PaperlayerView exact revision on arXiv

Source-linked AI summary

Neural-Symbolic Learning and Reasoning: A Survey and Interpretation

Tarek R. Besold, Artur d'Avila Garcez, Sebastian Bader, Howard Bowman, Pedro Domingos, Pascal Hitzler, Kai-Uwe Kuehnberger, Luis C. Lamb, Daniel Lowd, Priscila Machado Vieira Lima, Leo de Penning, Gadi Pinkas, Hoifung Poon, Gerson Zaverucha

arXiv:1711.03902v1cs.AI

TL;DR

The perceived brain–mind dichotomy separates symbolic reasoning from statistical machine learning, while integrating learning and cognitive reasoning in real-world applications remains difficult. This survey connects work across these fields and reports that ANNs combining machine learning and data learning are more effective than purely symbolic or connectionist systems, especially with noisy, unstructured real-world data.

  • Problem

    A perceived brain–mind dichotomy separates symbolic automated reasoning from statistical machine learning, while integrating learning and cognitive reasoning for real-world applications remains difficult amid noisy, incomplete, inconsistent data.

    Source: §1 · p. 2, §3 · p. 10, §11 · p. 46

  • Method

    The survey examines initial links among computer science and AI, cognitive science, cognitive neuroscience, and neural-symbolic computation in relation to the brain–mind dichotomy.

    Source: §11 · p. 46

  • Results

    ANNs combining machine learning and data learning are reported more effective than purely symbolic or purely connectionist systems, especially with noisy, unstructured real-world data (de Penning et al., 2010, 2011; Towell & Shavlik, 1994).

    Source: §1 · p. 4

  • Takeaways & Limitations

    The survey identifies initial links among AI, cognitive science, cognitive neuroscience, and neural-symbolic computation as addressing the perceived brain–mind dichotomy.

    Source: §11 · p. 46

    Broad-scope, general-purpose neural architectural models are scarce, and neural explanations may be less compelling for higher-level cognitive functions.

    Source: §4.1 · p. 20

Abstract

from arXiv · show

The study and understanding of human behaviour is relevant to computer science, artificial intelligence, neural computation, cognitive science, philosophy, psychology, and several other areas. Presupposing cognition as basis of behaviour, among the most prominent tools in the modelling of behaviour are computational-logic systems, connectionist models of cognition, and models of uncertainty. Recent studies in cognitive science, artificial intelligence, and psychology have produced a number of cognitive models of reasoning, learning, and language that are underpinned by computation. In addition, efforts in computer science research have led to the development of cognitive computational systems integrating machine learning and automated reasoning. Such systems have shown promise in a range of applications, including computational biology, fault diagnosis, training and assessment in simulators, and software verification. This joint survey reviews the personal ideas and views of several researchers on neural-symbolic learning and reasoning. The article is organised in three parts: Firstly, we frame the scope and goals of neural-symbolic computation and have a look at the theoretical foundations. We then proceed to describe the realisations of neural-symbolic computation, systems, and applications. Finally we present the challenges facing the area and avenues for further research.

1. Overview

The overview frames neural-symbolic computation as a way to combine logic, connectionist models, and uncertainty-aware approaches to cognition. It motivates this integration as a means of bringing learning and reasoning together in cognitive computational systems and applications.

1. Overview

  • Behaviour modelling draws on computational logic for high-level reasoning, connectionist models for lower-level dynamics, and uncertainty models for vague or probabilistic processing.Examples include classical and nonmonotonic logic, neural networks, Bayesian networks, and Markov decision processes.Source: §1 · p. 2
  • Integrated learning-and-reasoning systems show promise across fault diagnosis, computational biology, simulator training, software verification, and industry applications, with theory-plus-data learning more effective on noisy, unstructured real-world data (de Penning et al., 2010, 2011; Towell & Shavlik, 1994).Applications include business process modelling, fraud prevention, multi-agent coordination, and multimodal processing.Source: §1 · pp. 2, 4
  • Neural-symbolic integration seeks to combine logic’s role in modelling thought with connectionism’s adaptive learning and parallelism, translating principles between symbolic and subsymbolic computation (Valiant, 2008).The methodology also examines how symbolic principles can be implemented by connectionist mechanisms and subsymbolic computation described and analysed logically.Source: §1 · pp. 2–3
  • In a prototypical hybrid, neural networks provide parallel computation and robust learning, while logic explains network models through discrete interpretations at multiple abstraction levels (Garcez et al., 2009).The system remains continuous in its neural machinery while offering a discrete logical interpretation.Source: §1 · p. 3
  • Reconciling statistical concept acquisition with symbolic concept manipulation is a key research challenge, addressed through knowledge translation and extraction between logic systems and subsymbolic networks (Valiant, 2003).Rational agents are often conceptualised as cycling between concept acquisition and manipulation during perception and action.Source: §1 · p. 3
  • Multimodal settings may require systems to combine scene classification with temporal and nonmonotonic reasoning, revise answers as information changes, and control errors from real-world data.The passage identifies this need for rapid response and robustness as an opportunity for neural-symbolic systems.Source: §1 · p. 4

