arxiv: 2605.01797 · v1 · submitted 2026-05-03 · 💻 cs.AI

Recognition: unknown

Neural Decision-Propagation for Answer Set Programming

Thomas Eiter , Katsumi Inoue , Sota Moriyama

Authors on Pith no claims yet

Pith reviewed 2026-05-10 15:08 UTC · model grok-4.3

classification 💻 cs.AI

keywords answer set programmingstable modelsneuro-symbolic AIdecision propagationneural networksdifferentiable reasoning

0 comments

The pith

Decision-propagation computes stable models by alternating falsity decisions and truth propagations, and its neural version learns to do so efficiently.

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

The paper introduces decision-propagation as a procedure for computing stable models in answer set programming that works by repeatedly making falsity decisions and then propagating truths through the program rules. It proves that any complete successful computation under this procedure exactly matches a stable model. The authors then define a differentiable neural variant that uses a neural network to make the decisions and fuzzy logic to perform the propagations. This allows the procedure to be trained end-to-end so that it learns good decision heuristics and integrates directly with other neural components. A reader would care because the approach removes reliance on external classical solvers, which have been the main scalability bottleneck in neuro-symbolic ASP systems.

Core claim

The paper shows that successful decision-propagation computations, which alternate falsity decisions and truth propagations, capture the stable model semantics of an answer set program. It then presents Neural DProp, a differentiable extension that replaces the decisions with neural computation and the propagations with fuzzy evaluation, and demonstrates that this version can be trained to compute stable models efficiently while improving accuracy and scalability on neuro-symbolic benchmarks.

What carries the argument

Decision-propagation (DProp): the procedure that alternates falsity decisions with truth propagations to derive stable models.

If this is right

NDProp learns decision heuristics that compute stable models efficiently.
The approach improves accuracy and scalability compared with prior neuro-symbolic ASP methods.
Reasoning pipelines can integrate ASP directly with neural networks without calling classical solvers.
End-to-end differentiability enables training on the combined neural-symbolic system.

Where Pith is reading between the lines

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

Training on small programs might generalize to larger instances if the learned heuristics capture structural patterns.
The same propagation structure could be adapted to other non-monotonic reasoning formalisms.
Fuzzy propagation opens a route to handling uncertain or probabilistic variants of answer set programs.

Load-bearing premise

Neural decisions combined with fuzzy propagations can approximate exact stable-model computation closely enough to preserve correctness and generalize beyond the training distribution.

What would settle it

A counter-example in which NDProp returns an interpretation that violates the stable-model condition on a program outside the training set, or fails to produce any stable model on programs larger than those used in the reported benchmarks.

Figures

Figures reproduced from arXiv: 2605.01797 by Katsumi Inoue, Sota Moriyama, Thomas Eiter.

**Figure 2.** Figure 2: NDProp pipeline. The neuro-symbolic integration is added for specific instantiations, and the arrow leading from the neuro [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: Example instances of 4x4 visual sudoku with different numbers of clues. [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗

read the original abstract

Integration of Answer Set Programming (ASP) with neural networks has emerged as a promising tool in Neuro-symbolic AI. While existing approaches extend the capabilities of ASP to real world domains, their reasoning pipelines depend on classical solvers, which is a bottleneck for scalability. To tackle this problem, we propose a new method to compute stable models, called decision-propagation (DProp), which alternates falsity decisions and truth propagations. Successful DProp computations are shown to capture the stable model semantics. We then develop Neural DProp (NDProp), a differentiable extension of DProp with neural computation for decisions and fuzzy evaluation for propagations. We evaluate the capabilities of NDProp for learning decision heuristics as well as neuro-symbolic integration, and compare it with existing neuro-symbolic approaches. The results show that NDProp can learn to efficiently compute stable models, and it improves accuracy and scalability on neuro-symbolic benchmarks.

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.

