Source-linked AI summary
Neuro-Symbolic Artificial Intelligence: Current Trends
Md Kamruzzaman Sarker, Lu Zhou, Aaron Eberhart, Pascal Hitzler
TL;DR
NeSy AI research is growing, but how deeply complex formal logic and deep learning can be integrated remains an open question. This article surveys recent conference publications using two categorization schemes and finds shifts in the architectures, representations, and symbolic emphases of the work. It frames background knowledge as a natural path to attempt to improve deep-learning systems.
Problem
The extent to which complex formal logics and deep learning can be more strongly integrated remains unknown, motivating comparison of recent NeSy AI publications with the subfield’s broader vision.
Source: §4 · p. 7Method
Results
Recent top-conference work favors standard deep-learning architectures and distributed representations, with less emphasis on formal logic and extracting symbolic information from trained networks than earlier work.
Source: §3 · p. 6Takeaways & Limitations
The authors describe using additional background knowledge as a natural path to attempt to improve deep-learning systems.
Source: §5 · p. 8The focus on five conferences may miss relevant work, including papers not labeled as NeSy AI and research in other venues or fields.
Source: §3 · p. 4
Abstract
from arXiv · showhide
Neuro-Symbolic Artificial Intelligence -- the combination of symbolic methods with methods that are based on artificial neural networks -- has a long-standing history. In this article, we provide a structured overview of current trends, by means of categorizing recent publications from key conferences. The article is meant to serve as a convenient starting point for research on the general topic.
1. What Is Neuro-Symbolic Artificial Intelligence?
NeSy AI brings neural and symbolic AI traditions together, seeking to combine their complementary strengths. This overview focuses on recent research at leading AI conferences and organizes it using two existing categorization schemes.
What Is Neuro-Symbolic Artificial Intelligence?
- NeSy AI combines artificial neural-network methods with symbolic approaches involving explicit symbol manipulation; this overview includes symbolic methods beyond formal logic.The article treats neural-symbolic, neurosymbolic, and related spellings as equivalent.Source: §1 · p. 1
- The hoped-for combination pairs neural trainability from raw data and robustness with symbolic explainability, provable correctness, and use of expert knowledge.The authors describe potential benefits including out-of-vocabulary handling, learning from small datasets, error recovery, and explainability.Source: §1 · p. 1
- Neural systems typically encode information in weighted connections or distributed activations, unlike symbolic systems’ explicit, human-understandable representations.The authors note that a human observer may not readily recognize what a neural representation encodes.Source: §1 · pp. 1–2
- NeSy AI has roots predating AI and has long been discussed in workshops, but its terminology became systematic only more recently.The article notes a workshop series dating to 2005 and points readers to earlier overviews [3].Source: §1 · p. 2
- The overview categorizes recent papers from leading AI conferences using Kautz’s proposal and a 2005 scheme, then discusses and contrasts the classifications.The article presents the schemes in Section 2, their results in Section 3, and a future outlook in Section 4.Source: §1 · p. 2
2. Categorizing Neuro-Symbolic Artificial Intelligence
The section relates Kautz’s categories to an earlier diagram and contrasts their architectural emphasis with a 2005 survey’s eight dimensions. Together, the schemes describe different ways neural and symbolic processing can be combined and characterized.
Kautz categories
- Kautz’s categories include symbolic inputs and outputs with neural processing, and neural pattern recognition embedded within a symbolic problem solver.The latter pattern is illustrated by AlphaGo, AlphaZero, and self-driving-car approaches.Source: §2 · p. 2
- Another Kautz category compiles symbolic rules away during training, with Deep Learning for Symbolic Mathematics given as an example.The approach is denoted [Neuro ∪compile(Symbolic)].Source: §2 · p. 2
- Kautz also describes a cascade from a neural system into a symbolic reasoner and symbolic reasoning embedded in a neural engine as deliberative, type 2 reasoning.The latter description includes an attention schema in which a goal signals that deliberative reasoning should be executed.Source: §2 · p. 3
The 2005 survey dimensions
- The 2005 survey’s dimensions distinguish integrated from hybrid systems, neuronal from connectionist architectures, local from distributed representations, and standard from nonstandard networks.The article retains an emphasis on integrated systems, where symbolic processing functionalities emerge from neural structures and processes.Source: §2 · p. 3
- The remaining 2005 dimensions distinguish symbolic from logical processing, propositional from first-order logic, extraction from representation, and learning from reasoning.The survey presents these as dimensions rather than binary values; a system may lie along a continuum or span both aspects.Source: §2 · p. 3
Kautz categories and the 2005 diagram
- The 2005 diagram resembles Kautz’s presentation and can serve as an architectural substrate for his categories; it places I/O on the symbolic side, training on the neural side, and reasoning on either.This arrangement matches the article’s earlier best-of-both-worlds framing of neural training and symbolic knowledge and reasoning.Source: §2 · p. 4, Figure 1 · p. 4
3. Neuro-Symbolic Artificial Intelligence: Trends
The review classifies recent NeSy AI research at leading conferences using Kautz’s categories and a 2005 scheme, and reports growth and shifts in the newer sample.