2. Prolegomena of Neural-Symbolic Computation

Neural-symbolic computation seeks to integrate logic and connectionism for cognitive modeling and combined learning and reasoning. The section outlines its principles, computational mechanisms, applications, and remaining theoretical-to-practical challenges.

2. Prolegomena of Neural-Symbolic Computation

  • Neural-symbolic computation aims to unite logic and connectionism for cognitive modeling and integrated learning and reasoning, using translations to implement logic, characterize networks, or build hybrid learners.Theory suggests practical equivalence in computability and tractability, but results are coarse and largely existential, leaving constructive translation and implementation questions unresolved.Source: §2 · pp. 4–5
  • Systems range from feedforward and recurrent to symmetric networks; fibring recursively combines networks so one network’s activations can train another, enabling polynomial-function approximation.Fibring illustrates how symbolic principles such as recursion can advance neural-symbolic integration, whose research concerns principles, mechanisms, and applications.Source: §2 · pp. 5–6

2.1 Principles of Neural-Symbolic Integration

  • Neural-symbolic architectures commonly use modular ensembles and hierarchies that progressively abstract raw data, though unstructured networks are also possible.Networks at different levels may be trained independently, combining unsupervised and supervised learning; the resulting parallel model offers added expressiveness at comparatively low computational cost.Source: §2.1 · pp. 6–7

2.2 Mechanisms of Neural-Symbolic Integration

  • Integration depends on efficient computation as well as representation, cycling through knowledge translation, learning and generalization, reasoning, and symbolic extraction for explanation and maintenance.Fibring can integrate meta-level knowledge with object-level networks to represent and learn relations among concepts.Source: §2.2 · p. 7

2.3 Applications of Neural-Symbolic Integration

  • Applications span bioinformatics, fault diagnosis, robotics, verification, and simulators; research combines logic-network translation, empirical case studies, and application-specific design.CILP has contributed to bioinformatics and fault diagnosis [Garcez & Zaverucha, 1999; Garcez et al., 2002], while NSCA has been applied to behavior modeling in training simulators.Source: §2.3 · p. 9

2.4 Neural-Symbolic Integration in a Nutshell

  • The field integrates cognitive abilities such as induction, deduction, and abduction with computational work on logic, probabilities, and learning, and includes emotion and attention/utility modeling.The survey notes successes in simulation, bioinformatics, fault diagnosis, software engineering, model checking, visual information processing, and fraud prevention.Source: §2.4 · p. 9

3. NSCA as Application Example for Neural-Symbolic Computing

NSCA illustrates how neural learning and symbolic reasoning can be combined to learn from uncertain observations, reason about learned knowledge, and represent it in logic. Its applications span driver assessment, visual action understanding, and driving-style recognition, while further operational evaluation remains necessary.

3. NSCA as Application Example for Neural-Symbolic Computing

  • Learning and reasoning remain difficult in real-world applications because observations may be noisy or incomplete, while complex expert behavior can be subjective, non-deterministic, and hard to represent.Even simulator-based training of high-level cognitive skills remains guided by human experts.Source: §3 · p. 10

3.1 Principles of Neural-Symbolic Integration Exemplified

  • NSCA maps temporal logic knowledge into a neural network, using an RBM to integrate inductive learning from examples with deductive reasoning over symbolic knowledge.The architecture learns complex temporal relations from uncertain observations, reasons probabilistically about learned knowledge, and represents it in logic for validation.Source: §3.1 · pp. 10–11

