pith. sign in

arxiv: 2605.16383 · v1 · pith:2Q6GOGGEnew · submitted 2026-05-11 · 💻 cs.CV · cs.AI· stat.ML

A neurosymbolic Approach with Epistemic Deep Learning for Hierarchical Image Classification

Ezel Kilicdere , Shireen Kudukkil Manchingal , Fabio Cuzzolin This is my paper

Pith reviewed 2026-05-20 22:42 UTC · model grok-4.3

classification 💻 cs.CV cs.AIstat.ML
keywords neurosymbolic AIepistemic uncertaintyhierarchical image classificationfocal setsfuzzy logicSwin Transformersbelief theorycalibrated predictions
0
0 comments X

The pith

Augmenting Swin Transformers with focal set reasoning and fuzzy logic yields accurate hierarchical image classification that captures epistemic uncertainty and enforces logical consistency.

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

The paper establishes a neurosymbolic framework that adds focal set reasoning and differentiable fuzzy logic to Swin Transformers for hierarchical image classification. It induces data-driven focal sets in the embedding space to represent uncertainty over several plausible fine-grained classes, then routes them through a belief-theoretic layer that applies fuzzy membership functions and t-norm conjunctions to keep fine- and coarse-level predictions coherent. A learnable loss balances calibration, mass regularisation and logical consistency so the model can trade off symbolic structure against data evidence. A sympathetic reader would care because ordinary neural networks routinely produce overconfident and mutually inconsistent outputs when labels form hierarchies.

Core claim

Combining focal set reasoning with fuzzy logic supplies a practical step toward deep learning models that are both accurate and epistemically aware: the method maintains accuracy on par with transformer baselines while delivering more calibrated predictions and enforcing high logical consistency across hierarchical outputs.

What carries the argument

Data-driven focal sets induced inside the learnt embedding space, which capture epistemic uncertainty over multiple plausible fine-grained classes and supply the inputs to a belief-theoretic layer that uses fuzzy membership functions together with t-norm conjunctions to enforce consistency between fine- and coarse-grained predictions.

If this is right

  • Accuracy stays comparable to unmodified Swin Transformer baselines on hierarchical image tasks.
  • Predictions become more calibrated and reduce overconfidence on uncertain examples.
  • Logical consistency between fine-grained and coarse-grained outputs is enforced directly inside the model.
  • A single learnable loss lets the system adaptively balance data evidence against symbolic constraints.

Where Pith is reading between the lines

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

  • The same focal-set-plus-fuzzy-logic layer could be attached to other transformer backbones for structured prediction tasks that require both accuracy and constraint satisfaction.
  • Testing the approach on non-image hierarchical domains such as document classification or medical diagnosis would reveal whether the embedding-space focal sets generalise beyond visual features.
  • If focal sets prove robust, the method offers a route to embed epistemic uncertainty handling inside existing large-scale vision models without retraining from scratch.

Load-bearing premise

Data-driven focal sets in the embedding space combined with t-norm conjunctions in a belief-theoretic layer will reliably capture epistemic uncertainty and enforce hierarchical consistency without post-hoc adjustments or loss of accuracy.

What would settle it

An experiment on a standard hierarchical image dataset in which the augmented model either falls below baseline accuracy or produces fine- and coarse-level predictions that violate the logical hierarchy despite the belief layer.

Figures

Figures reproduced from arXiv: 2605.16383 by Ezel Kilicdere, Fabio Cuzzolin, Shireen Kudukkil Manchingal.

