Pith. sign in

REVIEW 3 major objections 4 minor 74 references

Pretrain on Small Synthetic Data, Scale Large for Free: Symmetry-Aware Foundation Model for Logic Rule Induction

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A symmetry-respecting decoder lets a rule inducer trained on 6–12 variables keep around 0.9 accuracy at 1,024 variables, with exported rules exactly equivariant under reordering, polarity flips, and label swaps.

desk verdict A genuinely useful discrete-export primitive wrapped in a theorem whose empirical premise is only as solid as the unpublished NRI description. read the letter →

arxiv 2608.00383 v1 pith:76LMYLB7 submitted 2026-08-01 cs.LO cs.AIcs.LG

Yin Jun Phua This is my paper
classification cs.LOcs.AIcs.LG MSC 68T2768T05
keywords logicruleinductionDNFgroupequivariancesymmetrybyconstructionzero-shotgeneralizationneural-symboliclearningvariable-schemascalingcanonicalization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that the reason small-data rule inducers fail to scale is broken symmetry, not lack of data. Binary rule induction carries four natural symmetries — example order, atom naming, polarity flips, and label swap — and the paper proves that if an inducer's literal scores respect all of them, a carefully constructed canonical export decodes a discrete rule that transforms exactly as it should under every one of those transformations, with no retraining. The proof is architectural: the export never breaks a tie by atom index, assigns no polarity at an exact tie, and abstains rather than choosing arbitrarily. Instantiated on the Neural Rule Inducer, the construction holds support-label accuracy near 0.89–0.93 out to schemas of 1,024 variables — 85 times the training maximum — while the unmodified model falls to 0.487; on 19 real datasets the zero-shot accuracy gain rises with schema size. If the construction is right, exact symmetry converts a fixed small-scale checkpoint into a reusable inducer whose error tracks rule complexity rather than atom count.

What carries the argument

The canonical export is the load-bearing object. For each atom j it compresses two literal probabilities into a polarity-invariant presence score sj = max(p+j, p−j) and a sign-flipping contrast dj = p+j − p−j; an atom enters a clause only when sj ≥ ½ and dj ≠ 0, with polarity set by the sign of dj, so an exact polarity tie omits the atom rather than choosing. Every arbitrary choice is replaced by a symmetry-respecting abstention or whole-bucket admission: tie-complete pair selection keeps every pair that ties the k-th best scored pair, the decoder admits whole presence-score tie-buckets in canonical order, and a dual-rail selector chooses between the positive rail's rule and the negative rai

What would settle it

Train a G-NRI checkpoint from the released code, swap two atom columns of a synthetic episode at N=128, and measure the worst-case change in the literal scores (p+, p−): Proposition 1 predicts drift at the floating-point floor (≈10⁻⁶), and if the residual exceeds the smallest decode margin — or the 13-transform rule-eq audit falls below 1.000 — the exact-equivariance claim fails for the shipped model.

Watch

Extended reading notes

Core claim

The paper's central claim is Theorem 2: if a model's paired literal scores (p+, p−) are G-equivariant — invariant under example permutations, permuted under atom renamings, swapped between xj and ¬xj under polarity flips, and exchanged between the two class rails under a label swap — then the canonical export yields a signed DNF R̂ with R̂(g·(X,Y)) logically equivalent to g·R̂(X,Y) for every symmetry g, exactly. The export compresses each atom into a presence score sj = max(p+j, p−j) and a signed contrast dj = p+j − p−j, admits atoms in whole tie-buckets in canonical order, omits an atom whose polarity ties rather than choosing a side, and selects between the positive rail's rule and the neg

Load-bearing premise

The exactness guarantee for the shipped model rests on the claim, taken from an unpublished description of the Neural Rule Inducer rather than verified from released code, that zeroing the literal-sign feature and disabling example-content keys leaves every per-literal statistic and the pair-memory selection free of any atom-indexed dependence — and on real data the guarantee covers only within-block permutations of one-hot blocks, which are trivial for 6 of the 19 datasets.