3.2 Mechanisms of Neural-Symbolic Integration Exemplified

  • In NSCA’s RTRBM, recurrent hidden-unit activations encode hypotheses about temporal rules; Bayesian inference selects rules, and observed–inferred belief differences train the model using Contrastive Divergence and backpropagation through time.Each hidden unit represents a rule hypothesis conditioned on observed beliefs and previously applied rules.Source: §3.2 · pp. 11–12
  • NSCA represents binary and continuous beliefs, including vague concepts through fuzzy-set interpretations, and uses cognitive temporal logic to express conditions involving past, future, and continuous variables.For example, learning can refine the subjective distance at which a driver is considered to be approaching an intersection into a more objective value from observations.Source: §3.2 · pp. 12–13

3.3 Applications of Neural-Symbolic Integration Exemplified

  • In simulator trials, NSCA learned from five students’ five-scenario driving tests and inferred assessment scores similar to instructors’ assessments (de Penning et al., 2010, 2011).It observed 43 simulator measurements and numerical assessments from three instructors in real time.Source: §3.3 · p. 13
  • In CORTEX’s Mind’s Eye visual-intelligence tasks, NSCA learned action semantics from observation for recognition, description, anomaly detection, and gap-filling (de Penning, den Hollander, Bouma, G., & Garcez, 2012).The system aimed to describe actions in terms of entities and their properties, and explain the reasoning behind them.Source: §3.3 · p. 13
  • NSCA recognized driving styles from real-time in-car sensor data, outperformed state-of-the-art methods, and described styles with multi-level temporal logic rules for interpretation and validation (de Penning, d’Avila Garcez, Lamb, Stuiver, & Meyer, 2014).The system was included in an Intelligent Transport System intended to reduce CO2 emissions.Source: §3.3 · p. 14

3.4 NSCA in a Nutshell

  • NSCA demonstrates an approach for learning rules from complex real-world data and extracting them for validation, reporting, and feedback, but operational evaluation with experts and larger trainee populations remains needed.Ongoing work includes using Deep Boltzmann Machines to find higher-level rules and applying RTRBMs to adaptive training.Source: §3.4 · p. 14

4. Neural-Symbolic Integration in and for Cognitive Science: Building Mental Models

The section frames mental models as constrained representations used in reasoning and considers how they might be neurally realized. It presents neural-symbolic methods as a way to connect symbolic models of cognition with neural implementation.

4. Neural-Symbolic Integration in and for Cognitive Science: Building Mental Models

  • Mental models represent real or imagined situations, their parts and relations, and inherent constraints; their construction and manipulation support reasoning and may contribute to abduction and creativity.How such models are neurally realized remains uncertain.Source: §4 · pp. 14–15
  • The binding problem is to represent which neural assemblies are associated; conjunctive codes seem readily integrable, but require scalable distributed resources and capacity for novel conjunctions.Randomly preconfigured resources are proposed to support combinations not previously experienced; the authors regard neural conjunctive coding as prerequisite to brain-based mental-model theories.Source: §4 · pp. 15–16
  • Reasoning may proceed through associative statistical inference or deliberate exploration of hypothetical possibilities, while emotion can bind multidimensional evaluations to nonaffective parts of mental models.The passage distinguishes emotions such as fear and disgust by their different response implications despite both being negatively valenced.Source: §4 · p. 16

4.1 Cognitive Neuroscience and Neural-Symbolic Methods

  • Although ANNs model sensory, attention, and memory capacities, perceived mismatches around language rules and compositionality, and difficulty with deliberative high-level functions, motivate symbolic cognitive models.The section suggests neural models are less natural for broad architectural and deliberative functions, without claiming those capacities cannot be neurally modelled.Source: §4.1 · pp. 19–20
  • A proposed cognitive-neuroscience route models behavior symbolically, then maps that model to neural networks through computational-logic-to-ANN transformation.The mapping should also be constrained by neurophysiological plausibility, including realistic activation dynamics, distinct excitation and inhibition, and biologically realistic learning.Source: §4.1 · pp. 20–21

5. Putting the Machinery to Work: Binding and First-Order Inference in a Neural-Symbolic Framework · 5.1 A Computational View on Fodor, Pylyshyn, and the Challenge(s) for Modelling Cognition

The section examines how neural-symbolic systems can represent and process complex combinatorial structures, especially first-order logic, while addressing cognitive-modeling requirements and shortcomings of earlier approaches.