Figure 1
Figure 1. Figure 1: Example image from iNaturalist dataset showing fine (family) and coarse (class) labels. Hierarchical image classification tasks organ￾ise labels across multiple levels of abstrac￾tion, such as fine-grained categories Y f = {1, . . . , nf } and coarse categories Y c = {1, . . . , nc}, linked by a deterministic map￾ping π : Y f → Y c . Datasets such as CIFAR-100 (Krizhevsky, 2012) and iNatural￾ist (Van Horn … view at source ↗
Figure 2
Figure 2. Figure 2: Computation flow of the Belief￾Based Logically Constrained Loss. The diagram shows how fine-level epistemic masses, coarse￾level fuzzified masses, semantic compatibility, specificity weighting, and t-norm interaction com￾bine to produce per-pair scores, semantic consis￾tency, and the final training loss. This design aligns with the structure of Generalised Modus Ponens (GMP) in fuzzy logic (Goguen, 1973; C… view at source ↗
read the original abstract

Deep neural networks achieve high accuracy on image classification tasks. Yet, they often produce overconfident predictions as which fail to express epistemic uncertainty, and frequently violate logical or structural constraints present in the data. These limitations are particularly pronounced in hierarchical classification, where predictions across fine and coarse levels must remain coherent. We propose, for the first time, a unified neurosymbolic and epistemic modelling framework that augments Swin Transformers with focal set reasoning and differentiable fuzzy logic. Rather than treating labels as isolated categories, our method induces data-driven focal sets within the learnt embedding space, which helps capture epistemic uncertainty over multiple plausible fine-grained classes. These focal sets form the basis of a belief-theoretic layer that uses fuzzy membership functions and t-norm conjunctions to encourage consistency between fine- and coarse-grained predictions. A learnable loss further balances calibration, mass regularisation, and logical consistency, allowing the model to adaptively trade off symbolic structure with data-driven evidence. In experiments on hierarchical image classification, our framework maintains accuracy on par with transformer baselines while providing more calibrated and interpretable predictions, reducing overconfidence and enforcing high logical consistency across hierarchical outputs. Our experimental results show that combining focal set reasoning with fuzzy logic provides a practical step toward deep learning models that are both accurate and epistemically aware.

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

3 major / 2 minor

Summary. The paper proposes a neurosymbolic epistemic framework that augments Swin Transformers with data-driven focal sets induced in the embedding space and a belief-theoretic layer using fuzzy membership functions and t-norm conjunctions. This is intended to capture epistemic uncertainty over plausible fine-grained classes while enforcing logical consistency between fine- and coarse-grained hierarchical predictions. A learnable loss balances calibration, mass regularisation, and logical consistency. Experiments claim that the approach maintains accuracy comparable to transformer baselines while delivering more calibrated predictions, reduced overconfidence, and high logical consistency across hierarchical outputs.

Significance. If the quantitative results and separation of epistemic from aleatoric uncertainty hold, the work offers a practical route to uncertainty-aware hierarchical classifiers that integrate symbolic constraints without sacrificing accuracy. The adaptive loss and differentiable fuzzy logic components are strengths that could generalise to other structured prediction tasks.

major comments (3)
  1. [Abstract] Abstract: the central claim that the method 'maintains accuracy on par with transformer baselines while providing more calibrated and interpretable predictions' is asserted without any reported accuracy numbers, error bars, dataset details, or baseline comparisons. This absence makes the no-accuracy-loss claim impossible to evaluate and is load-bearing for the paper's contribution.
  2. [Method / Focal Set Reasoning] The construction of focal sets from a single deterministic embedding (described in the method) is presented as capturing epistemic uncertainty, yet the manuscript provides no ensemble, posterior sampling, or OOD evaluation to distinguish this from aleatoric uncertainty or label ambiguity. This distinction is required for the epistemic-awareness claim.
  3. [Loss Function] The learnable loss that balances calibration, mass regularisation, and logical consistency (mentioned in the abstract and method) risks circularity: the consistency term appears to be optimised by parameters that also define the metric being reported, which could artificially inflate the 'high logical consistency' result.
minor comments (2)
  1. [Abstract] Abstract contains a grammatical error: 'overconfident predictions as which fail' should be 'overconfident predictions that fail'.
  2. [Introduction] The claim 'for the first time' is made without a literature comparison; a brief related-work paragraph would clarify novelty.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. The comments have helped us clarify key aspects of the contribution. We respond to each major comment below and have revised the manuscript where appropriate to address the concerns.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that the method 'maintains accuracy on par with transformer baselines while providing more calibrated and interpretable predictions' is asserted without any reported accuracy numbers, error bars, dataset details, or baseline comparisons. This absence makes the no-accuracy-loss claim impossible to evaluate and is load-bearing for the paper's contribution.

