Source-linked AI summary
Neurosymbolic AI: The 3rd Wave
Artur d'Avila Garcez, Luis C. Lamb
TL;DR
AI’s impact has intensified questions about trust, safety, interpretability, accountability, and the limits of deep learning alone. The paper reviews 20 years of neurosymbolic AI to identify ingredients, technical directions, and challenges for integrating neural learning with symbolic representation and reasoning.
Problem
Deep learning has driven major advances, but concerns about trust, safety, interpretability, accountability, and reasoning beyond training distributions motivate principled neural-symbolic integration.
Method
The paper places 20 years of neural-symbolic integration research in context, revisits foundational theoretical results, and identifies bottlenecks and promising directions.
Results
The review identifies key components and open challenges for neurosymbolic AI, including integrating symbolic descriptions with neural networks to support intervention and counterfactual reasoning.
Takeaways & Limitations
Neurosymbolic AI offers a cohesive research direction combining deep learning with symbolic reasoning for systems requiring sound representation, reasoning, and explainability.
Takeaways & Limitations
Neurosymbolic research faces a tension between principled integration and practical value, including a trade-off between effective learning and sound reasoning.
Abstract
from arXiv · showhide
Current advances in Artificial Intelligence (AI) and Machine Learning (ML) have achieved unprecedented impact across research communities and industry. Nevertheless, concerns about trust, safety, interpretability and accountability of AI were raised by influential thinkers. Many have identified the need for well-founded knowledge representation and reasoning to be integrated with deep learning and for sound explainability. Neural-symbolic computing has been an active area of research for many years seeking to bring together robust learning in neural networks with reasoning and explainability via symbolic representations for network models. In this paper, we relate recent and early research results in neurosymbolic AI with the objective of identifying the key ingredients of the next wave of AI systems. We focus on research that integrates in a principled way neural network-based learning with symbolic knowledge representation and logical reasoning. The insights provided by 20 years of neural-symbolic computing are shown to shed new light onto the increasingly prominent role of trust, safety, interpretability and accountability of AI. We also identify promising directions and challenges for the next decade of AI research from the perspective of neural-symbolic systems.
1 Introduction
AI and deep learning have transformed research and industry while intensifying debates about their broader impact. This paper situates two decades of neurosymbolic research within those debates and organizes key challenges and directions for combining learning, reasoning, and representation.
- Deep learning has reshaped computer science fields including speech recognition, computer vision, natural language processing, and machine translation.
- The paper revisits 20 years of neural-symbolic integration to identify foundational results, bottlenecks, and promising directions for representing learning and reasoning.
- The authors organize related AI, machine-learning, and deep-learning terminology for readers entering the field from other disciplines.
- The paper frames its discussion around AI building blocks, variable grounding, commonsense reasoning, and distributed versus localist representations.
2 Neurons and Symbols: Context and Current Debate
The current debate contrasts deep learning’s achievements with concerns about brittleness, explainability, robustness, and extrapolation. The paper reframes the issue as designing principled systems that combine distributed learning, symbolic representation, and reasoning while balancing scientific integration with practical value.
- Deep learning achieves groundbreaking results but remains vulnerable to adversarial attacks, limited explainability, and high data or computational demands.
- The paper asks how symbols can be computed and learned inside or outside neural networks under precise or approximate reasoning settings.
- Current neural networks are limited by propositional fixation, a limitation attributed to simple neuron models.
- Principled integration can clarify neural-network limits, whereas applied systems may favor effective combinations of existing knowledge bases, deep-learning tools, and software agents.
- A central design tension is whether components such as attention, probability, and theorem proving should reside in neural, symbolic, or hybrid systems.
- Neurosymbolic AI combines statistical machine learning, knowledge representation, and classical or non-classical logical reasoning.
3 Distributed and Localist Representation
Neurosymbolic systems aim to bridge distributed representations that support learning with localist or symbolic representations that support reasoning and explanation. The paper surveys embedding, tensorization, and taxonomy-based approaches while identifying unresolved tensions between effective learning and sound reasoning.
- Deep learning favors distributed representations for learning and optimization, while extrapolation, explainability, and goal-directed reasoning motivate a bridge to localist representations.
- Neural-symbolic computing can represent neural-network knowledge symbolically and perform reasoning symbolically or in distributed form within the network.
- Symbolic constraints are expected to supplement neural learning by addressing catastrophic forgetting, extrapolation difficulties, and out-of-distribution data.
- Embedding methods transform symbolic representations into vector spaces where reasoning uses matrix computations and distance functions.
- Logic Tensor Networks ground logical concepts in tensors, use logical statements as constraints, and support differentiable knowledge organization and querying.
- Kautz’s taxonomy distinguishes tightly coupled localist systems, distributed soft-constraint systems, and systems intended to perform symbolic reasoning inside neural engines.