5. Putting the Machinery to Work: Binding and First-Order Inference in a Neural-Symbolic Framework

  • Human cognition is argued to produce and process complex combinatorial structures, but their systematic, dynamic creation poses challenges for ANNs and neuro-cognitive theories.The section focuses on representing and efficiently learning and processing such structures, especially with first-order logic, for high-level cognitive tasks.Source: §5 · p. 21
  • The framework uses variable binding to encode first-order logic expressions compactly in activation, then encodes symbolic constraints in symmetric recurrent-network weights so inference solutions emerge in activation.The section presents this as a practical counterpart to earlier theoretical and cognitive-neuroscience discussions.Source: §5 · p. 21

5.1 A Computational View on Fodor, Pylyshyn, and the Challenge(s) for Modelling Cognition

  • Fodor and Pylyshyn (1988) identify combinatorial syntax and semantics as essential: complex mental representations are recursively built from atomic ones, with whole-structure semantics determined by component semantics.The section also identifies structure sensitivity as their second essential characteristic.Source: §5.1 · pp. 21–22
  • Structure sensitivity means that operations on a structured representation are applied based on its syntactic structure.Source: §5.1 · pp. 21–22
  • Earlier approaches only partially answered Fodor and Pylyshyn’s challenge and none captured full FOL expressive power; many were localist, fragile to neural damage, weak in learning, large in size-complexity, or required KB-specific engineering.The cited examples span Anandan, Letovsky, & Mjolsness (1989), Ballard (1986), Garcez et al. (2009), H¨olldobler (1990), H¨olldobler & Kurfess (1992), Stolcke (1989), and Van der Velde & De Kamps (2006).Source: §5.1 · p. 22
  • Only a few earlier systems used distributed representations, and those typically suffered information loss and limited processing ability.The passage cites Plate (1995), Pollack (1990), and Stewart & Eliasmith (2008).Source: §5.1 · p. 22

5.2 The Binding Problem Computationally Revisited

High-level cognitive modeling, especially language processing, remains difficult because binding must combine items encoded in distinct circuits, including through variable binding. Computational accounts must also explain how neural representations are encoded, processed, and used for reasoning.

The Binding Problem Computationally Revisited

  • Variable binding involves four challenges: massiveness, multiple instances of a fact, associating values with variables, and binding in short-term versus long-term memory (Jackendorff, 2002).The section frames binding as a fundamental obstacle because each item depends on it.Source: §5.2 · p. 22
  • For binary-valued neural units, computational accounts must explain how first-order-logic expressions are encoded, how constituents form complex structures, and how those structures support unification and reasoning.The section also asks how procedural knowledge for processing such structures is represented in synapses.Source: §5.2 · p. 22
  • Proposed variable-binding approaches (Anandan et al., 1989; Barrett et al., 2008; Browne & Sun, 1999; Shastri & Ajjanagadde, 1993; Van der Velde & De Kamps, 2006) still face limits in expressiveness, space requirements, or central control demands.A binding mechanism must glue together knowledge-base items such as predicates, functions, constants, and variables.Source: §5.2 · pp. 22–23

5.3 Inference Specifications as Fixed Points of ANNs

The section treats inference as a search for stable ANN states that encode valid solutions. Weighted Boolean constraints can shape energy minima, while FOL inference chains can be represented as proofs that emerge at equilibrium.

5.3 Inference Specifications as Fixed Points of ANNs

  • With clamped inputs, ANN fixed points can represent problem solutions; symmetric networks map valid FOL inference chains to global energy minima, while higher-order Boltzmann machines search for constraint-satisfying solutions (Pinkas, 1991a; Lima, 2000; Pinkas, Lima, & Cohen, 2012, 2013).Constraints are compiled or learned into weights. The KB may be compiled into weights or clamped onto visible neurons, while queries, short-term facts, or optional cues may be clamped onto working memory.Source: §5.3 · p. 23

5.3.1 ANN Specifications Using Weighted Boolean Constraints

  • Weighted Boolean constraints become an energy function whose minima maximize satisfaction, allowing partial satisfaction, preferred solutions, and non-monotonic penalty-logic inference (Pinkas, 1991b; Pinkas, 1995).A compiler generates ANNs for FOL retrieval, unification, and inference; constraints may alternatively be learned unsupervised with an anti-Hebbian rule.Source: §5.3.1 · pp. 24–25

5.3.2 ANN Specifications of FOL Inference Chains

  • Proof-theoretic FOL inference can use ungrounded formulae and yield explanations, avoiding grounding’s exponential variable growth or expressive limits; implementations include resolution and most general unification with occurrence checking.The passage contrasts proof theory with model-theoretic checking of grounded knowledge bases and notes that proofs can explain inferred statements.Source: §5.3.2 · p. 25
  • At a global energy minimum, an ANN’s stable state yields a proof chain built by retrieval, copying, or resolution; proof competition can favor likely explanations, probable inferences, least-cost plans, or preferred actions.Inference may search by refutation, a complete strategy, or directly infer a conclusion, which is incomplete but described as more cognitively plausible.Source: §5.3.2 · p. 25

5.4 A Fault Tolerant Mechanism for Dynamic Binding

The mechanism uses compact, reusable spatial binders to build working memory (WM) for retrieving long-term-memory (LTM) knowledge, unification, and inference. General-purpose binders provide fault tolerance while supporting graph-structured representations.