Editorial extensions

If this is right

  • Any inducer whose literal scores are G-equivariant inherits exact rule-level equivariance from the canonical export with no retraining and no added parameters, since the guarantee lives in the decoder, not the model.
  • A single checkpoint trained on 6–12 variables keeps support-set accuracy near 0.89–0.93 at 1,024 variables (85× its training range), while the unmodified model degrades to chance (0.487).
  • On 19 real datasets the zero-shot accuracy gain over the baseline grows with schema size (≈0 inside the training range, +5.7 percentage points above 100 atoms), and exported rules are far sparser than per-dataset trained CART and RIPPER rules (2.04 vs 8.01/4.56 mean clauses).
  • The construction costs 1.2–1.8× the single-pass inference time with no measurable peak-memory change, and an eval-time-only variant without symmetrised training reproduces most of the accuracy gain.

Reading between the lines

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

  • The paper's sparse-support argument predicts that the decoder's error tracks the number and size of clauses in the target rule, not the atom count; a direct test holding N fixed while varying kmax and ℓmax — which the current stress test deliberately does not do — would settle that prediction.
  • Because Theorem 2 constrains only the score interface, attaching the canonical export to any other G-equivariant literal scorer (for instance a transformer-based in-context learner) is a natural test the paper explicitly leaves open; the export itself should need no changes.
  • For datasets whose features carry semantic structure beyond one-hot blocks — such as MUTAG's 51 chemical indicators — the relevant symmetry group must come from the domain itself; learning or specifying such groups automatically would extend the exactness guarantee beyond categorical schemas.
  • The export's abstentions (exact polarity ties, rail ties, and empty rules) double as a free confidence signal: episodes that produce abstentions or empty rules could be flagged for human inspection rather than trusted, an operational use the paper does not discuss.
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 proposes a canonical, tie-complete decoder that turns per-literal inclusion scores (p+, p−) into a signed DNF rule, and proves (Theorem 2) that if the score map obeys the symmetry group G = SM × BN × Zy2, then the exported rule is exactly G-equivariant up to logical equivalence, with abstentions at polarity and rail ties. The authors instantiate this on a proposed Neural Rule Inducer (NRI), claiming to restore BN equivariance via architectural fixes (zeroing the literal-sign feature, disabling content keys), eval-time tie-complete pair selection and polarity averaging, a label-swap TTA, and symmetrized training. Experiments on synthetic schemas up to N=1024 show G-NRI holding support-label accuracy near 0.89–0.93 while the baseline degrades to ~0.49, and rule fidelity above baseline. On 19 real datasets, G-NRI improves over the baseline among zero-shot variants, especially on larger schemas. The paper is careful about paired seeds, pre-registered statistical families, rule-equality metrics excluding empty exports, and separates the conditional theorem from its empirical instantiation.

Significance. If the conditional theorem is taken as the core contribution, this is a clean and useful result: lifting equivariance from continuous scores to a discrete symbolic output by construction, without retraining, is a transferable idea that could benefit other neuro-symbolic inducers. The proof-by-construction in Appendix B is credible, and the authors are appropriately explicit that the guarantee is about commutation, not semantic correctness. The empirical protocol is stronger than typical for this area: paired seeds, a pre-registered analysis plan, careful treatment of the real-data unit of analysis, and explicit separation of empty-rule abstentions. The scaling experiments are also informative. The main weakness is that the applicability of Theorem 2 to the actual G-NRI model rests on unverified architecture assertions about an unpublished NRI, and the real-data 'exactness' is vacuous for a substantial subset of datasets.