    Authors: We agree that the abstract would be strengthened by including quantitative support for the claims. In the revised manuscript, we have updated the abstract to report specific accuracy values with standard deviations, the datasets used (CIFAR-100 and a hierarchical subset of ImageNet), and direct comparisons to the Swin Transformer baseline and other hierarchical methods. revision: yes

  2. Referee: [Method / Focal Set Reasoning] The construction of focal sets from a single deterministic embedding (described in the method) is presented as capturing epistemic uncertainty, yet the manuscript provides no ensemble, posterior sampling, or OOD evaluation to distinguish this from aleatoric uncertainty or label ambiguity. This distinction is required for the epistemic-awareness claim.

    Authors: The focal sets are induced directly from the embedding to assign belief mass to sets of plausible fine-grained classes, which by construction models epistemic uncertainty over which class is correct rather than label noise. This follows the epistemic interpretation in belief theory, where uncertainty is expressed via non-singleton focal elements. To address the distinction, we have added a dedicated paragraph in the Method section and included an OOD detection experiment demonstrating that the uncertainty mass behaves differently from aleatoric cases. While ensembles are not used, the single-pass focal set construction provides the epistemic separation without additional sampling. revision: partial

  3. Referee: [Loss Function] The learnable loss that balances calibration, mass regularisation, and logical consistency (mentioned in the abstract and method) risks circularity: the consistency term appears to be optimised by parameters that also define the metric being reported, which could artificially inflate the 'high logical consistency' result.

    Authors: The learnable parameters adaptively weight the loss terms during training, but the reported logical consistency metric is a fixed, post-training evaluation that counts the fraction of predictions satisfying the hierarchical constraints using an independent verification procedure. This metric does not depend on the loss parameters. We have clarified this separation in the revised Loss Function section and added an ablation that isolates the effect of the consistency term on both training dynamics and final reported consistency. revision: yes

Circularity Check

1 steps flagged

Learnable loss incorporates logical consistency as an optimized term, partially reducing the consistency claim to a fitted outcome

specific steps
  1. fitted input called prediction [Abstract]
    "A learnable loss further balances calibration, mass regularisation, and logical consistency, allowing the model to adaptively trade off symbolic structure with data-driven evidence."

    The logical consistency metric is one of the terms directly optimized within the learnable loss; therefore the reported enforcement of high logical consistency is a direct consequence of including and fitting that term, rather than a separate prediction derived from the focal-set or fuzzy-logic components.

full rationale

The paper's central neurosymbolic framework induces focal sets from embeddings and applies t-norm conjunctions in a belief layer, with a learnable loss that explicitly includes logical consistency as one of its balanced terms. This setup allows the model to trade off components during training, but the reported high logical consistency across hierarchical outputs is achieved by direct optimization of that term rather than emerging as an independent derivation from epistemic modeling alone. No self-citations or uniqueness theorems are load-bearing in the provided text; the derivation remains partially self-contained via the explicit loss design, though the epistemic uncertainty claim rests on the focal-set construction without separate verification against aleatoric alternatives. This yields moderate circularity without full reduction to inputs by definition.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 2 invented entities

The framework rests on several introduced components whose independence from prior literature is not demonstrated in the abstract; focal sets and the belief-theoretic layer appear as new constructs without external validation.