5.4 A Fault Tolerant Mechanism for Dynamic Binding

  • The binding mechanism underlies the FOL encoding and inference engine, constructing WM to retrieve needed LTM knowledge for unification and inference.The number of binders needed is proportional to proof size, not the entire knowledge base.Source: §5.4 · pp. 25–26
  • Spatial conjunctive binding encodes compositionality in stable neural states, avoiding timing synchronisation and time-slot allocation required by temporal binding (Shastri & Ajjanagadde, 1993).The compact mechanism uses only a few ensembles, each able to bind any of the many LTM items (Pinkas et al., 2012).Source: §5.4 · p. 26
  • General Purpose binders act as indirect pointers, dynamically linking retrieved FOL objects and other binders to represent nested terms, complete proofs, and arbitrary directed graphs.Weighted Boolean constraints compiled or learned into synapses manipulate these compound representations; arbitrary digraphs include FOL as a special case (Pinkas, 1991a; Lima, 2000).Source: §5.4 · p. 26
  • Only proof-needed binders are required, reducing WM size complexity to O(n · log(k)), where n is proof length and k is KB size (Pinkas et al., 2013).Fault tolerance comes from synaptic redundancy: any general-purpose binding ensemble can replace a failed one, rather than relying on redundant units.Source: §5.4 · p. 26

5.5 Examples of Inferences, Cues, and Deductive Associations

A Boltzmann-machine working memory using GP-binders can answer a grandparent query by binding a variable and constructing a resolution proof. The same mechanism can use cues to guide deduction and produce probabilistic associations.

5.5 Examples of Inferences, Cues, and Deductive Associations

  • Clamping GrandParent(V, Anne) lets the network bind V to Anne’s maternal grandmother and find a complete inference proof.The proof retrieves and unifies rules, resolves them, and concludes GrandParent(mother(mother(Anne)), Anne).Source: §5.5 · pp. 26–27
  • FOL clauses, literals, atoms, and terms are DAGs, while dynamically allocated GP-binders encode nodes and labeled arcs and retrieve only needed rules.Other KB rules remain in long-term memory rather than being retrieved into working memory when they are unnecessary for the inference.Source: §5.5 · p. 27
  • The query acts as a cue directing the network’s search, while specifying or clamping rules in working memory can accelerate convergence.The example suggests clamping rule 1, rule 2, or both before network operation.Source: §5.5 · p. 27
  • With “Anne” and “mother” as cues, the system can converge to several possible inferences, whose likelihood depends on KB-item weights and inference-chain depth.Possible associations include Anne having a mother or a grandmother, or Anne’s mother having a mother who also has a mother.Source: §5.5 · p. 27

5.6 Can Neurones Be Logical? Fundamental Negative and Positive Results

Compact attractor-based ANNs can represent memory-bound FOL inference, but cycles undermine guaranteed global solutions. Stochastic local search can also stall at local minima, motivating learning methods intended to improve search.

5.6 Can Neurones Be Logical? Fundamental Negative and Positive Results

  • Although compact attractor-based ANNs can represent memory-bound FOL inference, cyclic networks may oscillate and remain incomplete; Pinkas & Dechter (1995) show a global-solution guarantee is possible only without cycles.Plausible assumptions on the daemon can ensure pathological scenarios do not last forever.Source: §5.6 · p. 27
  • Boltzmann machines and other symmetric networks use stochastic local search, but local minima can prevent them from finding a solution in finite time.The search targets a global minimum of the network's energy function.Source: §5.6 · p. 28
  • Research explores Hebbian-learning and Boltzmann-machine variations to eliminate local minima and find global minima faster, an approach also associated with deep learning's success.These learning styles are called “speed-up practicing” because they use learning as a strategy for self-improvement.Source: §5.6 · p. 28

6. Connectionist First-Order Logic Learning Using the Core Method

The section examines how connectionist systems can represent and learn first-order rules, both to model cognition neurally and to support distributed, robust, self-improving symbolic computation. It focuses on approximating first-order logic programs’ immediate consequence operators and on learning them.