Desk Editor's Note private letter to a colleague

DProp gives a fresh propagation-based take on stable models and NDProp adds neural decisions plus fuzzy logic, but the neural version rests on empirical results without completeness or exact-semantics guarantees.

read the letter

The core of this paper is a decision-propagation procedure (DProp) that alternates falsity decisions with truth propagations, plus its neural extension (NDProp) that swaps in a neural network for the decisions and fuzzy logic for the propagations. The authors show that successful DProp runs match stable-model semantics and then demonstrate that NDProp can be trained to compute stable models on neuro-symbolic benchmarks while improving accuracy and speed over approaches that still invoke classical solvers. That is the actual novelty and the practical angle they emphasize. The work is clearly motivated by the scalability bottleneck in current neuro-symbolic ASP systems. What it does cleanly is separate the decision step from the propagation step so that learning can target the decisions directly. The benchmarks appear to support that the learned heuristics help on the reported tasks. The soft spots are exactly where the stress-test note flags them. DProp only claims to capture the semantics on successful runs, with no completeness argument that every program has such a run. NDProp replaces exact operations with approximations, so the output is no longer guaranteed to be a classical stable model; the paper validates this only empirically. There is no formal invariant tying the fuzzy fixed point back to the reduct, and generalization beyond the training distribution is not analyzed in depth. The experimental controls and variance details are also thin in the available description. This is aimed at researchers building hybrid reasoning systems who want differentiable alternatives to black-box solvers. It has enough new machinery and empirical signal to deserve peer review, though any referee will press for stronger theory on when NDProp stays correct.

Referee Report

3 major / 3 minor

Summary. The paper proposes Decision-Propagation (DProp), an iterative procedure that alternates falsity decisions with truth propagations to compute stable models of answer set programs, claiming that successful runs capture the stable-model semantics. It then introduces Neural DProp (NDProp), a differentiable variant that replaces decisions with a neural network and propagations with fuzzy logic, enabling learning of decision heuristics. The work evaluates NDProp on neuro-symbolic benchmarks, reporting gains in accuracy and scalability over prior neuro-symbolic ASP approaches.

Significance. If the central claims hold, the work offers a promising direction for scalable neuro-symbolic integration by avoiding reliance on classical ASP solvers. The DProp procedure provides an intuitive, non-backtracking view of stable-model computation, and the empirical results on learning heuristics and benchmark tasks indicate practical utility. Strengths include the explicit separation of decision and propagation steps and the end-to-end differentiability of NDProp, which could support further integration with gradient-based methods.

major comments (3)

[§3] §3: The manuscript establishes that successful DProp executions match stable-model semantics through alternating falsity decisions and truth propagations, yet provides no completeness argument showing that every consistent program admits at least one successful DProp path that recovers all stable models. Without this, the claim that DProp 'captures the stable model semantics' remains partial.
[§4.2] §4.2: NDProp substitutes classical propagations with fuzzy logic and decisions with a learned network to obtain differentiability. No formal invariant is supplied that relates the resulting fuzzy fixed point back to the classical reduct or guarantees that learned heuristics avoid non-successful branches, so correctness of NDProp rests entirely on the reported empirical results.
[§5] §5: The experimental section asserts that NDProp improves accuracy and scalability on neuro-symbolic benchmarks, but supplies neither error bars, details on random seeds or data splits, nor statistical significance tests for the reported gains. This absence makes it difficult to assess whether the observed improvements are robust or reproducible.

minor comments (3)

[§4] The definition of the fuzzy operators used in the propagation step (Eq. 7) would benefit from an explicit comparison table against the classical Boolean operators to clarify the approximation introduced.
[§5] Figure 3 (benchmark scalability plots) lacks y-axis scaling details and confidence intervals; adding these would improve interpretability of the runtime comparisons.
[§2] A short related-work subsection contrasting DProp with existing propagation-based ASP algorithms (e.g., unit propagation in CDCL solvers) would help situate the contribution.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We are grateful to the referee for the constructive comments, which help improve the clarity and rigor of our work. Below, we provide point-by-point responses to the major comments and outline the revisions we plan to make in the next version of the manuscript.

read point-by-point responses

Referee: §3: The manuscript establishes that successful DProp executions match stable-model semantics through alternating falsity decisions and truth propagations, yet provides no completeness argument showing that every consistent program admits at least one successful DProp path that recovers all stable models. Without this, the claim that DProp 'captures the stable model semantics' remains partial.

