% (o4-mini) versus 41.26% (FOMAML), which has no uncertainty attached. Without multiple runs or a significance test, the claim that VLLMs 'slightly outperform' meta-learning in 3-way tasks is not statistically grounded. Please report repeated runs with standard errors or confidence intervals, and apply a significance test where appropriate.
  3. [Section III-B and Section V] The ground-truth labels were produced by a single expert teacher who also defined the grading criteria, and the paper itself acknowledges that this 'introduces potential for human error and subjective bias.' Because the models are trained and evaluated against these labels, the reported accuracies and the practical claim of providing an 'objective and consistent' grading assistant are not yet established. While label noise may affect both model families symmetrically, the external validity of the task depends on label reproducibility. Please provide at least a subset double-annotated with inter-annotator agreement, or clearly scope the claims as reproducing a single teacher's judgments.
minor comments (4)
  1. [Section IV-B] The text reports Prototypical Network accuracy in 2-way 2-shot as '55.76% ± 0.01' and in 4-shot as '56.87% ± 3.62,' but Table III lists '55.76 ± 3.02' and '56.89 ± 3.61'; the numbers should be aligned.
  2. [Section VI] The conclusion refers to '4o-mini' instead of 'o4-mini' and contains the typo 'meat-learning' for 'meta-learning.'
  3. [Section V] The claim that converting images to binary format yields 'performance comparable' to the original color images is not supported by any table, figure, or quantitative result in the manuscript.
  4. [General] No data or code availability statement is included; given the real-world dataset and the need for reproducibility in educational machine learning, the authors should state access conditions or provide an anonymized data description.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's claims are empirical comparisons, not derived results, and no load-bearing premise reduces to its own inputs or to self-citation.

full rationale

The paper reports an empirical benchmark: meta-learning models versus VLLMs on a hand-annotated dataset of student graphs. Its central claim, that the best meta-learning models outperform VLLMs in 2-way tasks while the best VLLMs slightly outperform meta-learning models in 3-way tasks, is a summary of measured accuracies in Tables III and V. There is no derivation chain in which a predicted quantity is defined in terms of the quantity it claims to predict, no fitted parameter is renamed as a prediction, and no uniqueness theorem or prior-work assertion is invoked to force the choice of model family. The models are evaluated against ground-truth labels produced by an external expert annotation process; the few-shot support examples and query examples are distinct, and the VLLM prompts use in-context support pairs separate from the test queries. The authors' references contain no works by the present authors, so self-citation plays no role. The single-annotator labeling concern and the absence of an explicit meta-train/meta-test assignment-level split are validity risks for the empirical comparison, but they are not circularity: they concern label reliability and experimental protocol, not the logical equivalence of inputs and outputs. Under the given hard rules, concerns about generalization validity without a demonstrated self-referential reduction belong to correctness risk, not to the circularity score.

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

The paper is an empirical benchmark rather than a derivation, so the ledger records what the comparison rests on: a private single-annotator dataset, off-the-shelf pre-trained encoders, hand-chosen hyperparameters, and an implicit evaluation protocol. The costliest input the reader does not pay for upstream is label validity: one expert defined every rubric and labeled every graph with no inter-annotator agreement, which the authors themselves disclose in Section V.

free parameters (4)
  • VLLM sampling temperature = 0.1
    Chosen by hand in Section IV-D to make outputs more deterministic; o4-mini does not accept temperature, so models are not compared under identical sampling conditions.
  • Meta-learning training budget = 1,000 epochs; 100 inner gradient steps for FOMAML and ProtoFOMAML
    Stated in Section IV-B without optimizer, learning rate, or seeds, so the exact training protocol is not reproducible as written.
  • Data augmentation intensity = unspecified small rotations and perspective transformations
    Augmentation is applied during training (Section IV-A) but the ranges are never quantified, leaving an uncontrolled training degree of freedom.
  • Task restriction to usable grade classes = N in {2,3}, K in {1,2,4}
    Section IV-A restricts tasks to grades with at least N x (K+1) instances and highlights single-instance rows as unusable; this excludes the rarest classes from evaluation.
assumptions (5)
  • domain assumption One expert teacher's rubrics and labels are treated as ground truth for training and evaluation
    Section III-B builds labels from a single collaborating teacher; Section V discloses potential human error and subjective bias. If labels are wrong, all accuracy numbers are unanchored.
  • domain assumption The contour-detection step isolates the relevant graph region
    Section III-C and IV-A identify the graph as the largest square-like contour with annotator confirmation; missed contours would silently corrupt inputs.
  • domain assumption The binary-weighted grade encoding (Eq. 1) yields meaningful class labels
    Grade = sum of Ci times 2^i maps each fulfilled-criteria subset to a distinct integer; the 2-way and 3-way tasks presuppose these integers are the right classes.
  • domain assumption Episodes sampled from the same module assignment form a valid few-shot task distribution for training and evaluation
    Section III-E samples classes from one assignment per episode; no held-out assignment-level split is described, so cross-task generalization is not demonstrated.
  • domain assumption OCR-extracted German text is a useful grading signal when combined with the image
    Section III-C extracts text via Tesseract OCR; the ablation in Table IV shows text-only accuracy near chance, yet the multi-modal model assumes the text branch contributes signal.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Automated Grading of Students' Handwritten Graphs: A Comparison of Meta-Learning and Vision-Large Language Models." pith.science (2026). https://pith.science/paper/RMGJYHKE

