Pith. sign in

REVIEW 2 major objections 5 minor 1 cited by

Neural Logic Networks for Interpretable Classification

T0 review · 2 major / 5 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 generalized AND/OR network with NOT and unobserved-data biases learns readable IF-THEN rules that exactly match ground truth on tic-tac-toe and four Boolean networks.

desk verdict A useful generalization of NLNs with a strong Boolean-network demo, but the exact-recovery claim outruns the reported evidence. read the letter →

arxiv 2508.08172 v6 pith:SDKUVNBZ submitted 2025-08-11 cs.LG cs.AIcs.LO

Vincent Perreault , Katsumi Inoue , Richard Labib , Alain Hertz This is my paper · ORCID
classification cs.LGcs.AIcs.LO
keywords NeuralLogicNetworksinterpretableclassificationBooleannetworkdiscoveryrulelearningAND/ORneuro-symbolicfuzzyprobabilisticcircuits
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 builds a probabilistic and logical foundation for Neural Logic Networks—networks whose neurons compute AND and OR over concepts—by adding negative weights for NOT, biases that absorb unobserved relevant data, and a factorized IF-THEN rule structure. It claims that this generalized network, trained with a rule-reset scheme and simplified by discretization, pruning, and bias adjustment, can recover exact ground-truth logic programs: with 40% of the data it perfectly predicts four Boolean networks and identifies their rules, and on tic-tac-toe it discovers the exact 8-rule program for 'X wins.' The same pipeline yields compact, readable rules on tabular classification tasks, including a 4-rule sub-model that perfectly classifies a chronic-kidney-disease dataset and a 24-rule model for network-intrusion detection. The practical payoff, if these claims hold, is a classifier that is both accurate and genuinely inspectable—extract the logic, verify it, and reuse it across tasks.

What carries the argument

The engine is the probabilistic AND/OR neuron with signed weights: a positive weight $A^l_{i,j}$ is the probability that concept $j$ in layer $l-1$ is necessary to concept $i$, a negative weight is the probability that its negation is necessary, and the bias $a^l_i \in [0,1]$ is the probability that all unobserved necessary concepts are present. The forward equations are derived as a product-probability approximation to the logical definitions, and the paper builds around this neuron a two-layer DNF: an AND layer with negation forming rule modules, an OR layer combining rules per output, plus preprocessing modules (one-hot OR equivalency classes for categoricals, fuzzy-dichotomy intervals fo

What would settle it

Take a small NLN with two AND rules that share an input feature (so the third assumption fails) and continuous inputs; compute the exact joint probability by enumerating all states and compare with the product-form approximation from (P-AND)/(P-OR). Any gap on such a network shows that the approximation changes predictions; if instead the gap is zero even for shared ancestors, the assumption's violation is harmless in exactly the tested regime.

Watch

Extended reading notes

Core claim

The central claim is that a two-layer network of probabilistic AND/OR concepts—with a rule module per AND, shared input encodings, integer-discretized weights, and biases representing unobserved necessary/sufficient concepts—learns a logic program whose rules are the actual ground-truth rules when one exists. The authors derive the forward equations (P-AND) and (P-OR) from set-theoretic definitions of necessary and sufficient concepts under three independence assumptions, show they coincide exactly with classical AND/OR logic when inputs are binary and weights are in {−1,0,1}, and demonstrate exact recovery on mammalian, fission yeast, budding yeast, and arabidopsis Boolean networks from 40%

Load-bearing premise

The load-bearing premise is that concepts in the same layer are conditionally independent given the input; the authors state this is false whenever two concepts share a common necessary/sufficient concept, and the tractable (P-AND)/(P-OR) equations—and hence the whole learning pipeline—rest on it.

