pith. sign in

arxiv: 2606.05471 · v1 · pith:N2HTUNEMnew · submitted 2026-06-03 · 💻 cs.CV

Formal Concept Lattices are Good Semantic Scaffolds for Concept-Based Learning

Deepika SN Vemuri , Sayanta Adhikari , Ankit Saha , Krishn Vishwas Kher , Vineeth N Balasubramanian This is my paper

Pith reviewed 2026-06-28 06:19 UTC · model grok-4.3

classification 💻 cs.CV
keywords formal concept analysisconcept-based learningneural network interpretabilityhierarchical conceptssemantic scaffoldsconcept latticesconcept intervention
0
0 comments X

The pith

Formal concept lattices assign concepts to neural network layers by their level of generality to create staged semantic representations.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that concept-based models usually place all concepts at one layer as an unstructured set, which ignores how human understanding organizes ideas from general to specific. It uses formal concept lattices to capture those hierarchies directly from data and map general concepts to earlier layers and specific ones to later layers. This produces representations that develop in a semantically ordered way through the network depth. If the mapping holds, interventions become more effective because they can target concepts at their natural abstraction level, and embeddings become more interpretable without loss of task accuracy. Tests on real-world image datasets confirm the lattices yield hierarchically structured concept representations.

Core claim

Formal concept lattices provide principled semantic scaffolds that naturally identify where in the network concepts should be learned based on their level of generality, allowing the model to develop staged, semantically grounded representations throughout its depth.

What carries the argument

Formal concept lattices, the hierarchical structures produced by Formal Concept Analysis that order concepts by generality and specificity to determine their assigned network layer.

If this is right

  • Embeddings become more interpretable because each layer corresponds to a clear level of semantic abstraction.
  • Interventions on concepts become more effective by operating at the layer where the concept is explicitly represented.
  • Concept representations emerge as both meaningful and explicitly hierarchically structured rather than flat.
  • Task performance remains intact while the above properties are gained.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same lattice guidance could be applied during architecture search to decide layer widths based on concept count at each generality level.
  • If lattices are recomputed on new data distributions, the model could adapt its internal concept placement without full retraining.
  • The approach suggests a way to regularize networks so that feature emergence order matches an external semantic hierarchy rather than emerging only from data statistics.

Load-bearing premise

The hierarchical relations captured by formal concept lattices from data will align with the feature hierarchies that emerge across neural network layers.

What would settle it

Training the guided model on the same datasets and finding that concept interventions at the lattice-assigned layers produce no measurable gain in effectiveness over a flat single-layer baseline.

Figures

Figures reproduced from arXiv: 2606.05471 by Ankit Saha, Deepika SN Vemuri, Krishn Vishwas Kher, Sayanta Adhikari, Vineeth N Balasubramanian.