@misc{pith2026250703056,
  author       = {Pith},
  title        = {Pith review of: Automated Grading of Students' Handwritten Graphs: A Comparison of Meta-Learning and Vision-Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RMGJYHKE}},
  note         = {Machine review of arXiv:2507.03056}
}
read the original abstract

With the rise of online learning, the demand for efficient and consistent assessment in mathematics has significantly increased over the past decade. Machine Learning (ML), particularly Natural Language Processing (NLP), has been widely used for autograding student responses, particularly those involving text and/or mathematical expressions. However, there has been limited research on autograding responses involving students' handwritten graphs, despite their prevalence in Science, Technology, Engineering, and Mathematics (STEM) curricula. In this study, we implement multimodal meta-learning models for autograding images containing students' handwritten graphs and text. We further compare the performance of Vision Large Language Models (VLLMs) with these specially trained metalearning models. Our results, evaluated on a real-world dataset collected from our institution, show that the best-performing meta-learning models outperform VLLMs in 2-way classification tasks. In contrast, in more complex 3-way classification tasks, the best-performing VLLMs slightly outperform the meta-learning models. While VLLMs show promising results, their reliability and practical applicability remain uncertain and require further investigation.

Figures

Figures reproduced from arXiv: 2507.03056 by the authors.

Figure 1
Figure 1. Data preprocessing pipeline for an image. (Left) The original student-drawn image containing graph, text, and extraneous margins. (Top) OCR [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Performance of each meta-learning model across different module assignments for 2-way. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Performance of each meta-learning model across different module assignments for 3-way. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Performance of each LLM across different module assignments for 2-way. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Performance of each LLM across different module assignments for 3-way. [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Comparison of best-performing meta-learning models and VLLMs [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Comparison of image transformations applied to the original image. [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 20 canonical work pages

  1. [1]

    Knowledge tracing: Modeling the acquisition of procedural knowledge,

    A. T. Corbett and J. R. Anderson, “Knowledge tracing: Modeling the acquisition of procedural knowledge,” User modeling and user-adapted interaction, vol. 4, pp. 253–278, 1994

    work page 1994

  2. [2]

    Caa of short non-mcq answers,

    D. H. Callear, J. Jerrams-Smith, and V . Soh, “Caa of short non-mcq answers,” 2001

    work page 2001

  3. [3]

    Effective feature in- tegration for automated short answer scoring,

    K. Sakaguchi, M. Heilman, and N. Madnani, “Effective feature in- tegration for automated short answer scoring,” in Proceedings of the 2015 conference of the North American Chapter of the association for computational linguistics: Human language technologies , 2015, pp. 1049–1054

    work page 2015

  4. [4]

    Vector based techniques for short answer grading,

    A. E. Magooda, M. Zahran, M. Rashwan, H. Raafat, and M. Fayek, “Vector based techniques for short answer grading,” in The twenty-ninth international flairs conference , 2016

    work page 2016

  5. [5]

    An automatic short-answer grading model for semi-open-ended questions,

    L. Zhang, Y . Huang, X. Yang, S. Yu, and F. Zhuang, “An automatic short-answer grading model for semi-open-ended questions,” Interactive learning environments, vol. 30, no. 1, pp. 177–190, 2022

    work page 2022

  6. [6]

    The automated grading of student open responses in mathe- matics,

    J. A. Erickson, A. F. Botelho, S. McAteer, A. Varatharaj, and N. T. Heffernan, “The automated grading of student open responses in mathe- matics,” in Proceedings of the tenth international conference on learning analytics & knowledge , 2020, pp. 615–624

    work page 2020

  7. [7]

    Mathematical language processing: Automatic grading and feedback for open response mathematical questions,