major comments (3)
  1. [§3.1, Prop. 1; Appendix A; Reproducibility Statement] The central claim that G-NRI satisfies the premise of Theorem 2 rests on unverified assertions about the NRI architecture: that per-literal encoder statistics are computed by an index-free, polarity-symmetric function once the sign feature is zeroed and content keys are disabled, and that no atom-indexed parameter remains. The NRI is an unpublished citation (Phua 2026), and the released repository has no commit hash or checkpoints. The measured worst-case drift (~0.001) also does not by itself establish exactness; the margin-stability argument requires the drift to be below the smallest decode margin, but no margin values are reported. I request a versioned code snapshot with a runnable equivariance audit, or an explicit re-scoping of the empirical 'exact' claims to 'within measured drift'.
  2. [§5.1, Table 10; Abstract] The abstract and Section 5.1 claim exact rule equivariance on 'schema-valid real-data tests', but Table 10 shows that 6 of 19 datasets (breast-cancer-wisconsin, clevr-hans3, diabetes, ionosphere, mutag, spambase) have HΣ = {e}, so their schema rule eq of 1.000 is vacuous. In addition, the one-vs-rest positive-class-2 slice of nursery produces all-empty rules for two seeds, making the equivariance there vacuous. The claim should be qualified to datasets with a nontrivial schema subgroup and non-abstaining exports.
  3. [§3.5 and §5.2] The 'sparse-support stability' argument is stated as an explanation for why the decoder error should track the number of literals used rather than N, but no formal statement or proof is provided. If this is intended as a theoretical justification for the scaling result, it needs a precise formulation; otherwise, the scaling claim should be presented purely empirically, which would be acceptable given the data.
minor comments (4)
  1. [Title] The title contains a typo: 'Symmetry-A ware' should be 'Symmetry-Aware'.
  2. [§4, Table 6] The paper says 'twelve schema sizes from N=6 to N=1024', but the exact values are only in Table 6; consider listing them in the text for clarity.
  3. [§5.2, Table 3b] The claim 'no measurable change in peak memory' is not backed by a table or measurement description; either add the measurement or soften the statement.
  4. [§4, Appendix C] The pre-registration is described but no timestamp or public archive is given; a URL or DOI would strengthen the claim.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: Theorem 2 is a conditional equivariance proof that does not use its conclusion; the self-cited NRI is disclosed and is not the support for the theorem.

full rationale

The central derivation is Theorem 2 (Appendix B): assuming G-equivariant literal scores, it checks each generator of S_M, S_N, Z_2^N, Z_2^y against the canonical export's deterministic decode. The proof does not invoke the target equivariance; it derives it from the definitions of s_j=max(p+,p-), d_j=p+-p-, tie-complete selection, and dual-rail selection. No fitted parameter is renamed as a prediction: the paper freezes one checkpoint, adds no parameters, and compares against the same-recipe baseline. The self-citation to the author's NRI (Phua, 2026) is heavy but disclosed; Proposition 1's architectural premise (no atom-indexed parameter after fixes) is a claim about the model that is accompanied by released code and by direct drift measurements, not by the theorem. The vacuous H_Sigma={e} cases on 6 real datasets and the NRI-only validation are explicitly flagged in Appendix F.5 and Section 7; these are external-validity limitations, not circular reductions. No equation in the paper reduces to its own input by construction.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

No invented physical entities. The central theorem is conditional; the free parameters are decoder constants and inherited NRI hyperparameters. The empirical weight is carried by the NRI architecture and training, which are taken from prior work by the same author. The sparse-support stability claim is the least supported entry.