Editorial extensions

If this is right

  • With only 40% of the full state-transition data, the trained NLN attains 100% accuracy on all four Boolean networks considered and identifies all ground-truth rules; at 16% it already exceeds 97% accuracy.
  • On tic-tac-toe the NLN produces exactly the 8 winning-line rules of size 3, while the RRL baseline, at equal accuracy, requires hundreds of rules and recovers none of the ground truth.
  • For tabular problems without known logic programs, NLN rule sets are small enough to read (averaging 4–30 rules of size 1–22 across benchmarks), enabling direct inspection and expert validation.
  • Learned rules can be merged across cross-validation folds and transferred to new label structures: the kidney model collapses 34 rules to 11 (4 suffice for perfect classification), and binary NSL-KDD rules transfer to a 5-class intrusion-detection task with 17 rules after re-pruning and bias adjustment.
  • When inputs are binary/categorical with integer weights, the probabilistic forward pass coincides with the logical semantics, so extracted rules are exact rather than approximate.

Reading between the lines

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

  • The paper's own Section 2.1 concedes that the third independence assumption is false in most non-trivial networks; I would therefore expect the exact-recovery guarantees to hold only in the binary/integer-weight regime where the formulas coincide with logic. For continuous inputs the forward probabilities are approximations, so the post-hoc bias adjustment is a heuristic—a direct test would compar
  • The rule-reset mechanism is effectively a random-restart search over the rule space; coupling it with a SAT-based verifier after discretization could certify whether the learned rule set is logically equivalent to (or implied by) the training data, turning high predictive accuracy into a guarantee.
  • The successful transfer of binary NSL-KDD rules to the multiclass setting suggests a general template: learn an interpretable coarse model, then specialize it by re-pruning and re-adjusting biases, instead of training a multiclass network from scratch—an approach that could extend to streaming or federated settings where re-training is costly.
  • The kidney 4-rule model classifies the 400-sample dataset perfectly, but the authors themselves caution that this is not proof of clinical truth; the natural next step is an independent cohort study checking whether the discovered feature thresholds (e.g., hemoglobin < 14.9, packed cell volume < 48.3) replicate.
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

2 major / 5 minor

Summary. The paper generalizes Neural Logic Networks (NLNs) for interpretable tabular classification. It introduces a probabilistic and logical formalism for AND/OR nodes with signed weights (positive for a concept, negative for its negation) and biases that account for unobserved data. It then proposes a factorized rule-module architecture, fuzzy-dichotomy pre-processing for continuous features, a rule-reset training procedure, and a post-processing pipeline (weight discretization, pruning, coverage analysis, bias adjustment). Experiments cover Boolean network discovery on four gene-regulatory benchmarks and tabular classification on eight UCI-style data sets, with additional medical and industrial case studies. The headline claims are that the method recovers exact ground-truth Boolean rules from as little as 40% of the data and that it 'exactly discovered the ground truth' on tic-tac-toe, while producing far sparser rule sets than the RRL baseline.

Significance. If the claims hold, this is a useful contribution to interpretable neuro-symbolic learning: it formalizes and extends earlier NLN formulations, provides an openly available implementation, and demonstrates that a constrained AND/OR architecture can yield small, human-readable rule sets on several benchmarks. The theoretical derivation in Section 2 and Appendix B is detailed and the authors are unusually candid about the limitations of the independence assumptions. However, the two strongest claims—rigorous probabilistic grounding and exact ground-truth rule recovery—are substantially weaker than stated: the tractable formulas rest on an approximation the authors themselves call 'catastrophic' outside a narrow binary/integer case, and the exact-recovery results are reported only as mean accuracy over two cross-validation repeats, without any seed-level rule-recovery statistics. These issues are load-bearing for the paper's central contributions, though they are fixable with additional experiments and a more careful framing.