Authors: We thank the referee for this observation. Section 3 provides a proof that any successful execution of DProp produces a stable model, establishing soundness with respect to the stable model semantics. However, we acknowledge that we do not present a completeness argument demonstrating that for every consistent program and every stable model, there exists a sequence of decisions leading to it via DProp. We will revise the abstract and Section 3 to state that 'successful DProp computations correspond to stable models' rather than claiming to fully 'capture the stable model semantics'. We will also add a remark that establishing completeness is an interesting direction for future work. revision: partial
Referee: §4.2: NDProp substitutes classical propagations with fuzzy logic and decisions with a learned network to obtain differentiability. No formal invariant is supplied that relates the resulting fuzzy fixed point back to the classical reduct or guarantees that learned heuristics avoid non-successful branches, so correctness of NDProp rests entirely on the reported empirical results.

Authors: We agree with this assessment. NDProp is designed as a differentiable, learned approximation to DProp, and we do not provide formal guarantees that the fuzzy fixed points correspond to classical stable models or that the learned heuristics always lead to successful branches. The end-to-end differentiability is intended to facilitate learning effective heuristics from data. We will expand Section 4.2 to explicitly state that the correctness of NDProp is supported by the empirical results on the benchmarks, and clarify the approximate nature of the fuzzy propagation. revision: yes
Referee: §5: The experimental section asserts that NDProp improves accuracy and scalability on neuro-symbolic benchmarks, but supplies neither error bars, details on random seeds or data splits, nor statistical significance tests for the reported gains. This absence makes it difficult to assess whether the observed improvements are robust or reproducible.

Authors: We appreciate this feedback on the experimental reporting. In the revised manuscript, we will include error bars computed over multiple independent runs with different random seeds, provide details on the data splits and preprocessing, and report statistical significance tests (such as Student's t-test) comparing NDProp against the baselines to substantiate the improvements in accuracy and scalability. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper defines DProp as an independent alternating procedure of falsity decisions and truth propagations, then states that successful executions capture stable model semantics (a separate proof claim). NDProp is introduced as a differentiable extension using neural decisions and fuzzy propagations, with learning of heuristics evaluated empirically on external neuro-symbolic benchmarks. No equations reduce a prediction to a fitted parameter by construction, no self-citation bears the load of the central semantics claim, and no ansatz or uniqueness result is smuggled in. The derivation remains self-contained against the stated benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no explicit free parameters, axioms, or invented entities; the approach rests on standard ASP stable-model semantics and differentiability of the neural/fuzzy components.

pith-pipeline@v0.9.0 · 5449 in / 1052 out tokens · 57467 ms · 2026-05-10T15:08:45.540414+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

49 extracted references · 41 canonical work pages

[1]

Neurosymbolic

Artur d'Avila Garcez and Lu. Neurosymbolic. Artif. Intell. Rev. , volume =. 2023 , url =. doi:10.1007/S10462-023-10448-W , timestamp =

work page doi:10.1007/s10462-023-10448-w 2023
[2]

NeurASP: Embracing Neural Networks into Answer Set Programming , booktitle =

Zhun Yang and Adam Ishay and Joohyung Lee , editor =. NeurASP: Embracing Neural Networks into Answer Set Programming , booktitle =. 2020 , url =. doi:10.24963/IJCAI.2020/243 , timestamp =

work page doi:10.24963/ijcai.2020/243 2020
[3]

Robin Manhaeve and Sebastijan Dumancic and Angelika Kimmig and Thomas Demeester and Luc De Raedt , title =. Artif. Intell. , volume =. 2021 , url =. doi:10.1016/J.ARTINT.2021.103504 , timestamp =

work page doi:10.1016/j.artint.2021.103504 2021
[4]

Richard Evans and Edward Grefenstette , title =. J. Artif. Intell. Res. , volume =. 2018 , url =. doi:10.1613/JAIR.5714 , timestamp =

work page doi:10.1613/jair.5714 2018
[5]

What Is Answer Set Programming? , booktitle =

Vladimir Lifschitz , editor =. What Is Answer Set Programming? , booktitle =. 2008 , url =

2008
[6]

Gerhard Brewka and Thomas Eiter and Miroslaw Truszczynski , title =. Commun. 2011 , url =. doi:10.1145/2043174.2043195 , timestamp =

work page doi:10.1145/2043174.2043195 2011
[7]

Graph-Based Attention for Differentiable Max

Sota Moriyama and Katsumi Inoue , booktitle=. Graph-Based Attention for Differentiable Max. 2025 , url=

2025
[8]

Learning a

Daniel Selsam and Matthew Lamm and Benedikt B. Learning a. 7th International Conference on Learning Representations,. 2019 , url =