6. Connectionist First-Order Logic Learning Using the Core Method

  • A single consequence-operator application can yield infinitely many atoms, and rational models offer a finite representation whose size has no computable upper bound.The section therefore focuses on Herbrand interpretations and contrasts approximation in space with approximation in time: increasing hidden units versus small networks that improve with runtime. Rational models are discussed by Bornscheuer (1996); the space approach by Bader (2009a) and Bader, Hitzler, Hölldobler, and Witzel (2007), and the time approach by Bader and Hitzler (2004) and Bader, Hitzler, and Garcez (2005).Source: §6 · pp. 28–29

6.1 Feasibility of the First-Order Core Method and the Embedding of First-Order Rules into Neural Networks

  • A level-mapping embedding represents interpretations as base-b encodings and preserves structure, enabling continuous real-valued versions of consequence operators to be approximated by connectionist networks.For any given approximation accuracy, standard sigmoidal and radial basis function networks can be constructed.Source: §6.1 · pp. 29–30
  • Iterating the approximating function converges to a state corresponding to a program model when the immediate consequence operator is contractive; the approximation is contractive on the embedded interpretation set, though not generally on R.The convergence claim follows from the Banach contraction mapping principle.Source: §6.1 · p. 30

6.2 Learning in a First-Order Setting

  • Training on interpretation–consequence pairs enabled a randomly initialized network to learn the immediate consequence operator sufficiently accurately for iterated outputs to converge to the program’s model.The network was evaluated by feeding random interpretations into it and feeding each output back as the next input.Source: §6.2 · p. 30
  • Practical first-order learning also draws on inductive logic programming and propositionalisation, while connectionist modal logic can learn a two-variable fragment of first-order logic as a representationally capable, efficient middle ground.The section describes early propositionalisation results as promising and notes that propositional modal logic is robustly decidable.Source: §6.2 · p. 30

7. Markov Logic Networks Combining Probabilities and First-Order Logic

The section presents Markov logic networks (MLNs) as a way to combine first-order logic with probabilistic graphical models, then discusses their inference and learning challenges alongside sum-product networks (SPNs) and relational SPNs (RSPNs). It frames these formalisms as examples of symbolic-numerical integration with applications in AI.

7. Markov Logic Networks Combining Probabilities and First-Order Logic

  • MLNs soften first-order formulae with weights that serve as features in a Markov network, combining logic’s high-order dependencies with probabilistic handling of noise and uncertainty (Richardson & Domingos, 2006; Domingos & Lowd, 2009).Constants ground the formulae into network nodes and features, and state probabilities depend on the weights of satisfied formulae.Source: §7 · p. 31
  • Grounding can make MLNs prohibitively large: 1,000 people yield 1 million possible friendships and 1 billion transitivity relations, motivating lifted and query-focused inference.Lifted inference groups symmetric variables or configurations; query-focused, lazy, and coarse-to-fine methods limit the network or computation needed (Braz, Amir, & Roth, 2005; Singla & Domingos, 2008; Riedel, 2008; Singla & Domingos, 2006; Poon, Domingos, & Sumner, 2008; Kiddon & Domingos, 2011).Source: §7 · pp. 31–32
  • SPNs answer arbitrary probability queries by evaluating sums and products bottom-up in linear time; the worked query P(X1 = true, X3 = false) evaluates to 0.49.SPNs are motivated by seeking conditions for a tractable partition function; their internal sums and products reuse intermediate results.Source: §7 · pp. 32–33
  • Under completeness and consistency, SPNs represent a graphical model’s partition function and marginals, and experiments report faster, more accurate inference and learning than standard deep networks, including image completion improvements.SPNs support learning with backpropagation and EM and are strictly more general than tractable graphical models cast as SPNs.Source: §7 · p. 33
  • RSPNs extend SPNs to relational domains with always-tractable inference, and applications in social networks and automated debugging were much faster and more accurate than MLNs (Nath & Domingos, 2015).They specify classes of objects, attributes, parts, and relations, with a tractable distribution learned from data.Source: §7 · pp. 33–34
  • The section concludes that MLNs and related formalisms are increasingly popular AI applications that integrate symbolic representations with subsymbolic graphical features and reasoning.The authors present them as systems-level neural-symbolic examples that complement other approaches, including NSCA and Penalty Logic.Source: §7 · p. 34

8. Relating Neural-Symbolic Systems to Human-Level Artificial Intelligence

The section presents neural-symbolic integration as a research direction for human-level artificial intelligence (HLAI), arguing that progress requires reconciling neural and symbolic approaches in theory and computation.