free parameters (1)
  • loss balancing weights
    Learnable loss that adaptively trades off calibration, mass regularisation, and logical consistency.
axioms (2)
  • domain assumption Focal sets induced in the learnt embedding space can capture epistemic uncertainty over multiple plausible fine-grained classes.
    Central to the neurosymbolic layer described in the abstract.
  • domain assumption Fuzzy membership functions and t-norm conjunctions can enforce consistency between fine- and coarse-grained predictions.
    Basis of the belief-theoretic layer.
invented entities (2)
  • focal sets no independent evidence
    purpose: Capture epistemic uncertainty over multiple fine-grained classes in embedding space
    Data-driven constructs introduced to augment the transformer.
  • belief-theoretic layer no independent evidence
    purpose: Use fuzzy logic to enforce hierarchical consistency
    New layer added on top of the transformer output.

pith-pipeline@v0.9.0 · 5772 in / 1438 out tokens · 28522 ms · 2026-05-20T22:42:31.424587+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

300 extracted references · 300 canonical work pages · 7 internal anchors

  1. [1]

    2025 , eprint=

    Neuro-Symbolic AI in 2024: A Systematic Review , author=. 2025 , eprint=

    work page 2024

  2. [2]

    arXiv preprint arXiv:2504.07640 , year=

    Enhancing large language models through neuro-symbolic integration and ontological reasoning , author=. arXiv preprint arXiv:2504.07640 , year=

    work page arXiv

  3. [3]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Swin transformer: Hierarchical vision transformer using shifted windows , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

    work page

  4. [4]

    IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics) , volume=

    Two new Bayesian approximations of belief functions based on convex geometry , author=. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics) , volume=. 2007 , publisher=

    work page 2007

  5. [5]

    Fuzzy sets and systems , volume=

    Theory of T-norms and fuzzy inference methods , author=. Fuzzy sets and systems , volume=. 1991 , publisher=

    work page 1991

  6. [6]

    2011 , publisher=

    Fuzzy set theory—and its applications , author=. 2011 , publisher=

    work page 2011

  7. [7]

    arXiv preprint arXiv:2501.03282 , year=

    From aleatoric to epistemic: Exploring uncertainty quantification techniques in artificial intelligence , author=. arXiv preprint arXiv:2501.03282 , year=

    work page arXiv

  8. [8]

    2017 , eprint=

    Learning to Reason: End-to-End Module Networks for Visual Question Answering , author=. 2017 , eprint=

    work page 2017

  9. [9]

    DeepProbLog: Neural Probabilistic Logic Programming , url =

    Manhaeve, Robin and Dumancic, Sebastijan and Kimmig, Angelika and Demeester, Thomas and De Raedt, Luc , booktitle =. DeepProbLog: Neural Probabilistic Logic Programming , url =

    work page

  10. [10]

    2023 , eprint=

    Neurosymbolic AI -- Why, What, and How , author=. 2023 , eprint=

    work page 2023

  11. [11]

    2021 , eprint=

    Neuro-Symbolic AI: An Emerging Class of AI Workloads and their Characterization , author=. 2021 , eprint=

    work page 2021

  12. [12]

    2024 , eprint=

    Neuro-Symbolic AI: Explainability, Challenges, and Future Trends , author=. 2024 , eprint=

    work page 2024

  13. [13]

    2020 , eprint=

    GLUCOSE: GeneraLized and COntextualized Story Explanations , author=. 2020 , eprint=

    work page 2020

  14. [14]

    A Survey on Verification and Validation, Testing and Evaluations of Neurosymbolic Artificial Intelligence , volume=

    Renkhoff, Justus and Feng, Ke and Meier-Doernberg, Marc and Velasquez, Alvaro and Song, Houbing Herbert , year=. A Survey on Verification and Validation, Testing and Evaluations of Neurosymbolic Artificial Intelligence , volume=. IEEE Transactions on Artificial Intelligence , publisher=. doi:10.1109/tai.2024.3351798 , number=

    work page doi:10.1109/tai.2024.3351798 2024

  15. [15]

    arXiv preprint arXiv:2401.01040 , year=

    Towards cognitive ai systems: a survey and prospective on neuro-symbolic ai , author=. arXiv preprint arXiv:2401.01040 , year=

    work page arXiv

  16. [16]