    A. S. Lan, D. Vats, A. E. Waters, and R. G. Baraniuk, “Mathematical language processing: Automatic grading and feedback for open response mathematical questions,” in Proceedings of the second (2015) ACM conference on learning@ scale , 2015, pp. 167–176

    work page 2015

  8. [8]

    Efficient estimation of word representations in vector space,

    T. Mikolov, K. Chen, G. S. Corrado, and J. Dean, “Efficient estimation of word representations in vector space,” in International Conference on Learning Representations , 2013. [Online]. Available: https://api.semanticscholar.org/CorpusID:5959482

    work page 2013

Show all 38 references
  1. [9]

    Glove: Global vectors for word representation,

    J. Pennington, R. Socher, and C. D. Manning, “Glove: Global vectors for word representation,” in Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP) , 2014, pp. 1532–1543

    2014

  2. [10]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” arXiv preprint arXiv:1810.04805, 2018

    2018 arXiv

  3. [11]

    Sentence-bert: Sentence embeddings using siamese bert-networks,

    N. Reimers and I. Gurevych, “Sentence-bert: Sentence embeddings using siamese bert-networks,” arXiv preprint arXiv:1908.10084 , 2019

    1908 arXiv

  4. [12]

    Pre-training bert on domain resources for short answer grading,

    C. Sung, T. Dhamecha, S. Saha, T. Ma, V . Reddy, and R. Arora, “Pre-training bert on domain resources for short answer grading,” in Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Lang...

    2019

  5. [13]

    Exploring automatic short answer grading as a tool to assist in human rating,

    A. Condor, “Exploring automatic short answer grading as a tool to assist in human rating,” in Artificial Intelligence in Education: 21st International Conference, AIED 2020, Ifrane, Morocco, July 6–10, 2020, Proceedings, Part II 21 . Springer, 2020, pp. 74–79

    2020

  6. [14]

    Improving automated scoring of student open responses in mathematics

    S. Baral, A. F. Botelho, J. A. Erickson, P. Benachamardi, and N. T. Heffernan, “Improving automated scoring of student open responses in mathematics.” International Educational Data Mining Society , 2021

    2021

  7. [15]

    Mathbert: A pre-trained language model for general nlp tasks in mathematics education,

    J. T. Shen, M. Yamashita, E. Prihar, N. Heffernan, X. Wu, B. Graff, and D. Lee, “Mathbert: A pre-trained language model for general nlp tasks in mathematics education,” arXiv preprint arXiv:2106.07340 , 2021

    2021 arXiv

  8. [16]

    A detailed analysis of optical character recognition technology,

    K. Hamad and M. Kaya, “A detailed analysis of optical character recognition technology,” International Journal of Applied Mathematics Electronics and Computers , no. Special Issue-1, pp. 244–249, 2016

    2016

  9. [17]

    Watch, attend and parse: An end-to-end neural network based approach to handwritten mathematical expression recognition,

    J. Zhang, J. Du, S. Zhang, D. Liu, Y . Hu, J. Hu, S. Wei, and L. Dai, “Watch, attend and parse: An end-to-end neural network based approach to handwritten mathematical expression recognition,” Pattern Recognition, vol. 71, pp. 196–206, 2017

    2017

  10. [18]

    Track, attend, and parse (tap): An end-to-end framework for online handwritten mathematical expression recognition,

    J. Zhang, J. Du, and L. Dai, “Track, attend, and parse (tap): An end-to-end framework for online handwritten mathematical expression recognition,” IEEE Transactions on Multimedia, vol. 21, no. 1, pp. 221– 233, 2018

    2018

  11. [19]

    Ai-assisted automated short answer grading of handwritten university level mathematics exams,

    T. Liu, J. Chatain, L. Kobel-Keller, G. Kortemeyer, T. Willwacher, and M. Sachan, “Ai-assisted automated short answer grading of handwritten university level mathematics exams,” arXiv preprint arXiv:2408.11728 , 2024

    2024 arXiv

  12. [20]

    Toward ai grading of student problem solutions in introductory physics: A feasibility study,

    G. Kortemeyer, “Toward ai grading of student problem solutions in introductory physics: A feasibility study,” Physical Review Physics Education Research, vol. 19, no. 2, p. 020163, 2023

    2023

  13. [21]

    Drawedumath: Evaluating vision language models with expert-annotated students’ hand-drawn math images,

    S. Baral, L. Lucy, R. Knight, A. Ng, L. Soldaini, N. T. Heffernan, and K. Lo, “Drawedumath: Evaluating vision language models with expert-annotated students’ hand-drawn math images,” arXiv preprint arXiv:2501.14877, 2025