Figure 1
Figure 1. Figure 1: Classes (orange) and attributes (green) (a) An illustrative example of how attributes shared by more classes are more general; those shared by fewer are more specific, naturally forming a subset￾superset hierarchy; (b) Attributes learned by a FoCA CBM at different layers for the class White Stork in ImageNet100. Recent work has focused on making models inherently in￾terpretable (Chen et al., 2019; Sarkar e… view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of our overall approach. A formal concept lattice is constructed for a concept-based setting using class and attribute sets and the binary relation between them (top). A neural network is supervised at intermediate blocks using the information extracted from specific levels in the lattice (bottom). Mi = [ fc∈Li fc.intent (2) where fc.intent denotes the intent of the formal concept. By repeatin… view at source ↗
Figure 3
Figure 3. Figure 3: An illustrative example on 10 classes of how lattice levels are selected to supervise layers (blocks in our implementation) in the network. We perform clustering with n (here 10) centers at each layer and obtain the average number of unique classes present per cluster, compute the same over the formal concept extents at each lattice level, and then choose closest alignment. Dj = 1 n Xn k=1 |{Y(Kk)} | (3) w… view at source ↗
Figure 4
Figure 4. Figure 4: t-SNE plots of sample embeddings from AwA2 obtained from trained ResNet-18 backbones of Vanilla CBM and FoCA CBM. On a CBM, the clusters separate at the final block; in FoCA CBM, the separation happens gradually over blocks, showing graded semantic learning. Pr f foca  Gˆ i+1(x) ̸⊆ Gˆ i(x)  ≪ Pr f rnd  Gˆ i+1(x) ̸⊆ Gˆ i(x)  . We provide a brief proof sketch herein, while deferring the detailed proof to… view at source ↗
Figure 5
Figure 5. Figure 5: (a) Severity of misclassification informs which level of attributes to intervene on. Ground truth class for current input is c2, which is not in prediction set at layer 2; we hence intervene at 2nd layer; (b) Comparison of average number of corrections on interventions at last attribute layer (blue) and at appropriate intermediate layer on Imagenet100 (orange). Intervening at apt level improves performance… view at source ↗
Figure 6
Figure 6. Figure 6: Visualizing some class clusters at different blocks. Classes from the clusters of a Vanilla CBM are mixed up across blocks, whereas a FoCA CBM gradually telescopes relevant concepts. O(|L| · |G| 2 · |M|), where |L| is the number of formal concepts; however, real-world class–attribute relations are typically sparse (fill ratio < 0.1), resulting in near-quadratic growth in practice (Table A12) [PITH_FULL_IM… view at source ↗
read the original abstract

Learning semantics is essential for deep learning models to be interpretable and better aligned with human reasoning. Concept-based models approach this by representing classes through meaningful semantic abstractions, but typically treat all concepts as a flat, unstructured set learned at a single neural network layer. This overlooks a fundamental property of human semantic understanding: concepts being organized hierarchically, from general to specific. While deep networks do learn a hierarchy of visual features, this structure is rarely aligned with explicit semantic hierarchies. Drawing on Formal Concept Analysis, we demonstrate that formal concept lattices provide principled semantic scaffolds to guide neural network learning. These lattices naturally identify where in the network concepts should be learned based on their level of generality. This allows the model to develop staged, semantically grounded representations throughout its depth. Empirical results on real-world datasets show that our models produce more interpretable embeddings, support more effective interventions, and learn concept representations that are both meaningful and hierarchically structured.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 1 minor

Summary. The paper claims that formal concept lattices derived via Formal Concept Analysis supply a hierarchy of concept generality that can be used to assign concepts to specific layers in a neural network, thereby producing staged, semantically grounded representations. The authors assert that this alignment yields more interpretable embeddings, more effective interventions, and hierarchically structured concept representations, with supporting empirical results on real-world datasets.

Significance. If the empirical claims hold, the work supplies a principled, lattice-based mechanism for injecting explicit semantic hierarchy into concept-based models, addressing the common limitation of flat concept sets. The approach is grounded in established FCA theory and offers a concrete layer-assignment rule that could improve both interpretability and intervention efficacy while preserving task performance.

minor comments (1)
  1. Abstract: the summary of empirical results would be strengthened by naming the specific datasets, metrics, and baselines even at a high level, as the current phrasing leaves the strength of the reported gains difficult to gauge from the abstract alone.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive assessment of the work and for recommending acceptance. The provided summary accurately reflects the manuscript's central claim that formal concept lattices can serve as semantic scaffolds for assigning concepts to network layers in a hierarchy-aware manner.

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The provided abstract and description contain no equations, derivations, or explicit construction steps that could reduce to self-definition, fitted inputs, or self-citation chains. The central claim—that formal concept lattices supply a hierarchy for assigning concepts to network layers—is presented as a methodological proposal supported by empirical results on real-world datasets, without any visible load-bearing step that equates a prediction to its own input by construction. No self-citation is invoked to justify uniqueness or an ansatz, and the mapping from lattice order to layer depth is not shown to be forced by prior author work. This is the expected outcome when the manuscript supplies no internal derivation chain to inspect.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Only the abstract is available, so the ledger is necessarily incomplete. No free parameters, invented entities, or explicit axioms are stated in the provided text.