free parameters (4)
  • presence threshold 1/2 = 0.5
    Chosen constant in Algorithm 1 canonical decode; an atom enters a clause when s_j >= 1/2 and d_j != 0. Not fitted to data but set by hand.
  • budget b (max literals per clause) = not stated
    Algorithm 1 takes a budget b for literals per clause; the paper does not report its value or whether it was tuned.
  • tie tolerance tie eps = not stated
    Algorithm 1 groups tied presence scores within tie eps; value not reported, chosen by hand.
  • pair memory top-k k = 8
    NRI native forward uses k=8 (Appendix A); tie-complete selection replaces it, but the cutoff is inherited from the NRI and affects which pairs are considered.
assumptions (6)
  • domain assumption The NRI per-literal encoder is index-free and polarity-symmetric after zeroing the sign feature and disabling content keys.
    Proposition 1 proof (Appendix A) relies on this claim about the architecture; it is asserted from the unpublished NRI description, not verified from shipped code.
  • domain assumption The literal statistics used by the encoder (truth-rate, marginal, entropy, co-occurrence) are symmetric under swapping a literal with its negation.
    Appendix A asserts this so the shared MLP and constant-zero sign feature commute with polarity flips.
  • domain assumption Training drives each atom and its negation to exactly equal pair scores, so self-pairs tie and the tie-complete selection admits exactly the boundary bucket.
    Appendix A states that training drives self-pairs to exactly equal scores; this is an empirical property of the training dynamics, not proven.
  • domain assumption Real-data schema-valid symmetries are exactly the one-hot block subgroups H_Sigma = product of Sym(B_b) with no polarity flips; raw BN is off-manifold.
    Appendix F defines H_Sigma and uses it to explain raw BN failures as score-equivariance violations under schema-invalid transforms. This limits the real-data guarantee to within-block permutations.
  • domain assumption The score-map drift stays below the decode margin on all audited inputs in floating point.
    The margin-stability argument in Appendix B converts approximate score equivariance to exact discrete equivariance; the margin is not computed directly, only inferred from rule eq = 1.000.
  • ad hoc to paper Sparse-support stability: decoder error tracks the number of literals the rule uses, not N.
    Section 3.5 gives this as an unproved argument for why scaling should work; it is not established, and the stress test keeps rule complexity fixed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pretrain on Small Synthetic Data, Scale Large for Free: Symmetry-Aware Foundation Model for Logic Rule Induction." pith.science (2026). https://pith.science/paper/76LMYLB7

@misc{pith2026260800383,
  author       = {Pith},
  title        = {Pith review of: Pretrain on Small Synthetic Data, Scale Large for Free: Symmetry-Aware Foundation Model for Logic Rule Induction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/76LMYLB7}},
  note         = {Machine review of arXiv:2608.00383}
}
read the original abstract

Logical rule induction seeks interpretable rules that transfer across propositional schemas. This requires respecting symmetries: atom naming, example order, polarity flips, and label swap. Enforcing exact symmetry by construction lets one trained inducer scale beyond its training schemas. Our central contribution is a canonical export that decodes a discrete rule from literal scores. It needs no retraining and is exactly equivariant whenever those scores respect the symmetries. We instantiate it on the Neural Rule Inducer, a disjunctive-normal-form (DNF) foundation model that natively respects only example order. We restore the remaining symmetries through architecture, inference, and training. On synthetic stress tests, accuracy on the support labels stays stable at much larger schemas, and rule fidelity on fresh inputs remains above the unmodified model. On real data, accuracy improves most on larger schemas. The exported rule is exact on synthetic full-group tests and on schema-valid real-data tests. This is a mathematical property of the export rather than of a specific model, and we validate it empirically only on the NRI. Enforcing symmetry by construction turns this small-data pretrained model into a reusable, interpretable inducer that transfers to larger schemas.

Figures

Figures reproduced from arXiv: 2608.00383 by the authors.

Figure 1
Figure 1. G-NRI, our symmetry-restored NRI. The export commutes with every example, atom, polarity, and label transform in G, defined below. per-example component makes its output depend on atom ordering again. So beyond the training range its accuracy drops and its exported rules stop transforming correctly under label swaps. Enforcing these symmetries by construction removes such shortcuts rather than asking training to app… view at source ↗
Figure 2
Figure 2. Variable-schema scaling and drift stratification [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

74 extracted references · 69 canonical work pages

  1. [1]

    A Foundation Model for Zero-Shot Logical Rule Induction

    Phua, Yin Jun , title =. Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI) , year =. 2605.04916 , archivePrefix =

    work page Pith review arXiv

  2. [2]