major comments (2)
  1. [Section 2.1, Eqs. (P-AND)/(P-OR)] The derivation of (P-AND) and (P-OR) requires three independence assumptions. The third, conditional independence between concepts in the same layer given the input, is stated in the paper to be false in most practical cases and 'catastrophic' for NLNs as a purely probabilistic model. The formulas coincide with the logical semantics only when all inputs are binary, all weights are in {-1,0,1}, and biases are full (a=1, o=0). For continuous features and non-binary biases—the general setting used in the tabular experiments—the forward probabilities are therefore an approximation, not the exact logical probability. This affects not only classification outputs but also the post-hoc coverage analysis and bias adjustment in Section 3.2.2, which are presented as statistical estimates of rule probabilities. I ask the authors to state this limitation in the abstract and introduction, and to eithe
  2. [Section 4.1, Table D.1, and Section 4.2, Table 5] The claim that the method 'correctly identifies the rules of the ground-truth boolean networks' at 40% data, and that it 'has exactly discovered the ground truth' on tic-tac-toe, is not supported by the reported metrics. Table D.1 reports only mean prediction accuracy over two repeats of five-fold cross-validation; it reports no standard deviations, no rule-recovery frequency, and no measure of logical equivalence between the learned rule set and the ground truth. Perfect prediction accuracy does not imply rule-set equivalence: the authors' own examples in Appendix D.1.1 show learned rules that are subsumed by other rules or are over-general (e.g., the budding yeast case where one rule replaces two ground-truth rules) while achieving 100% accuracy. Moreover, Section 4.2 explicitly acknowledges that the stochastic search yields 'differing sets of rules' across runs. To establish exact rec
minor comments (5)
  1. [Table 3] The F1-score table is difficult to read because entries are not visually separated (e.g., '10099.31' should be '100, 99.31'). Please add spacing or LaTeX column separators.
  2. [Appendix B.1.1, Figure B.1] The PGM diagram contains corrupted characters (e.g., '⪅⌢⌞⋌ⓈⓈ⋋≫®⋔1') that make it unreadable. This figure is important for the probabilistic-modeling section and should be regenerated with proper math rendering.
  3. [Section 1, first paragraph] Typo: 'medecine' should be 'medicine'.
  4. [Code Availability] The repository URL is broken across a line break ('NeuralLogic Networks'). Also, consider adding a license and version tag for reproducibility.
  5. [Section 4.2.1] The merged kidney model with 'perfect classification' is selected after inspecting the full data set and then pruned on the same full data set. The authors do caution that perfect classification should not be over-interpreted, but the phrasing 'perfect classification on the kidney data set' in Figure 6 could be misread as a general property. Suggest explicitly labeling this as a post-hoc demonstration on the training set.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: NLN equations are derived from stated assumptions, and rule-recovery claims are validated against external ground-truth logic programs.

full rationale

The paper's central derivation (P-AND)/(P-OR) is obtained from explicit modeling assumptions in Sections 2.1 and B.1.2, not from the target labels or from the claims being predicted. The weights and biases are learned parameters; the bias-adjustment step in Section 3.2.2 estimates a_i and o_k as empirical proportions of coverage, which is parameter fitting rather than a 'prediction' of the same quantity. Boolean-network rule recovery (Section 4.1 and Appendix D.1) is compared directly against known ground-truth logic programs (mammalian, fission, budding, arabidopsis), and tic-tac-toe recovery (Section 4.2, Table 5) is compared against the known 8 winning lines; these are external benchmarks, not the model's own fitted values. The only self-citations (LF1T, D-LFIT, Sato & Inoue 2023) are used as baselines or related work and are not load-bearing for the derivation. The paper explicitly acknowledges the falsehood of the third independence assumption and the stochastic variability of the search; these are correctness/robustness limitations, not circular reductions. No equation reduces to its inputs by construction, and no fitted parameter is renamed as an independent prediction.