    2025 arXiv

  14. [22]

    Auto-scoring student responses with images in mathematics

    S. Baral, A. Botelho, A. Santhanam, A. Gurung, L. Cheng, and N. Hef- fernan, “Auto-scoring student responses with images in mathematics.” The Proceedings of the 16th International Conference on Educational Data Mining., 2023

    2023

  15. [23]

    Sympy: symbolic computing in python,

    A. Meurer, C. P. Smith, M. Paprocki, O. ˇCert´ık, S. B. Kirpichev, M. Rocklin, A. Kumar, S. Ivanov, J. K. Moore, S. Singh, T. Rathnayake, S. Vig, B. E. Granger, R. P. Muller, F. Bonazzi, H. Gupta, S. Vats, F. Johansson, F. Pedregosa, M. J. Curry, A. R. Terrel, v. Rou ˇcka, A. ...

    2017 doi

  16. [24]

    Automatic short math answer grading via in-context meta-learning,

    M. Zhang, S. Baral, N. Heffernan, and A. Lan, “Automatic short math answer grading via in-context meta-learning,” arXiv preprint arXiv:2205.15219, 2022

    2022 arXiv

  17. [25]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark et al., “Learning transferable visual models from natural language supervision,” in International conference on machine learning . PmLR, 2021, pp. 8748–8763

    2021

  18. [26]

    Evaluating gpt-4 at grading handwritten solutions in math exams,

    A. Caraeni, A. Scarlatos, and A. Lan, “Evaluating gpt-4 at grading handwritten solutions in math exams,” arXiv preprint arXiv:2411.05231, 2024

    2024 arXiv

  19. [27]

    The assistments ecosystem: Building a platform that brings scientists and teachers together for min- imally invasive research on human learning and teaching,

    N. T. Heffernan and C. L. Heffernan, “The assistments ecosystem: Building a platform that brings scientists and teachers together for min- imally invasive research on human learning and teaching,” International Journal of Artificial Intelligence in Education , vol. 24, pp. 470...

    2014

  20. [28]

    An overview of the tesseract ocr engine,

    R. Smith, “An overview of the tesseract ocr engine,” in Ninth interna- tional conference on document analysis and recognition (ICDAR 2007) , vol. 2. IEEE, 2007, pp. 629–633

    2007

  21. [29]

    Imagenet classification with deep convolutional neural networks,

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” Advances in neural informa- tion processing systems , vol. 25, 2012

    2012

  22. [30]

    Visualizing and understanding convo- lutional networks,

    M. D. Zeiler and R. Fergus, “Visualizing and understanding convo- lutional networks,” in Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13. Springer, 2014, pp. 818–833

    2014

  23. [31]

    Very deep convolutional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” in 3rd International Conference on Learning Representations (ICLR 2015) . Computational and Biological Learning Society, 2015

    2015

  24. [32]

    Going deeper with convolutions,

    C. Szegedy, W. Liu, Y . Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V . Vanhoucke, and A. Rabinovich, “Going deeper with convolutions,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2015, pp. 1–9

    2015

  25. [33]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 770–778

    2016

  26. [34]

    Matching net- works for one shot learning,

    O. Vinyals, C. Blundell, T. Lillicrap, D. Wierstra et al., “Matching net- works for one shot learning,” Advances in neural information processing systems, vol. 29, 2016

    2016

  27. [35]

    Prototypical networks for few-shot learning,

    J. Snell, K. Swersky, and R. Zemel, “Prototypical networks for few-shot learning,” Advances in neural information processing systems , vol. 30, 2017

    2017

  28. [36]

    Learning to compare: Relation network for few-shot learning,

    F. Sung, Y . Yang, L. Zhang, T. Xiang, P. H. Torr, and T. M. Hospedales, “Learning to compare: Relation network for few-shot learning,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 1199–1208

    2018

  29. [37]

    Model-agnostic meta-learning for fast adaptation of deep networks,

    C. Finn, P. Abbeel, and S. Levine, “Model-agnostic meta-learning for fast adaptation of deep networks,” in International conference on machine learning. PMLR, 2017, pp. 1126–1135

    2017

  30. [38]

    Meta-dataset: A dataset of datasets for learning to learn from few examples,

    E. Triantafillou, T. Zhu, V . Dumoulin, P. Lamblin, U. Evci, K. Xu, R. Goroshin, C. Gelada, K. Swersky, P.-A. Manzagol, and H. Larochelle, “Meta-dataset: A dataset of datasets for learning to learn from few examples,” in International Conference on Learning Representations , 2...

    2020

Pith tools

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