    2024 , eprint=

    Towards Data-and Knowledge-Driven Artificial Intelligence: A Survey on Neuro-Symbolic Computing , author=. 2024 , eprint=

    work page 2024

  17. [17]

    Multisensor, Multisource Information Fusion: Architectures, Algorithms, and Applications 2003 , editor =

    Florea, Mihai Cristian and Jousselme, Anne-Laure and Grenier, Dominic and Bosse, Eloi , title =. Multisensor, Multisource Information Fusion: Architectures, Algorithms, and Applications 2003 , editor =. 2003 , doi =

    work page 2003

  18. [18]

    T-norms driven loss functions for machine learning , volume=

    Giannini, Francesco and Diligenti, Michelangelo and Maggini, Marco and Gori, Marco and Marra, Giuseppe , year=. T-norms driven loss functions for machine learning , volume=. Applied Intelligence , publisher=. doi:10.1007/s10489-022-04383-6 , number=

    work page doi:10.1007/s10489-022-04383-6

  19. [19]

    Logistic regression, neural networks and Dempster–Shafer theory: A new perspective , volume=

    Denœux, Thierry , year=. Logistic regression, neural networks and Dempster–Shafer theory: A new perspective , volume=. doi:10.1016/j.knosys.2019.03.030 , journal=

    work page doi:10.1016/j.knosys.2019.03.030 2019

  20. [20]

    2013 , eprint=

    Hierarchical Evidence and Belief Functions , author=. 2013 , eprint=

    work page 2013

  21. [21]

    2019 , eprint=

    On the relation between Loss Functions and T-Norms , author=. 2019 , eprint=

    work page 2019

  22. [22]

    1995 , eprint=

    Learning Membership Functions in a Function-Based Object Recognition System , author=. 1995 , eprint=

    work page 1995

  23. [23]

    T-norm and uninorm-based combination of belief functions , year=

    Pichon, Frederic and Denoeux, Thierry , booktitle=. T-norm and uninorm-based combination of belief functions , year=

    work page

  24. [24]

    Reasoning with fuzzy and uncertain evidence using epistemic random fuzzy sets: General framework and practical models , volume=

    Denœux, Thierry , year=. Reasoning with fuzzy and uncertain evidence using epistemic random fuzzy sets: General framework and practical models , volume=. doi:10.1016/j.fss.2022.06.004 , journal=

    work page doi:10.1016/j.fss.2022.06.004 2022

  25. [25]

    Belief functions induced by random fuzzy sets: A general framework for representing uncertain and fuzzy evidence , volume=

    Denœux, Thierry , year=. Belief functions induced by random fuzzy sets: A general framework for representing uncertain and fuzzy evidence , volume=. doi:10.1016/j.fss.2020.12.004 , journal=

    work page doi:10.1016/j.fss.2020.12.004 2020

  26. [26]

    Fuzzy-Set Based Logics — an History-Oriented Presentation of their Main Developments , volume =

    Dubois, Didier and Esteva, Francesc and Godo, Lluis and Prade, Henri , year =. Fuzzy-Set Based Logics — an History-Oriented Presentation of their Main Developments , volume =. Handbook of the History of Logic , doi =

    work page

  27. [27]

    The Unnormalized Dempster’s Rule of Combination: A New Justification from the Least Commitment Principle and Some Extensions , volume =

    Pichon, Frédéric and Denœux, Thierry , year =. The Unnormalized Dempster’s Rule of Combination: A New Justification from the Least Commitment Principle and Some Extensions , volume =. Journal of Automated Reasoning , doi =

    work page

  28. [28]

    Belief functions

    Cuzzolin, Fabio. Belief functions. The Geometry of Uncertainty: The Geometry of Imprecise Probabilities. 2021. doi:10.1007/978-3-030-63153-6_2

    work page doi:10.1007/978-3-030-63153-6_2 2021

  29. [29]