4 Neurosymbolic Computing Systems: Technical Aspects
Neurosymbolic systems combine neural learning with symbolic representations and reasoning, with design choices spanning integrated neural approaches and hybrid neural-symbolic interfaces. Their technical promise includes learning relations from few examples, deriving symbolic descriptions for extrapolative reasoning, and organizing neural models within established symbolic hierarchies.
- Representations: Symbolic machine learning searches over representations for classification or regression, including relational and probabilistic logical rules.Decision trees are a familiar example; richer approaches use first-order logic and probabilistic extensions to represent uncertainty.
- Representations: Deep learning searches continuous spaces with neural networks and stochastic gradient descent, producing vector-based distributed representations for classification or regression.These sub-symbolic systems handle large-scale multimodal noisy perceptual data and use efficient propagation and tensor processing.
- Reasoning: Recurrent neural networks can compute logical consequences of propositional modal logic programs and other forms of non-classical reasoning.The paper also identifies many-valued logics as languages for describing neural networks.
- Reasoning: Learning relations from relatively few examples can produce symbolic descriptions that support reasoning beyond the data distribution and extrapolation.For example, an acquired ancestor relation can support reasoning over arbitrarily long chains of family relationships.
- System designs: Neurosymbolic AI combines neural and symbolic strengths by translating symbols into neural networks or connecting neural networks to symbolic reasoning systems.The second option preserves a neurosymbolic interface between continuous approximate learning and discrete exact reasoning.
- Open challenges: Neurosymbolic computing can map newer neural models into existing symbolic hierarchies, helping organize otherwise ad hoc work in neural computation.The choice between integrated and hybrid designs depends partly on the application and the availability of quality data and knowledge.
5 Challenges for the Principled Combination of Reasoning and Learning
The paper frames principled neurosymbolic AI as a combination of neural learning and symbolic reasoning that must address mismatched computational regimes, representation, explainability, and scalability. It argues that grounded symbols, modularity, and reasoning beyond the training distribution are central to this agenda.
- Computational Foundations: Existential quantifiers and function symbols challenge reasoning systems, while universal quantification challenges learning systems.The paper identifies these mismatches as a central difficulty in combining reasoning and learning.
- Computational Foundations: Learning operates over continuous, differentiable spaces, whereas reasoning generally operates over discrete spaces, motivating hybrid neural-symbolic systems.Neural networks handle continuous search and probability learning; symbolic systems handle logical descriptions and discrete reasoning.
- Architectural Design: Modularity remains important because compositional rule extraction from stacked networks can compound accuracy loss when learned knowledge is not modular.The authors therefore reject monoblock networks with millions of parameters as a preferred direction.
- Representation and Generalization: Knowledge should be grounded in vector representations for efficient learning through neural-network message passing.The paper presents vector grounding as a basis for connecting knowledge with data-driven learning.
- Representation and Generalization: Symbols extracted from trained networks can support compositional discrete reasoning and extrapolation beyond the data distribution.The proposed role for symbols includes providing an abstract description language while retaining links to learned knowledge.
- Explainability: Knowledge extraction supports explainability, but compact, correct, and complete explanations are difficult to obtain efficiently.When full extraction is impractical, the paper notes that local explanations may be necessary; fidelity measures agreement with the neural network.
6 AAAI 2020, a Turning Point
The AAAI 2020 discussions marked convergence around integrating symbolic reasoning with deep learning while exposing unresolved choices about reasoning, representation, and AI’s goals. The paper distills a practical neurosymbolic recipe and identifies formal, commonsense, and human-network challenges.
- AAAI 2020 discussions: AAAI 2020 featured convergence toward integrating symbolic reasoning with deep learning, alongside debates about whether the goal is brain modeling or robust AI.The paper notes these goals may lead to separate research directions, although both confront the emergence of symbolic meaning from neural networks.
- Reasoning choices: A central design choice is combining efficient approximate reasoning with more deliberative, precise symbolic reasoning, allowing conclusions to be revised through new observations.The paper contrasts combinatorial exact reasoning with commonsense approximate reasoning and envisions their combination.
- Core ingredients: A practical recipe assigns data learning and efficient propositional reasoning to neural networks, while symbolic descriptions support rich first-order reasoning and extrapolation.The recipe separates approximate neural learning and reasoning from richer symbolic reasoning over extracted descriptions.
- Core ingredients: Neurosymbolic AI combines gradient-based learning, modularity, symbolic language, reasoning, and constraint satisfaction as five core ingredients.These ingredients are presented as the basis for combining neural structure with symbolic language and reasoning.
- Open technical directions: The paper calls for principled formal correspondences between neural architectures and logical languages, supported by proofs of representational capability.It cites existing correspondences for propositional, nonmonotonic, modal, and epistemic logic as foundations for this direction.
- Open technical directions: Three immediate challenges are provably sound and efficient knowledge extraction, goal-directed commonsense reasoning, and human-network communication through questioning and argumentation.Together, these challenges span model explanation, reasoning over multimodal learning, and interactive multi-agent use.