2019
[9]

Graph Neural Networks for Maximum Constraint Satisfaction , journal =

Jan T. Graph Neural Networks for Maximum Constraint Satisfaction , journal =. 2020 , url =. doi:10.3389/FRAI.2020.580607 , timestamp =

work page doi:10.3389/frai.2020.580607 2020
[10]

Answer Set Computation of Negative Two-Literal Programs Based on Graph Neural Networks: Preliminary Results , booktitle =

Antonio Ielo and Francesco Ricca , editor =. Answer Set Computation of Negative Two-Literal Programs Based on Graph Neural Networks: Preliminary Results , booktitle =. 2021 , url =

2021
[11]

Deep Learning for the Generation of Heuristics in Answer Set Programming:

Carmine Dodaro and Davide Ilardi and Luca Oneto and Francesco Ricca , editor =. Deep Learning for the Generation of Heuristics in Answer Set Programming:. Logic Programming and Nonmonotonic Reasoning - 16th International Conference,. 2022 , url =. doi:10.1007/978-3-031-15707-3\_12 , timestamp =

work page doi:10.1007/978-3-031-15707-3 2022
[12]

Theory Pract

Kewen Wang and Lian Wen and Kedian Mu , title =. Theory Pract. Log. Program. , volume =. 2015 , url =. doi:10.1017/S1471068414000611 , timestamp =

work page doi:10.1017/s1471068414000611 2015
[13]

Stable Models and Non-Determinism in Logic Programs with Negation , booktitle =

Domenico Sacc. Stable Models and Non-Determinism in Logic Programs with Negation , booktitle =. 1990 , url =. doi:10.1145/298514.298572 , timestamp =

work page doi:10.1145/298514.298572 1990
[14]

A New Fixpoint Semantics for General Logic Programs Compared with the Well-Founded and the Stable Model Semantics , journal =

Fran. A New Fixpoint Semantics for General Logic Programs Compared with the Well-Founded and the Stable Model Semantics , journal =. 1991 , url =. doi:10.1007/BF03037172 , timestamp =

work page doi:10.1007/bf03037172 1991
[15]

Embedding Negation as Failure into a Model Generation Theorem Prover , booktitle =

Katsumi Inoue and Miyuki Koshimura and Ryuzo Hasegawa , editor =. Embedding Negation as Failure into a Model Generation Theorem Prover , booktitle =. 1992 , url =. doi:10.1007/3-540-55602-8\_180 , timestamp =

work page doi:10.1007/3-540-55602-8 1992
[16]

Katsumi Inoue and Chiaki Sakama , title =. J. Log. Program. , volume =. 1996 , url =. doi:10.1016/0743-1066(95)00119-0 , timestamp =

work page doi:10.1016/0743-1066(95)00119-0 1996
[17]

2004 , url =

Yasuyuki Shirai and Ryuzo Hasegawa , title =. 2004 , url =. doi:10.1007/978-3-540-28633-2\_7 , timestamp =

work page doi:10.1007/978-3-540-28633-2 2004
[18]

Theory Solving Made Easy with Clingo 5 , booktitle =

Martin Gebser and Roland Kaminski and Benjamin Kaufmann and Max Ostrowski and Torsten Schaub and Philipp Wanko , editor =. Theory Solving Made Easy with Clingo 5 , booktitle =. 2016 , url =. doi:10.4230/OASICS.ICLP.2016.2 , timestamp =

work page doi:10.4230/oasics.iclp.2016.2 2016
[19]

2006 , url =

Nicola Leone and Gerald Pfeifer and Wolfgang Faber and Thomas Eiter and Georg Gottlob and Simona Perri and Francesco Scarcello , title =. 2006 , url =. doi:10.1145/1149114.1149117 , timestamp =

work page doi:10.1145/1149114.1149117 2006
[20]

Logic Programming and Nonmonotonic Reasoning, 12th International Conference,

Mario Alviano and Carmine Dodaro and Wolfgang Faber and Nicola Leone and Francesco Ricca , editor =. Logic Programming and Nonmonotonic Reasoning, 12th International Conference,. 2013 , url =. doi:10.1007/978-3-642-40564-8\_6 , timestamp =

work page doi:10.1007/978-3-642-40564-8 2013
[21]