Assumptions & free parameters 7 free parameters · 7 assumptions · 1 invented entities

The learning pipeline uses standard gradient descent over the model's weights; the more problematic fixed choices are the lambda regularizers, 128 rule modules, 32 fuzzy dichotomies, and the initial/final bias settings. The theoretical derivation rests on three independence assumptions (one admittedly false) plus a 'cognitive bias' assumption. Unobserved concepts are invented latent constructs with no independent empirical handle.

free parameters (7)
  • lambda_non_empty = 1e-1
    Regularization coefficient forcing non-empty definitions; chosen by hand for all experiments (Section 4).
  • lambda_sparsity = 1e-3
    L1 sparsity regularization coefficient; chosen by hand (Section 4).
  • num_rule_modules = 128
    Number of AND rule modules in the hidden layer for all experiments (Section 4).
  • num_fuzzy_dichotomies_per_feature = 32
    Default number of fuzzy dichotomies per continuous feature, with 33 interval encodings (Section 4 and Section 3.1.2).
  • initial_AND_bias a_i = 1
    Initial bias representing fully observed concepts; set by hand to maximize initial gradients.
  • initial_OR_bias o_j = 0
    Initial bias representing no unobserved sufficient concepts; set by hand.
  • posthoc_bias_adjustments = empirical proportions
    Rule biases a_i and output biases o_k are set to empirical coverage proportions over the training set (Section 3.2.2), i.e., fitted to data after training.
assumptions (7)
  • domain assumption First independence assumption: presence of a concept in a realization is independent of its general role (necessary/sufficient) in the next layer.
    Invoked in Appendix B.1.2 to factor terms; stated in Section 2.1.
  • domain assumption Second independence assumption: the necessary/sufficient roles of different concepts for the same node are mutually independent.
    Stated in Section 2.1; described as a cognitive bias of 'total open-mindedness'.
  • domain assumption Third independence assumption: concepts in the same layer are conditionally independent given the input.
    Section 2.1; the authors state it is false in most cases and only an approximation.
  • domain assumption At most one of P[C_j contains Y] and P[(C_j)^c contains Y] is nonzero for each weight.
    Section 2.1 and Section 2.4; used to encode positive/negative weight with a single parameter.
  • standard math DNF completeness: every Boolean function can be represented as a two-layer AND-OR network.
    Standard result from classical logic, used to justify the network structure in Section 3.1.1.
  • standard math Product t-norm/t-conorm semantics coincide with probability under independence.
    Used in Section 2.2 to rewrite the probabilistic equations as fuzzy logic operators.
  • domain assumption Boolean network ground truths in the experiments are correct.
    Used in Section 4.1 and Appendix D.1; the ground-truth logic programs are taken from published biology papers.
invented entities (1)
  • Unobserved relevant data u and unobserved concepts ~N_i^l, ~S_i^l
    purpose: Represent missing causes/effects that justify the bias parameters a_i and o_i in the probabilistic model.
    Introduced in Section 2.1 as latent quantities; no falsifiable handle is given outside the fitted biases. The biases are set empirically to training-set proportions in Section 3.2.2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Neural Logic Networks for Interpretable Classification." pith.science (2026). https://pith.science/paper/SDKUVNBZ

@misc{pith2026250808172,
  author       = {Pith},
  title        = {Pith review of: Neural Logic Networks for Interpretable Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SDKUVNBZ}},
  note         = {Machine review of arXiv:2508.08172}
}
read the original abstract

Traditional neural networks have an impressive classification performance, but what they learn cannot be inspected, verified or extracted. Neural Logic Networks on the other hand have an interpretable structure that enables them to learn a logical mechanism relating the inputs and outputs with AND and OR operations. We generalize these networks with NOT operations and biases that take into account unobserved data and develop a rigorous logical and probabilistic modeling in terms of concept combinations to motivate their use. We also propose a novel factorized IF-THEN rule structure for the model as well as a modified learning algorithm. Our method improves the state-of-the-art in Boolean networks discovery and is able to learn relevant, interpretable rules in tabular classification, notably on examples from the medical and industrial fields where interpretability has tangible value.

Figures

Figures reproduced from arXiv: 2508.08172 by the authors.