axioms (1)
  • domain assumption Formal concept lattices capture the hierarchical organization of semantic concepts in a way that can be aligned with neural network layer depths.
    Invoked implicitly when the abstract states that lattices 'naturally identify where in the network concepts should be learned based on their level of generality.'

pith-pipeline@v0.9.1-grok · 5708 in / 1261 out tokens · 30081 ms · 2026-06-28T06:19:01.299679+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 8 canonical work pages

  1. [1]

    Relational concept bottleneck models

    Barbiero, P., Giannini, F., Ciravegna, G., Diligenti, M., and Marra, G. Relational concept bottleneck models. Advances in Neural Information Processing Systems, 37: 0 77663--77685, 2024

    2024

  2. [2]

    and Kontorovich, A

    Berend, D. and Kontorovich, A. A finite sample analysis of the naive bayes classifier. Journal of Machine Learning Research, 16 0 (44): 0 1519--1545, 2015. URL http://jmlr.org/papers/v16/berend15a.html

    2015

  3. [3]

    Classification and Regression Trees

    Breiman, L., Friedman, J., Stone, C., and Olshen, R. Classification and Regression Trees. Taylor & Francis, 1984. ISBN 9780412048418. URL https://books.google.co.in/books?id=JwQx-WOmSyQC

    1984

  4. [4]

    Chen, C., Li, O., Tao, D., Barnett, A., Rudin, C., and Su, J. K. This looks like that: deep learning for interpretable image recognition. Advances in neural information processing systems, 32, 2019

    2019

  5. [5]

    Davies, D. L. and Bouldin, D. W. A cluster separation measure. IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-1 0 (2): 0 224--227, 1979. doi:10.1109/TPAMI.1979.4766909

    work page doi:10.1109/tpami.1979.4766909 1979

  6. [6]

    Causally reliable concept bottleneck models

    De Felice, G., Casanova Flores, A., De Santis, F., Santini, S., Schneider, J., Barbiero, P., and Termine, A. Causally reliable concept bottleneck models. Advances in Neural Information Processing Systems, 38: 0 149099--149139, 2026

    2026

  7. [7]

    Fca2vec: Embedding techniques for formal concept analysis

    Durrschnabel, D., Hanika, T., and Stubbemann, M. Fca2vec: Embedding techniques for formal concept analysis. In Complex Data Analytics with Formal Concept Analysis, 2019. URL https://api.semanticscholar.org/CorpusID:208291462

    2019

  8. [8]

    and Lov \'a sz, L

    Erd o s, P. and Lov \'a sz, L. Problems and results on 3-chromatic hypergraphs and some related questions. In Infinite and F inite S ets ( C olloq., K eszthely, 1973; dedicated to P . E rd o s on his 60th birthday) , volume 10 of Colloq. Math. Soc. J \'a nos Bolyai , pp.\ 609--627. North-Holland, Amsterdam, 1975

    1973

  9. [9]

    and Lov\'asz, L

    Erdos, P. and Lov\'asz, L. Problems and results on 3 -chromatic hypergraphs and some related questions. In Infinite and finite sets ( C olloq., K eszthely, 1973; dedicated to P . E rd os on his 60th birthday), V ols. I , II , III , volume Vol. 10 of Colloq. Math. Soc. J\'anos Bolyai, pp.\ 609--627. North-Holland, Amsterdam-London, 1975

    1973

  10. [10]

    Hyperbolic entailment cones for learning hierarchical embeddings

    Ganea, O., Becigneul, G., and Hofmann, T. Hyperbolic entailment cones for learning hierarchical embeddings. In Dy, J. and Krause, A. (eds.), Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pp.\ 1646--1655. PMLR, 10--15 Jul 2018. URL https://proceedings.mlr.press/v80/ganea18a.html

    2018

  11. [11]

    and Wille, R

    Ganter, B. and Wille, R. Formal concept analysis: mathematical foundations. Springer Nature, 2024