HLAI and interdisciplinary research

  • HLAI research is regaining attention as specialised systems leave some complex tasks beyond their scope and newer methods and neuroscience developments offer ways to address general intelligence.The section calls for substantial convergence among artificial intelligence, neuroscience, psychology, and computational linguistics, pursued through tractable next steps.Source: §8 · p. 35

The proposal for neural-symbolic integration

  • The proposal combines neural models of biological minds with symbolic representations and logic, which the section regards as necessary for encoding complex tasks.The two perspectives are orthogonal in the current state of the art: biological minds are implemented by neural systems, while complex task modelling relies on symbolic structures and logic.Source: §8 · p. 35

The gap between neural and symbolic approaches

  • Human brains support symbolic reasoning, but current neural and symbolic theories remain substantially disconnected, creating a mismatch between human capabilities and computational models.Symbolically understanding neural systems remains difficult, while current artificial neural approaches fall short of representing and manipulating complex symbolic knowledge at the level envisioned for HLAI.Source: §8 · p. 36

The proposal for neural-symbolic integration

  • The authors argue that reconciling neural and symbolic theories and computational models is central to HLAI, cognitive science, and cognitive neuroscience, and is timely and feasible.They frame neural-symbolic integration as a practical research direction toward HLAI rather than a single leap to an all-encompassing theory.Source: §8 · pp. 35–36

Developments and challenges

  • Neural-symbolic systems have progressed from propositional logic to first-order logic, but remain severely limited in scope and applicability.The line of work formally defines and implements artificial neural networks that learn and reason with logic representations of complex domains.Source: §8 · p. 36
  • Advancing these systems toward HLAI requires ideas from cognitive psychology, neuroscience-inspired neural paradigms, dynamical systems theory, and organic computing.The selected research line exemplifies computer science, machine learning, and classical AI augmented by cognitive science and neuroscience.Source: §8 · pp. 36–37

9. (Most) Recent Developments and Work in Progress from the Neural-Symbolic Neighbourhood

This section surveys work near the neural-symbolic core: new computational architectures, connectionist systems for symbolic tasks, and research programs probing interaction, interpretability, and complexity.

9.1 Other Paradigms in Computation and Representation Narrowing the Neural-Symbolic Gap

  • Conceptor filters select and stabilize RNN processing modes, while Neural Turing Machines (NTMs) couple a controller network to addressable memory for differentiable computation (Jaeger, 2014; Graves, Wayne, & Danihelka, 2014).Conceptors characterize modes through state-cloud geometry and can be combined using operations akin to Boolean logic. NTMs use selective, probabilistic read/write operations, and experiments show they can learn simple algorithms from examples and generalize beyond training.Source: §9.1 · pp. 37–38, Figure 6 · p. 38
  • Memory networks combine indexed memory with input, generalization, output, and response components, and can use learned or standard machine-learning mechanisms (Weston, Chopra, & Bordes, 2015).A text-based question-answering implementation was evaluated in large-scale and simulated settings, including questions with previously unseen words, and in a hybrid system proof of concept.Source: §9.1 · pp. 38–39

9.2 Other Application Systems Narrowing the Neural-Symbolic Gap

  • Connectionist systems addressed image description, image-region and sentence alignment, and visual analogy tasks that had often been treated as symbolic (Vinyals et al., 2014; Karpathy & Fei-Fei, 2015; Reed et al., 2015).The approaches used CNN/RNN language generation, shared image-text embeddings with alignment objectives, and encoder-decoder networks for analogies involving appearance, rotation, 3D pose, and object attributes; the image-sentence system reported state-of-the-art ranking performance and beat retrieval baselines.Source: §9.2 · pp. 39–40
  • Deep distributed recurrent Q-networks learned communication protocols and solved two classical logic riddles requiring collaborative information exchange, though scalability remained unproven (Foerster et al., 2016).The approach shared agent-network weights while conditioning on agent ID, supplied each agent’s previous action as input, and disabled experience replay to avoid non-stationarity.Source: §9.2 · p. 40
  • AlphaGo combined learned policy and value networks with Monte Carlo tree search to play Go at a level on par with the world’s best human players (Silver et al., 2016).The networks learned from human expert games and massively parallel self-play, guiding search that pruned and narrowed the game tree before lookahead selected actions.Source: §9.2 · p. 41