Figure 1
Figure 1. Interpretability of the learned AND/OR networks on tic-tac-toe, when trying to [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Toy examples of interpretations for AND and OR concepts [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. It contains two fully-connected layers arranged in DNF, i.e. an AND layer with negation followed by an OR layer without negation, in order to learn a logic program for each target (section 3.1.1). The input features that are not binary are pre-processed with appropriate input modules, one for categorical features, and another for continuous features (section 3.1.2). 3.1.1 Fully-Connected DNF Layers NLNs, as we have … view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: NLN structure used for learning as in the clustering of a binary classification problem (Ruspini, 1970) or in measuring the “dichotomousness” of a fuzzy set (Kitainik, 1987). 3.1.3 Input Encodings and Rule Modules To help the learning process, we do not learn the NLN d…
Figure 5
Figure 5. Figure 5: NLN initialization our chances to find potential rules that can be further massaged towards relevant rules, with respect to the target concepts. We begin with full binary biases a 1 i = 1 and o 2 j = 0, i.e. without unobserved effects. These ensure that the initial gra…
Figure 6
Figure 6. Figure 6: Merged NLN’s 4-rule sub-model with perfect classification on the kidney data set [PITH_FULL_IMAGE:figures/full_fig_p026_6.png]
Figure 7
Figure 7. Figure 7: Merged RRL’s 13-rule sub-model with perfect classification on the kidney data set [PITH_FULL_IMAGE:figures/full_fig_p027_7.png]
Figure 8
Figure 8. Figure 8: Breakdown of rule coverage by class on the full data set (training + test) [PITH_FULL_IMAGE:figures/full_fig_p029_8.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Training, Reading, and Editing Legible Transformers

    cs.LG 2026-07 conditional novelty 6.5 of 10

    A variance-floor objective plus learned operator gates produce an end-to-end legible transformer whose crisp units are 50–184× more local to edit and can be reshaped from fan-out to fan-in circuits without quality loss.

Reference graph

Works this paper leans on

4 extracted references · 3 canonical work pages · cited by 1 Pith paper

  1. [1980]

    Prithviraj Sen, Breno W

    doi: 10.1017/S0140525X00005756. Prithviraj Sen, Breno W. S. R. de Carvalho, Ryan Riegel, and Alexander Gray. Neuro- symbolic inductive logic programming with logical neural networks.Proceedings of the AAAI Conference on Artificial Intelligence, 36(8):8212–8219, Jun. 2022. Alen Shapiro. Chess (King-Rook vs. King-Pawn on a7). UCI Machine Learning Repository,

    doi:10.1017/s0140525x00005756 2022

  2. [1983]

    DOI: https://doi.org/10.24432/C5DK5C. R. Siegler. Balance Scale. UCI Machine Learning Repository, 1976. DOI: https://doi.org/10.24432/C5488X. Mahbod Tavallaee, Ebrahim Bagheri, Wei Lu, and Ali A. Ghorbani. A detailed analysis of the kdd cup 99 data set. In2009 IEEE Symposium on Computational Intelligence for Security and Defense Applications, pages 1–6, 2...

    work page arXiv 1976

  3. [2020]

    Incorporating Relational Background Knowledge into Reinforcement Learning via Differentiable Inductive Logic Programming

    URLhttps://arxiv.org/abs/2003.10386. Witold Pedrycz. Fuzzy neural networks and neurocomputations.Fuzzy Sets and Systems, 56 (1):1–28, 1993. Jonas Peters, Dominik Janzing, and Bernhard Schölkopf.Elements of Causal Inference - Foundations and Learning Algorithms. The MIT Press, 2017. Ryan Riegel, Alexander Gray, Francois Luus, Naweed Khan, Ndivhuwo Makondo,...

    work page Pith review arXiv 2003

  4. [2024]

    Curran Associates Inc. M.M. Gupta. Fuzzy logic, neural networks and virtual cognitive systems. In1993 (2nd) International Symposium on Uncertainty Modeling and Analysis, pages 90–97, 1993. M.M. Gupta and D.H. Rao. On the principles of fuzzy neural networks.Fuzzy Sets and Systems, 61(1):1–18, 1994. K. Hirota and W. Pedrycz. Or/and neuron in modeling fuzzy ...

    work page arXiv 1993

Pith tools

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