    2024

  12. [12]

    and Lukasiewicz, T

    Giunchiglia, E. and Lukasiewicz, T. Coherent hierarchical multi-label classification networks. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS '20, Red Hook, NY, USA, 2020. Curran Associates Inc. ISBN 9781713829546

    2020

  13. [13]

    and Lukasiewicz, T

    Giunchiglia, E. and Lukasiewicz, T. Multi-label classification neural networks with hard logical constraints. J. Artif. Int. Res., 72: 0 759–818, January 2022. ISSN 1076-9757. doi:10.1613/jair.1.12850. URL https://doi.org/10.1613/jair.1.12850

    work page doi:10.1613/jair.1.12850 2022

  14. [14]

    Gyurek, C., Talukder, N., and Hasan, M. A. Binder: Hierarchical concept representation through order embedding of binary vectors. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp.\ 980--991, 2024

    2024

  15. [15]

    Harvey, N. J. and Vondrák, J. An algorithmic proof of the lovasz local lemma via resampling oracles. In 2015 IEEE 56th Annual Symposium on Foundations of Computer Science, pp.\ 1327--1346, 2015. doi:10.1109/FOCS.2015.85

    work page doi:10.1109/focs.2015.85 2015

  16. [16]

    and Saul, L

    Kearns, M. and Saul, L. Large deviation methods for approximate probabilistic inference. In Proceedings of the Fourteenth Conference on Uncertainty in Artificial Intelligence, UAI'98, pp.\ 311–319, San Francisco, CA, USA, 1998. Morgan Kaufmann Publishers Inc. ISBN 155860555X

    1998

  17. [17]

    Probabilistic concept bottleneck models

    Kim, E., Jung, D., Park, S., Kim, S., and Yoon, S.-H. Probabilistic concept bottleneck models. ArXiv, abs/2306.01574, 2023. URL https://api.semanticscholar.org/CorpusID:259063823

    arXiv 2023

  18. [18]

    W., Nguyen, T., Tang, Y

    Koh, P. W., Nguyen, T., Tang, Y. S., Mussmann, S., Pierson, E., Kim, B., and Liang, P. Concept bottleneck models. In International conference on machine learning, pp.\ 5338--5348. PMLR, 2020

    2020

  19. [19]

    Learning multiple layers of features from tiny images

    Krizhevsky, A., Hinton, G., et al. Learning multiple layers of features from tiny images. 2009

    2009

  20. [20]

    Kuznetsov, S. O. On stability of a formal concept. Annals of Mathematics and Artificial Intelligence, 49 0 (1): 0 101--115, 2007

    2007

  21. [21]

    Logicseg: Parsing visual semantics with neural logic learning and reasoning

    Li, L., Wang, W., and Yang, Y. Logicseg: Parsing visual semantics with neural logic learning and reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp.\ 4122--4133, October 2023

    2023

  22. [22]

    Hybrid concept bottleneck models

    Liu, Y., Zhang, T., and Gu, S. Hybrid concept bottleneck models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pp.\ 20179--20189, 2025

    2025

  23. [23]

    Glancenets: Interpretabile, leak-proof concept-based models

    Marconato, E., Passerini, A., and Teso, S. Glancenets: Interpretabile, leak-proof concept-based models. In Neural Information Processing Systems, 2022. URL https://api.semanticscholar.org/CorpusID:249209924

    2022

  24. [24]

    and Kiela, D

    Nickel, M. and Kiela, D. Poincar\' e embeddings for learning hierarchical representations. In Guyon, I., Luxburg, U. V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., and Garnett, R. (eds.), Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. URL https://proceedings.neurips.cc/paper_files/paper/2017/file/5...