    New Generation Computing , volume =

    Muggleton, Stephen , title =. New Generation Computing , volume =. 1991 , doi =

    work page 1991

  3. [3]

    Inductive Logic Programming: Theory and Methods , journal =

    Muggleton, Stephen and. Inductive Logic Programming: Theory and Methods , journal =. 1994 , doi =

    work page 1994

  4. [4]

    Ross , title =

    Quinlan, J. Ross , title =. Machine Learning , volume =. 1990 , doi =

    work page 1990

  5. [5]

    Machine Learning , volume =

    Cropper, Andrew and Morel, Rolf , title =. Machine Learning , volume =. 2021 , doi =

    work page 2021

  6. [6]

    Machine Learning , volume =

    Inoue, Katsumi and Ribeiro, Tony and Sakama, Chiaki , title =. Machine Learning , volume =. 2014 , doi =

    work page 2014

  7. [7]

    Journal of Artificial Intelligence Research , volume =

    Evans, Richard and Grefenstette, Edward , title =. Journal of Artificial Intelligence Research , volume =. 2018 , doi =

    work page 2018

  8. [8]

    , title =

    Yang, Fan and Yang, Zhilin and Cohen, William W. , title =. Advances in Neural Information Processing Systems (NeurIPS) , volume =

    work page

Show all 74 references
  1. [9]

    End-to-End Differentiable Proving , booktitle =

    Rockt. End-to-End Differentiable Proving , booktitle =

  2. [10]

    Artificial Intelligence Review , volume =

    Neurosymbolic. Artificial Intelligence Review , volume =. 2023 , doi =

    2023

  3. [11]

    2020 , doi =

    From Statistical Relational to Neuro-Symbolic Artificial Intelligence , booktitle =. 2020 , doi =

    2020

  4. [12]

    Shortcut Learning in Deep Neural Networks , journal =

    Geirhos, Robert and Jacobsen, J. Shortcut Learning in Deep Neural Networks , journal =. 2020 , doi =

    2020

  5. [13]

    and Liang, Percy , title =

    Sagawa, Shiori and Koh, Pang Wei and Hashimoto, Tatsunori B. and Liang, Percy , title =. International Conference on Learning Representations (ICLR) , year =

  6. [14]

    Invariant Risk Minimization , year =

    Arjovsky, Martin and Bottou, L. Invariant Risk Minimization , year =. 1907.02893 , archivePrefix =

    1907 arXiv

  7. [15]

    Out-of-Distribution Generalization via Risk Extrapolation (

    Krueger, David and Caballero, Ethan and Jacobsen, J. Out-of-Distribution Generalization via Risk Extrapolation (. Proceedings of the 38th International Conference on Machine Learning , series =. 2021 , url =

    2021

  8. [16]

    International Conference on Learning Representations (ICLR) , year =

    Gulrajani, Ishaan and Lopez-Paz, David , title =. International Conference on Learning Representations (ICLR) , year =

  9. [17]

    2014 , note =

    Graves, Alex and Wayne, Greg and Danihelka, Ivo , title =. 2014 , note =. 1410.5401 , archivePrefix =

    2014 arXiv

  10. [18]

    Hybrid Computing using a Neural Network with Dynamic External Memory , journal =

    Graves, Alex and Wayne, Greg and Reynolds, Malcolm and Harley, Tim and Danihelka, Ivo and Grabska-Barwi. Hybrid Computing using a Neural Network with Dynamic External Memory , journal =. 2016 , doi =

    2016

  11. [19]

    Proceedings of The 33rd International Conference on Machine Learning , series =

    Santoro, Adam and Bartunov, Sergey and Botvinick, Matthew and Wierstra, Daan and Lillicrap, Timothy , title =. Proceedings of The 33rd International Conference on Machine Learning , series =. 2016 , url =

    2016

  12. [20]