Lampert and Bernt Schiele and Zeynep Akata , title =

Yongqin Xian and Christoph H. Lampert and Bernt Schiele and Zeynep Akata , title =. 2019 , url =. doi:10.1109/TPAMI.2018.2857768 , timestamp =

work page doi:10.1109/tpami.2018.2857768 2019
[22]

Przymusinski , title =

Teodor C. Przymusinski , title =. Artif. Intell. , volume =. 1991 , url =. doi:10.1016/0004-3702(91)90013-A , timestamp =

work page doi:10.1016/0004-3702(91)90013-a 1991
[23]

Melvin Fitting , title =. Theor. Comput. Sci. , volume =. 2002 , url =. doi:10.1016/S0304-3975(00)00330-3 , timestamp =

work page doi:10.1016/s0304-3975(00)00330-3 2002
[24]

Approximations, Stable Operators, Well-Founded Fixpoints and Applications in Nonmonotonic Reasoning

Denecker, Marc and Marek, Victor and Truszczy \' n ski, Miros aw. Approximations, Stable Operators, Well-Founded Fixpoints and Applications in Nonmonotonic Reasoning. Logic-Based Artificial Intelligence. 2000. doi:10.1007/978-1-4615-1567-8_6

work page doi:10.1007/978-1-4615-1567-8_6 2000
[25]

2024 , url =

Akihiro Takemura and Katsumi Inoue , title =. 2024 , url =. doi:10.3233/FAIA240628 , timestamp =

work page doi:10.3233/faia240628 2024
[26]

CoRR , volume =

Taisuke Sato and Akihiro Takemura and Katsumi Inoue , title =. CoRR , volume =. 2023 , url =. doi:10.48550/ARXIV.2306.06821 , note =. 2306.06821 , timestamp =

work page doi:10.48550/arxiv.2306.06821 2023
[27]

Proceedings of the 21st International Conference on Principles of Knowledge Representation and Reasoning,

Renato Lui Geh and Jonas Gon. Proceedings of the 21st International Conference on Principles of Knowledge Representation and Reasoning,. 2024 , url =. doi:10.24963/KR.2024/69 , timestamp =

work page doi:10.24963/kr.2024/69 2024
[28]

ASPeRiX, a first-order forward chaining approach for answer set computing , journal =

Claire Lef. ASPeRiX, a first-order forward chaining approach for answer set computing , journal =. 2017 , url =. doi:10.1017/S1471068416000569 , timestamp =

work page doi:10.1017/s1471068416000569 2017
[29]

Fuzzy Answer Set Programming , booktitle =

Davy Van Nieuwenborgh and Martine De Cock and Dirk Vermeir , editor =. Fuzzy Answer Set Programming , booktitle =. 2006 , url =. doi:10.1007/11853886\_30 , timestamp =

work page doi:10.1007/11853886 2006
[30]

Fuzzy answer sets approximations , journal =

Mario Alviano and Rafael Pe. Fuzzy answer sets approximations , journal =. 2013 , url =. doi:10.1017/S1471068413000471 , timestamp =

work page doi:10.1017/s1471068413000471 2013
[31]

Ross, and John S

Allen. The Well-Founded Semantics for General Logic Programs , journal =. 1991 , url =. doi:10.1145/116825.116838 , timestamp =

work page doi:10.1145/116825.116838 1991
[32]

Arseny Skryagin and Daniel Ochs and Devendra Singh Dhami and Kristian Kersting , title =. J. Artif. Intell. Res. , volume =. 2023 , url =. doi:10.1613/JAIR.1.15027 , timestamp =

work page doi:10.1613/jair.1.15027 2023
[33]

Alessandro. Fundam. Informaticae , volume =. 2009 , url =. doi:10.3233/FI-2009-180 , timestamp =

work page doi:10.3233/fi-2009-180 2009
[34]

Lengning Liu and Enrico Pontelli and Tran Cao Son and Miroslaw Truszczynski , title =. Artif. Intell. , volume =. 2010 , url =. doi:10.1016/J.ARTINT.2009.11.016 , timestamp =

work page doi:10.1016/j.artint.2009.11.016 2010
[35]

Fuzzy Description Logic Programs under the Answer Set Semantics for the Semantic Web , booktitle =

Thomas Lukasiewicz , editor =. Fuzzy Description Logic Programs under the Answer Set Semantics for the Semantic Web , booktitle =. 2006 , url =. doi:10.1109/RULEML.2006.12 , timestamp =

work page doi:10.1109/ruleml.2006.12 2006
[36]