    2017

  25. [25]

    M., and Weng, T.-W

    Oikarinen, T., Das, S., Nguyen, L. M., and Weng, T.-W. Label-free concept bottleneck models. arXiv preprint arXiv:2304.06129, 2023

    arXiv 2023

  26. [26]

    The building blocks of interpretability

    Olah, C., Satyanarayan, A., Johnson, I., Carter, S., Schubert, L., Ye, K., and Mordvintsev, A. The building blocks of interpretability. Distill, 3 0 (3): 0 e10, 2018

    2018

  27. [27]

    P., Ienco, D., and Marcos, D

    Panousis, K. P., Ienco, D., and Marcos, D. Coarse-to-fine concept bottleneck models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/forum?id=RMdnTnffou

    2024

  28. [28]

    J., Jiang, Y., and Veitch, V

    Park, K., Choe, Y. J., Jiang, Y., and Veitch, V. The geometry of categorical and hierarchical concepts in large language models. In International Conference on Learning Representations, volume 2025, pp.\ 76441--76463, 2025

    2025

  29. [29]

    Bert4fca: A method for bipartite link prediction using formal concept analysis and bert

    Peng, S., Yang, H., and Yamamoto, A. Bert4fca: A method for bipartite link prediction using formal concept analysis and bert. Plos one, 19 0 (6): 0 e0304858, 2024

    2024

  30. [30]

    J., Espinosa Zarlenga, M., and Jamnik, M

    Raman, N. J., Espinosa Zarlenga, M., and Jamnik, M. Understanding inter-concept relationships in concept-based models. In Salakhutdinov, R., Kolter, Z., Heller, K., Weller, A., Oliver, N., Scarlett, J., and Berkenkamp, F. (eds.), Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, p...

    2024

  31. [31]

    Using fca for encoding closure operators into neural networks

    Rudolph, S. Using fca for encoding closure operators into neural networks. In International Conference on Conceptual Structures, pp.\ 321--332. Springer, 2007

    2007

  32. [32]

    Bernstein, Alexander C

    Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A. C., and Fei-Fei, L. ImageNet Large Scale Visual Recognition Challenge . International Journal of Computer Vision (IJCV), 115 0 (3): 0 211--252, 2015. doi:10.1007/s11263-015-0816-y

    work page doi:10.1007/s11263-015-0816-y 2015

  33. [33]

    Sarkar, A., Vijaykeerthy, D., Sarkar, A., and Balasubramanian, V. N. A framework for learning ante-hoc explainable models via concepts. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 10276--10285, 2022

    2022

  34. [34]

    and Nakamura, K

    Sawada, Y. and Nakamura, K. Concept bottleneck model with additional unsupervised concepts. IEEE Access, 10: 0 41758--41765, 2022. doi:10.1109/ACCESS.2022.3167702

    work page doi:10.1109/access.2022.3167702 2022

  35. [35]

    Understanding and enhancing robustness of concept-based models

    Sinha, S., Huai, M., Sun, J., and Zhang, A. Understanding and enhancing robustness of concept-based models. In AAAI Conference on Artificial Intelligence, 2022. URL https://api.semanticscholar.org/CorpusID:254069697

    2022

  36. [36]

    Vlg-cbm: Training concept bottleneck models with vision-language guidance

    Srivastava, D., Yan, G., and Weng, T.-W. Vlg-cbm: Training concept bottleneck models with vision-language guidance. In Globerson, A., Mackey, L., Belgrave, D., Fan, A., Paquet, U., Tomczak, J., and Zhang, C. (eds.), Advances in Neural Information Processing Systems, volume 37, pp.\ 79057--79094. Curran Associates, Inc., 2024. URL https://proceedings.neuri...

    2024

  37. [37]

    Eliminating information leakage in hard concept bottleneck models with supervised, hierarchical concept learning

    Sun, A., Yuan, Y., Ma, P., and Wang, S. Eliminating information leakage in hard concept bottleneck models with supervised, hierarchical concept learning. arXiv preprint arXiv:2402.05945, 2024

    arXiv 2024

  38. [38]

    A., Fern \'a ndez, G., and Doeller, C

    Theves, S., Neville, D. A., Fern \'a ndez, G., and Doeller, C. F. Learning and representation of hierarchical concepts in hippocampus and prefrontal cortex. Journal of Neuroscience, 41 0 (36): 0 7675--7686, 2021