    Deep Sets , booktitle =

    Zaheer, Manzil and Kottur, Satwik and Ravanbakhsh, Siamak and P. Deep Sets , booktitle =

  13. [21]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Henderson, Peter and Islam, Riashat and Bachman, Philip and Pineau, Joelle and Precup, Doina and Meger, David , title =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2018 , doi =

    2018

  14. [22]

    Proceedings of Machine Learning and Systems (MLSys) , volume =

    Bouthillier, Xavier and Delaunay, Pierre and Bronzi, Mirko and Trofimov, Assya and Nichyporuk, Brennan and Szeto, Justin and others , title =. Proceedings of Machine Learning and Systems (MLSys) , volume =. 2021 , url =

    2021

  15. [23]

    2021 , note =

    Picard, David , title =. 2021 , note =. 2109.08203 , archivePrefix =

    2021 arXiv

  16. [24]

    , title =

    Perez, Ethan and Strub, Florian and de Vries, Harm and Dumoulin, Vincent and Courville, Aaron C. , title =. Proceedings of the 32nd AAAI Conference on Artificial Intelligence , volume =. 2018 , doi =

    2018

  17. [25]

    and Kaiser,

    Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N. and Kaiser,. Attention Is All You Need , booktitle =

  18. [26]

    2023 , institution =

    Kelly, Markelle and Longjohn, Rachel and Nottingham, Kolby , title =. 2023 , institution =

    2023

  19. [27]

    Proceedings of The 33rd International Conference on Machine Learning , year =

    Group Equivariant Convolutional Networks , author =. Proceedings of The 33rd International Conference on Machine Learning , year =

  20. [28]

    7th International Conference on Learning Representations , year =

    Invariant and Equivariant Graph Networks , author =. 7th International Conference on Learning Representations , year =

  21. [29]

    2021 , note =

    Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges , author =. 2021 , note =. 2104.13478 , archivePrefix =

    2021 arXiv

  22. [30]

    Proceedings of the 35th International Conference on Machine Learning , year =

    Deep Models of Interactions Across Sets , author =. Proceedings of the 35th International Conference on Machine Learning , year =

  23. [31]

    Proceedings of the 36th International Conference on Machine Learning , year =

    Set Transformer: A Framework for Attention-based Permutation-Invariant Neural Networks , author =. Proceedings of the 36th International Conference on Machine Learning , year =

  24. [32]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

    Better Aggregation in Test-Time Augmentation , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

  25. [33]

    Neural-Symbolic Learning and Reasoning , year =

    Variable Assignment Invariant Neural Networks for Learning Logic Programs , author =. Neural-Symbolic Learning and Reasoning , year =

  26. [34]

    Scandinavian Journal of Statistics , volume =

    A Simple Sequentially Rejective Multiple Test Procedure , author =. Scandinavian Journal of Statistics , volume =. 1979 , publisher =

    1979

  27. [35]

    Journal of the American Statistical Association , volume =

    Robust Tests for the Equality of Variances , author =. Journal of the American Statistical Association , volume =. 1974 , publisher =

    1974

  28. [36]

    and Bala, Jerzy W

    Thrun, Sebastian B. and Bala, Jerzy W. and Bloedorn, Eric and Bratko, Ivan and Cestnik, Bojan and Cheng, John and De Jong, Kenneth A. and D. The. 1991 , month =. doi:10.13021/MARS/3370 , url =

    1991 doi

  29. [37]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Right for the Right Concept: Revising Neuro-Symbolic Concepts by Interacting with their Explanations , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =. 2021 , doi =

    2021

  30. [38]

    Correlation with Molecular Orbital Energies and Hydrophobicity , author =

    Structure-Activity Relationship of Mutagenic Aromatic and Heteroaromatic Nitro Compounds. Correlation with Molecular Orbital Energies and Hydrophobicity , author =. Journal of Medicinal Chemistry , volume =. 1991 , doi =

    1991

  31. [39]