9.3 Other Research Efforts and Programs Narrowing the Neural-Symbolic Gap

  • A proposed situated-agent architecture links environmental sensing and interaction with knowledge formation, symbolic reasoning, and feedback to lower layers (Besold et al., 2015; Besold & K¨uhnberger, 2015).Its layers ground and dynamically repair mappings between sensed objects and symbols, lift representations through neural learning and temporal knowledge representation, then apply symbolic analogy-making, concept blending, and concept reformation.Source: §9.3 · pp. 41–42
  • LSTM analyses found interpretable cells tracking long-range dependencies and task-specific structure, and suggested methods for further knowledge-extraction studies (Karpathy et al., 2015).Character-level language models revealed cells tracking code properties such as parentheses, line length, and indentation; LSTMs also kept track of long-range interactions remarkably well compared with finite-horizon n-gram models.Source: §9.3 · pp. 42–43
  • A complexity-theory research program proposes examining polynomial overhead and additional resources to explain practical differences between neural and logical approaches (Besold & K¨uhnberger, 2015).It considers connectionist resources such as spike counts and samples to convergence, alongside symbolic-side tools including parameterised and descriptive complexity; the work is intended to clarify theoretical and practical tensions and inform integration challenges.Source: §9.3 · pp. 43–44

10. Challenges and Future Directions

The section identifies unresolved challenges in integrating neural and symbolic learning and reasoning, and argues for foundational research alongside promising multimodal applications.

10. Challenges and Future Directions

  • Hybrid architectures and formal connections offer routes toward convergence, but integrated systems remain immature and mechanisms often remain insufficiently unified.The survey notes examples including hybrid architectures (Sun, 2002), CLARION (Sun, 2003), symbolic/sub-symbolic multi-context systems, and links among neural, probabilistic, and logical models.Source: §10 · p. 44
  • Key challenges are defining representational limits, finding representations that support integrated learning and reasoning, and extracting knowledge from very large networks.Current efforts often focus on learning or reasoning separately rather than principled integration that contributes to both areas.Source: §10 · pp. 44–45
  • Open research includes reconciling first-order logic learning and reasoning, semi-supervised and incremental learning, parallelism gains, agent learning-reasoning-acting cycles, and ensemble-structure learning.Source: §10 · p. 45
  • Further issues include scalable rule extraction and interpretability, practical fibring, explaining neural-symbolic application differences, ANN proof and type theories, scientific-method automation, attention, emotions, and utility.The listed directions include analogical and abductive neuro-symbolism, and models of attention focus, emotions, and utility functions.Source: §10 · p. 45
  • Some topics are gaining attention, but knowledge remains limited; progress calls for foundational synthesis across computer science, cognitive science, and cognitive neuroscience, not brain-modeling or engineering alone.Source: §10 · p. 45
  • Multimodal processing could combine neural image models with legacy symbolic text systems through loose functional integration such as fibring, using inconsistencies to trigger learning.The survey presents this as a potential high-impact application for areas including the web, intelligent applications and tools, and security.Source: §10 · pp. 45–46

11. Concluding Remarks

The survey connects AI, cognitive science, cognitive neuroscience, and neural-symbolic computation to address the perceived divide between brain and mind, symbol-based reasoning, and statistical learning. It identifies effective integration, expressive reasoning, and robust learning as goals, while emphasizing practical and computational challenges.

11. Concluding Remarks

  • The survey links AI, cognitive science, cognitive neuroscience, and neural-symbolic computation to address the perceived dichotomy between brain and mind, symbolic reasoning, and statistical machine learning.Source: §11 · p. 46
  • Neural-symbolic systems combine connectionist machines with logical abstractions.Source: §11 · p. 46
  • Integrating symbolic and connectionist approaches may benefit computing and cognitive sciences, with practical advantage to be demonstrated against purely symbolic or connectionist systems.Source: §11 · p. 46
  • Effective integration requires expressive reasoning and robust learning while balancing tractability with expressiveness and addressing efficient logical-model computation and comprehensible knowledge extraction.Practical application challenges arise in engineering, robotics, and the semantic web.Source: §11 · p. 46
  • The ultimate goal is a unifying theory, or faster principled development of cognitive and computing sciences and AI, through neural-symbolic systems with expressive reasoning and robust learning.Source: §11 · p. 47

Related papers

Loading 1711.03902v1…