    2021

  39. [39]

    and Zaslavsky, N

    Tishby, N. and Zaslavsky, N. Deep learning and the information bottleneck principle, 2015. URL https://arxiv.org/abs/1503.02406

    Pith/arXiv arXiv 2015

  40. [40]

    D., Zhang, G.-Q., and Tian, Y

    Troy, A. D., Zhang, G.-Q., and Tian, Y. Faster concept analysis. In Proceedings of the 15th International Conference on Conceptual Structures: Knowledge Architectures for Smart Applications, ICCS '07, pp.\ 206–219, Berlin, Heidelberg, 2007. Springer-Verlag. ISBN 9783540736806. doi:10.1007/978-3-540-73681-3_16. URL https://doi.org/10.1007/978-3-540-73681-3_16

    work page doi:10.1007/978-3-540-73681-3_16 2007

  41. [41]

    Stochastic concept bottleneck models

    Vandenhirtz, M., Laguna, S., Marcinkevi c s, R., and Vogt, J. Stochastic concept bottleneck models. Advances in Neural Information Processing Systems, 37: 0 51787--51810, 2024 a

    2024

  42. [42]

    Stochastic concept bottleneck models

    Vandenhirtz, M., Laguna, S., Marcinkevi c s, R., and Vogt, J. Stochastic concept bottleneck models. Advances in Neural Information Processing Systems, 37: 0 51787--51810, 2024 b

    2024

  43. [43]

    Order-embeddings of images and language

    Vendrov, I., Kiros, R., Fidler, S., and Urtasun, R. Order-embeddings of images and language. arXiv preprint arXiv:1511.06361, 2015

    Pith/arXiv arXiv 2015

  44. [44]

    Lampert, Bernt Schiele, and Zeynep Akata

    Xian, Y., Lampert, C. H., Schiele, B., and Akata, Z. Zero-shot learning—a comprehensive evaluation of the good, the bad and the ugly. IEEE Transactions on Pattern Analysis &; Machine Intelligence, 41 0 (09): 0 2251--2265, sep 2019. ISSN 1939-3539. doi:10.1109/TPAMI.2018.2857768

    work page doi:10.1109/tpami.2018.2857768 2019

  45. [45]

    Language in a bottle: Language model guided concept bottlenecks for interpretable image classification

    Yang, Y., Panagopoulou, A., Zhou, S., Jin, D., Callison-Burch, C., and Yatskar, M. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 19187--19197, 2023

    2023

  46. [46]

    Post-hoc concept bottleneck models

    Yuksekgonul, M., Wang, M., and Zou, J. Post-hoc concept bottleneck models. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=nA5AZ8CEyow

    2023

  47. [47]

    E., Barbiero, P., Ciravegna, G., Marra, G., Giannini, F., Diligenti, M., Shams, Z., Precioso, F., Melacci, S., Weller, A., Lio, P., and Jamnik, M

    Zarlenga, M. E., Barbiero, P., Ciravegna, G., Marra, G., Giannini, F., Diligenti, M., Shams, Z., Precioso, F., Melacci, S., Weller, A., Lio, P., and Jamnik, M. Concept embedding models. In Oh, A. H., Agarwal, A., Belgrave, D., and Cho, K. (eds.), Advances in Neural Information Processing Systems, 2022. URL https://openreview.net/forum?id=HXCPA2GXf_

    2022

  48. [48]

    Zeiler, M. D. and Fergus, R. Visualizing and understanding convolutional networks. In Computer Vision--ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13, pp.\ 818--833. Springer, 2014

    2014

  49. [49]

    2021 , issue_date =

    Zhang, C., Bengio, S., Hardt, M., Recht, B., and Vinyals, O. Understanding deep learning (still) requires rethinking generalization. Commun. ACM, 64 0 (3): 0 107–115, February 2021. ISSN 0001-0782. doi:10.1145/3446776. URL https://doi.org/10.1145/3446776

    work page doi:10.1145/3446776 2021