    Artificial Intelligence , volume =

    Theories for Mutagenicity: A Study in First-Order and Feature-Based Induction , author =. Artificial Intelligence , volume =. 1996 , doi =

    1996

  32. [40]

    Advances in Neural Information Processing Systems (NeurIPS) , volume =

    Manhaeve, Robin and Duman. Advances in Neural Information Processing Systems (NeurIPS) , volume =

  33. [41]

    The Annals of Statistics , volume =

    Bootstrap Methods: Another Look at the Jackknife , author =. The Annals of Statistics , volume =. 1979 , publisher =

    1979

  34. [42]

    International Conference on Learning Representations (ICLR) , year =

    Neural Logic Machines , author =. International Conference on Learning Representations (ICLR) , year =

  35. [43]

    International Conference on Learning Representations (ICLR) , year =

    The Neuro-Symbolic Concept Learner: Interpreting Scenes, Words, and Sentences From Natural Supervision , author =. International Conference on Learning Representations (ICLR) , year =

  36. [44]

    Artificial Intelligence , volume =

    Logic Tensor Networks , author =. Artificial Intelligence , volume =. 2022 , doi =

    2022

  37. [45]

    and Valiant, Gregory , booktitle =

    Garg, Shivam and Tsipras, Dimitris and Liang, Percy S. and Valiant, Gregory , booktitle =. What Can Transformers Learn In-Context?

  38. [46]

    Nature , volume =

    Human-like systematic generalization through a meta-learning neural network , author =. Nature , volume =. 2023 , doi =

    2023

  39. [47]

    Proceedings of the 34th International Conference on Machine Learning , series =

    Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks , author =. Proceedings of the 34th International Conference on Machine Learning , series =. 2017 , volume =

    2017

  40. [48]

    Advances in Neural Information Processing Systems (NeurIPS) , volume =

    Learning Compositional Rules via Neural Program Synthesis , author =. Advances in Neural Information Processing Systems (NeurIPS) , volume =. 2020 , url =

    2020

  41. [49]

    Clark, Peter and Niblett, Tim , journal =. The. 1989 , doi =

    1989

  42. [50]

    Machine Learning Proceedings 1995 , pages =

    Fast Effective Rule Induction , author =. Machine Learning Proceedings 1995 , pages =. 1995 , publisher =. doi:10.1016/B978-1-55860-377-6.50023-2 , isbn =

    1995 doi

  43. [51]

    Classification and Regression Trees , author =

  44. [52]

    and Madigan, David , journal =

    Letham, Benjamin and Rudin, Cynthia and McCormick, Tyler H. and Madigan, David , journal =. Interpretable Classifiers Using Rules and. 2015 , doi =

    2015

  45. [53]

    2020 , eprint =

    On the Benefits of Invariance in Neural Networks , author =. 2020 , eprint =

    2020

  46. [54]

    Journal of Machine Learning Research , volume =

    A Group-Theoretic Framework for Data Augmentation , author =. Journal of Machine Learning Research , volume =

  47. [55]

    Advances in Neural Information Processing Systems (NeurIPS) , volume =

    Learning Invariances in Neural Networks from Training Data , author =. Advances in Neural Information Processing Systems (NeurIPS) , volume =. 2020 , url =

    2020

  48. [56]

    Proceedings of the 34th International Conference on Machine Learning , series =

    Equivariance Through Parameter-Sharing , author =. Proceedings of the 34th International Conference on Machine Learning , series =. 2017 , volume =

    2017

  49. [57]

    and Le Lan, Charline and Zaidi, Sheheryar and Dupont, Emilien and Teh, Yee Whye and Kim, Hyunjik , booktitle =

    Hutchinson, Michael J. and Le Lan, Charline and Zaidi, Sheheryar and Dupont, Emilien and Teh, Yee Whye and Kim, Hyunjik , booktitle =. 2021 , volume =

    2021

  50. [58]