    HCAM-CL: A Novel Method Integrating a Hierarchical Cross-Attention Mechanism with CNN-LSTM for Hierarchical Image Classification , volume =

    Su, Jing and Liang, Jianmin and Zhu, Jiayi and Li, Yongjiang , year =. HCAM-CL: A Novel Method Integrating a Hierarchical Cross-Attention Mechanism with CNN-LSTM for Hierarchical Image Classification , volume =. Symmetry , doi =

    work page

  30. [30]

    2024 , eprint=

    Informed deep hierarchical classification: a non-standard analysis inspired approach , author=. 2024 , eprint=

    work page 2024

  31. [31]

    2021 , eprint=

    MultiplexNet: Towards Fully Satisfied Logical Constraints in Neural Networks , author=. 2021 , eprint=

    work page 2021

  32. [32]

    2024 , eprint=

    Learning with Logical Constraints but without Shortcut Satisfaction , author=. 2024 , eprint=

    work page 2024

  33. [33]

    Combination of dependent and partially reliable Gaussian random fuzzy numbers , journal =

    Denœux, Thierry , keywords =. Combination of dependent and partially reliable Gaussian random fuzzy numbers , journal =. 2024 , issn =. doi:https://doi.org/10.1016/j.ins.2024.121208 , url =

    work page doi:10.1016/j.ins.2024.121208 2024

  34. [34]

    An Evidential Neural Network Model for Regression Based on Random Fuzzy Numbers , ISBN=

    Denœux, Thierry , year=. An Evidential Neural Network Model for Regression Based on Random Fuzzy Numbers , ISBN=. doi:10.1007/978-3-031-17801-6_6 , booktitle=

    work page doi:10.1007/978-3-031-17801-6_6

  35. [35]

    2021 , eprint=

    Evaluating Relaxations of Logic for Neural Networks: A Comprehensive Study , author=. 2021 , eprint=

    work page 2021

  36. [36]

    2023 , MONTH = Jul, KEYWORDS =

    Ott, Johanna and Ledaguenel, Arthur and Hudelot, C. 2023 , MONTH = Jul, KEYWORDS =

    work page 2023

  37. [37]

    and Monahan, Rosemary , year=

    Flinkow, Thomas and Pearlmutter, Barak A. and Monahan, Rosemary , year=. Comparing differentiable logics for learning with logical constraints , volume=. doi:10.1016/j.scico.2025.103280 , journal=

    work page doi:10.1016/j.scico.2025.103280 2025

  38. [38]

    Regularizing deep networks with prior knowledge: A constraint-based approach , journal =

    Roychowdhury, Soumali and Diligenti, Michelangelo and Gori, Marco , keywords =. Regularizing deep networks with prior knowledge: A constraint-based approach , journal =. 2021 , issn =. doi:https://doi.org/10.1016/j.knosys.2021.106989 , url =

    work page doi:10.1016/j.knosys.2021.106989 2021

  39. [39]

    2025 , eprint=

    MSCViT: A Small-size ViT architecture with Multi-Scale Self-Attention Mechanism for Tiny Datasets , author=. 2025 , eprint=

    work page 2025

  40. [40]

    2021 , eprint=

    Swin Transformer: Hierarchical Vision Transformer using Shifted Windows , author=. 2021 , eprint=

    work page 2021

  41. [41]

    2024 , eprint=

    GCI-ViTAL: Gradual Confidence Improvement with Vision Transformers for Active Learning on Label Noise , author=. 2024 , eprint=

    work page 2024

  42. [42]

    2020 , eprint=

    Integrating domain knowledge: using hierarchies to improve deep classifiers , author=. 2020 , eprint=

    work page 2020

  43. [43]