Answer Set Programming Phase Transition:

Yuting Zhao and Fangzhen Lin , editor =. Answer Set Programming Phase Transition:. Logic Programming, 19th International Conference,. 2003 , url =. doi:10.1007/978-3-540-24599-5\_17 , timestamp =

work page doi:10.1007/978-3-540-24599-5 2003
[37]

Answer Set Programming and Neurosymbolic AI: Applications and Future Perspectives (Invited Talk)

Borroto, Manuel and Ielo, Antonio and Mazzotta, Giuseppe and Ricca, Francesco. Answer Set Programming and Neurosymbolic AI: Applications and Future Perspectives (Invited Talk). Hybrid Models for Coupling Deductive and Inductive Reasoning. 2025

2025
[38]

2025 , url =

Wenguan Wang and Yi Yang and Fei Wu , title =. 2025 , url =. doi:10.1109/TPAMI.2024.3483273 , timestamp =

work page doi:10.1109/tpami.2024.3483273 2025
[39]

Theory Pract

Thomas Eiter and Nelson Higuera and Johannes Oetsch and Michael Pritz , title =. Theory Pract. Log. Program. , volume =. 2022 , url =. doi:10.1017/S1471068422000229 , timestamp =

work page doi:10.1017/s1471068422000229 2022
[40]

Theory Pract

Vito Barbara and Massimo Guarascio and Nicola Leone and Giuseppe Manco and Alessandro Quarta and Francesco Ricca and Ettore Ritacco , title =. Theory Pract. Log. Program. , volume =. 2023 , url =. doi:10.1017/S1471068423000170 , timestamp =

work page doi:10.1017/s1471068423000170 2023
[41]

Tenenbaum and Jiajun Wu , title =

Jiayuan Mao and Chuang Gan and Pushmeet Kohli and Joshua B. Tenenbaum and Jiajun Wu , title =. 7th International Conference on Learning Representations,. 2019 , url =

2019
[42]

Chen , title =

Michael K. Chen , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2508.03366 , eprinttype =. 2508.03366 , timestamp =

work page doi:10.48550/arxiv.2508.03366 2025
[43]

Po-Wei Wang, Priya L

Maarten H. van Emden and Robert A. Kowalski , title =. J. 1976 , url =. doi:10.1145/321978.321991 , timestamp =

work page doi:10.1145/321978.321991 1976
[44]

The Alternating Fixpoint of Logic Programs with Negation , journal =

Allen. The Alternating Fixpoint of Logic Programs with Negation , journal =. 1993 , url =. doi:10.1016/0022-0000(93)90024-Q , timestamp =

work page doi:10.1016/0022-0000(93)90024-q 1993
[45]

Improving the Alternating Fixpoint: The Transformation Approach , booktitle =

Ulrich Zukowski and Burkhard Freitag and Stefan Brass , editor =. Improving the Alternating Fixpoint: The Transformation Approach , booktitle =. 1997 , url =. doi:10.1007/3-540-63255-7\_4 , timestamp =

work page doi:10.1007/3-540-63255-7 1997
[46]

Apt and Howard A

Krzysztof R. Apt and Howard A. Blair and Adrian Walker , editor =. Towards a Theory of Declarative Knowledge , booktitle =. 1988 , url =. doi:10.1016/B978-0-934613-40-8.50006-3 , timestamp =

work page doi:10.1016/b978-0-934613-40-8.50006-3 1988
[47]

Gaunt and Marc Brockschmidt and Nate Kushman and Daniel Tarlow , editor =

Alexander L. Gaunt and Marc Brockschmidt and Nate Kushman and Daniel Tarlow , editor =. Differentiable Programs with Neural Libraries , booktitle =. 2017 , url =

2017
[48]

Hospedales and Loizos Michael , title =

Efthymia Tsamoura and Timothy M. Hospedales and Loizos Michael , title =. Thirty-Fifth. 2021 , url =. doi:10.1609/AAAI.V35I6.16639 , timestamp =

work page doi:10.1609/aaai.v35i6.16639 2021
[49]

Differentiable

Matthias Nickles , editor =. Differentiable. Proceedings of the 5th International Workshop on Probabilistic Logic Programming,. 2018 , url =

2018