    Densely Connected

    Agrawal, Devanshu and Ostrowski, James , journal =. Densely Connected

  51. [59]

    Artificial Intelligence , volume =

    Analyzing Differentiable Fuzzy Logic Operators , author =. Artificial Intelligence , volume =. 2022 , doi =

    2022

  52. [60]

    Proceedings of the 35th International Conference on Machine Learning , series =

    Generalization without Systematicity: On the Compositional Skills of Sequence-to-Sequence Recurrent Networks , author =. Proceedings of the 35th International Conference on Machine Learning , series =. 2018 , volume =

    2018

  53. [61]

    2020 , doi =

    Kim, Najoung and Linzen, Tal , booktitle =. 2020 , doi =

    2020

  54. [62]

    Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , address =

    Harnessing Deep Neural Networks with Logic Rules , author =. Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , address =. 2016 , doi =

    2016

  55. [63]

    Proceedings of the 35th International Conference on Machine Learning , series =

    A Semantic Loss Function for Deep Learning with Symbolic Knowledge , author =. Proceedings of the 35th International Conference on Machine Learning , series =. 2018 , volume =

    2018

  56. [64]

    Equivariance with Learned Canonicalization Functions , booktitle =

    Kaba, S. Equivariance with Learned Canonicalization Functions , booktitle =

  57. [65]

    and Misra, Ishan and Grover, Aditya and Ben-Hamu, Heli and Lipman, Yaron , title =

    Puny, Omri and Atzmon, Matan and Smith, Edward J. and Misra, Ishan and Grover, Aditya and Ben-Hamu, Heli and Lipman, Yaron , title =. International Conference on Learning Representations (ICLR) , year =

  58. [66]

    , title =

    Golomb, Solomon W. , title =. IRE Transactions on Information Theory , volume =. 1959 , publisher =

    1959

  59. [67]

    International Conference on Field-Programmable Technology (FPT) , pages =

    Huang, Zheng and Wang, Lingli and Nasikovskiy, Yakov and Mishchenko, Alan , title =. International Conference on Field-Programmable Technology (FPT) , pages =. 2013 , organization =

    2013

  60. [68]

    International Conference on Machine Learning (ICML) , series =

    Maron, Haggai and Litany, Or and Chechik, Gal and Fetaya, Ethan , title =. International Conference on Machine Learning (ICML) , series =. 2020 , publisher =

    2020

  61. [69]

    and Chechik, Gal and Maron, Haggai , title =

    Yehudai, Gilad and Fetaya, Ethan and Meirom, Eli A. and Chechik, Gal and Maron, Haggai , title =. International Conference on Machine Learning (ICML) , series =. 2021 , publisher =

    2021

  62. [70]

    , title =

    Gordon, Jonathan and Bronskill, John and Bauer, Matthias and Nowozin, Sebastian and Turner, Richard E. , title =. International Conference on Learning Representations (ICLR) , year =

  63. [71]

    and Sra, Suvrit and Maron, Haggai and Jegelka, Stefanie , title =

    Lim, Derek and Robinson, Joshua David and Zhao, Lingxiao and Smidt, Tess E. and Sra, Suvrit and Maron, Haggai and Jegelka, Stefanie , title =. International Conference on Learning Representations (ICLR) , year =

  64. [72]

    End-to-end Symbolic Regression with Transformers , booktitle =

    Kamienny, Pierre-Alexandre and d'Ascoli, St. End-to-end Symbolic Regression with Transformers , booktitle =. 2022 , url =

    2022

  65. [73]

    Logics in Artificial Intelligence (JELIA) , year =

    Law, Mark and Russo, Alessandra and Broda, Krysia , title =. Logics in Artificial Intelligence (JELIA) , year =

  66. [74]

    Machine Learning , volume =

    Cunnington, Daniel and Law, Mark and Lobo, Jorge and Russo, Alessandra , title =. Machine Learning , volume =. 2023 , pages =

    2023

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.