    SparseSwin: Swin transformer with sparse transformer block , volume=

    Pinasthika, Krisna and Laksono, Blessius Sheldo Putra and Irsal, Riyandi Banovbi Putera and Shabiyya, Syifa’ Hukma and Yudistira, Novanto , year=. SparseSwin: Swin transformer with sparse transformer block , volume=. doi:10.1016/j.neucom.2024.127433 , journal=

    work page doi:10.1016/j.neucom.2024.127433 2024

  44. [44]

    1998 , publisher=

    Metamathematics of Fuzzy Logic , author=. 1998 , publisher=

    work page 1998

  45. [45]

    2022 , eprint=

    Survey on Applications of Neurosymbolic Artificial Intelligence , author=. 2022 , eprint=

    work page 2022

  46. [46]

    2022 , eprint=

    Some properties of fuzzy t -norm and vague t -norm , author=. 2022 , eprint=

    work page 2022

  47. [47]

    2021 , eprint=

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale , author=. 2021 , eprint=

    work page 2021

  48. [48]

    Efficient adaptive ensembling for image classification , volume=

    Antonio, Bruno and Moroni, Davide and Martinelli, Massimo , year=. Efficient adaptive ensembling for image classification , volume=. Expert Systems , publisher=. doi:10.1111/exsy.13424 , number=

    work page doi:10.1111/exsy.13424

  49. [49]

    2019 , eprint=

    Explainable Artificial Intelligence (XAI): Concepts, Taxonomies, Opportunities and Challenges toward Responsible AI , author=. 2019 , eprint=

    work page 2019

  50. [50]

    and Prade, H

    Dubois, D. and Prade, H. , keywords =. Upper and Lower Possibilities Induced by a Multivalued Mapping , journal =. 1983 , note =. doi:https://doi.org/10.1016/S1474-6670(17)62022-6 , url =

    work page doi:10.1016/s1474-6670(17)62022-6 1983

  51. [51]

    2023 , eprint=

    GLUECons: A Generic Benchmark for Learning Under Constraints , author=. 2023 , eprint=

    work page 2023

  52. [52]

    2019 , eprint=

    'In-Between' Uncertainty in Bayesian Neural Networks , author=. 2019 , eprint=

    work page 2019

  53. [53]

    2016 , eprint=

    Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning , author=. 2016 , eprint=

    work page 2016

  54. [54]

    ROAD-R: the autonomous driving dataset with logical requirements , volume=

    Giunchiglia, Eleonora and Stoian, Mihaela Cătălina and Khan, Salman and Cuzzolin, Fabio and Lukasiewicz, Thomas , year=. ROAD-R: the autonomous driving dataset with logical requirements , volume=. Machine Learning , publisher=. doi:10.1007/s10994-023-06322-z , number=

    work page doi:10.1007/s10994-023-06322-z

  55. [55]

    Deep Learning , author=

    work page

  56. [56]

    2017 , eprint=

    On Calibration of Modern Neural Networks , author=. 2017 , eprint=

    work page 2017

  57. [57]

    timm - PyTorch Image Models

    Hugging Face. timm - PyTorch Image Models

    work page

  58. [58]

    2023 , isbn =

    Jo, Suhyeon and Shin, DongHyeok and Na, Byeonghu and Jang, JoonHo and Moon, Il-Chul , title =. 2023 , isbn =. doi:10.1145/3583780.3614912 , booktitle =

    work page doi:10.1145/3583780.3614912 2023

  59. [59]

    2017 , eprint=

    What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision? , author=. 2017 , eprint=

    work page 2017

  60. [60]

    2017 , eprint=

    Adam: A Method for Stochastic Optimization , author=. 2017 , eprint=

    work page 2017

  61. [61]

    Learning Multiple Layers of Features from Tiny Images , journal =

    Krizhevsky, Alex , year =. Learning Multiple Layers of Features from Tiny Images , journal =

    work page

  62. [62]

    ImageNet Classification with Deep Convolutional Neural Networks , volume =

    Krizhevsky, Alex and Sutskever, Ilya and Hinton, Geoffrey , year =. ImageNet Classification with Deep Convolutional Neural Networks , volume =. Neural Information Processing Systems , doi =

    work page

  63. [63]

    The annals of mathematical statistics , keywords =

    Kullback, Solomon and Leibler, Richard A , biburl =. The annals of mathematical statistics , keywords =

    work page

  64. [64]

    Comprehensive Integration of Hyperdimensional Computing with Deep Learning towards Neuro-Symbolic AI , year=

    Lee, Hyunsei and Kim, Jiseung and Chen, Hanning and Zeira, Ariela and Srinivasa, Narayan and Imani, Mohsen and Kim, Yeseong , booktitle=. Comprehensive Integration of Hyperdimensional Computing with Deep Learning towards Neuro-Symbolic AI , year=

    work page

  65. [65]

    Symbolic Graph Reasoning Meets Convolutions , url =

    Liang, Xiaodan and Hu, Zhiting and Zhang, Hao and Lin, Liang and Xing, Eric P , booktitle =. Symbolic Graph Reasoning Meets Convolutions , url =

    work page

  66. [66]

    Surveying neuro-symbolic approaches for reliable artificial intelligence of things , volume =

    Lu, Zhen and Afridi, Imran and Kang, Hong and Ruchkin, Ivan and Zheng, Xi , year =. Surveying neuro-symbolic approaches for reliable artificial intelligence of things , volume =. Journal of Reliable Intelligent Environments , doi =

    work page

  67. [67]

    2019 , eprint=

    Can You Trust Your Model's Uncertainty? Evaluating Predictive Uncertainty Under Dataset Shift , author=. 2019 , eprint=

    work page 2019

  68. [68]

    2019 , eprint=

    PyTorch: An Imperative Style, High-Performance Deep Learning Library , author=. 2019 , eprint=

    work page 2019

  69. [69]

    Early Stopping - But When?

    Prechelt, Lutz. Early Stopping - But When?. Neural Networks: Tricks of the Trade. 1998. doi:10.1007/3-540-49430-8_3

    work page doi:10.1007/3-540-49430-8_3 1998

  70. [70]

    A Mathematical Theory of Evidence , urldate =

    Shafer, Glenn , publisher =. A Mathematical Theory of Evidence , urldate =

    work page

  71. [71]

    2015 , eprint=

    Very Deep Convolutional Networks for Large-Scale Image Recognition , author=. 2015 , eprint=

    work page 2015

  72. [72]

    2013 , eprint=

    The Transferable Belief Model and Other Interpretations of Dempster-Shafer's Model , author=. 2013 , eprint=

    work page 2013

  73. [73]

    2017 , eprint=

    Neural-Symbolic Learning and Reasoning: A Survey and Interpretation , author=. 2017 , eprint=

    work page 2017

  74. [74]

    Viualizing data using t-SNE , volume =

    van der Maaten, Laurens and Hinton, Geoffrey , year =. Viualizing data using t-SNE , volume =

    work page

  75. [75]

    2022 , eprint=

    Symbolic Learning to Optimize: Towards Interpretability and Scalability , author=. 2022 , eprint=

    work page 2022

  76. [76]

    2020 , eprint=

    Are we done with ImageNet? , author=. 2020 , eprint=

    work page 2020

  77. [77]

    2025 , eprint=

    The Impact of the Single-Label Assumption in Image Recognition Benchmarking , author=. 2025 , eprint=

    work page 2025

  78. [78]

    2021 , eprint=

    Multi-label Iterated Learning for Image Classification with Label Ambiguity , author=. 2021 , eprint=

    work page 2021

  79. [79]

    2024 , eprint=

    Diverse and Tailored Image Generation for Zero-shot Multi-label Classification , author=. 2024 , eprint=

    work page 2024

  80. [80]

    2023 , eprint=

    A Comprehensive Study of Vision Transformers in Image Classification Tasks , author=. 2023 , eprint=

    work page 2023

Showing first 80 references.