3. Neuro-Symbolic Artificial Intelligence: Trends
- The review covers original NeSy AI research papers published since 2011 at five leading AI conferences, while acknowledging that its selection may miss relevant work and other venues could show a different picture.Papers were identified by reading abstracts and sometimes introductions.Source: §3 · p. 4
- 40 papers appeared since 2017, compared with only one before 2016, indicating a shift toward more mainstream adoption of NeSy AI research.The counts cover the selected conferences.Source: §3 · p. 4, Table 1 · p. 5
- The first four Kautz categories provide a reasonably balanced view of recent publications, while the absent fifth category was intended as a forward-looking goal.Two papers fit two categories.Source: §3 · p. 5, Table 2 · p. 5
- The newer sample has no emphasis on biological plausibility and favors standard deep-learning architectures with distributed representations, unlike the more balanced, localist, and nonstandard work in the 2005 survey.The authors suggest deep learning’s influence and the top-conference restriction may explain the lopsided counts.Source: §3 · p. 6
- Compared with the 2005 survey’s stronger emphasis on formal logic and propositional work, first-order logic now features more prominently and is relevant to Knowledge Representation and Reasoning.Some papers use symbolic information that is not based on formal logic.Source: §3 · p. 6
- Explicit extraction of symbolic information from trained networks is less emphasized than in the past; the authors connect this partly to Kautz categories that often do not involve extraction.The first three categories would usually omit extraction, and the fourth may as well.Source: §3 · p. 6
- The sample shows a rather balanced emphasis on learning and reasoning, with the authors noting strides since deep learning and growing recognition of both aspects’ importance.Both learning first-order logic and reasoning with it had been particularly difficult for neural-network-based systems in the past.Source: §3 · pp. 6–7
- Among desirable neuro-symbolic features, interpretability is most common, while few papers address out-of-distribution handling or learning from small data and hardly any address error recovery.The authors note that classification decisions were often not clear-cut.Source: §3 · p. 7, Table 5 · p. 7
4. Paths Forward
The authors argue that deep learning’s rise calls for systematic study of symbol manipulation and deeper integration with complex logics. They identify deep deductive reasoning as a promising but difficult direction, with scalability issues as logical complexity increases.
4. Paths Forward
- The authors suggest deep learning’s success in areas such as Chess and Go may explain its growing influence on NeSy AI.They describe deep learning as potentially more suitable than earlier connectionist technology for addressing the subsymbolic-symbolic gap.Source: §4 · p. 7
- The authors call for systematic studies of deep-learning systems on elementary symbol-manipulation tasks, whose capabilities could provide components for larger systems.Suggested tasks include term rewriting, list, tree, and graph manipulation, formal grammars, elementary algebra, and logical deduction.Source: §4 · p. 7
- Current top-conference NeSy work involves complex logics more than two decades ago, but much remains non-logical and desired best-of-both-worlds features often remain unmet.Whether deep learning can integrate more strongly with complex formal logics, and to what extent, remains an open research question.Source: §4 · pp. 7–8
- Experiments that increase logical sophistication—from flat annotation tags to class hierarchies and richer Semantic Web logics—could help build a toolbox for complex-logical NeSy.The proposed progression tests whether systems benefit from simple logical implications before attempting more complex background knowledge.Source: §4 · p. 8
- Deep deductive reasoners train neural systems to perform deduction over formal logics, but published results show the task is difficult and faces scalability issues as logic complexity rises.The task can be framed as classifying whether a statement follows from a theory or generating all consequences under finiteness constraints.Source: §4 · p. 8
5. Conclusions
The authors see rising NeSy AI publications, events, and public addresses as evidence of growing awareness in mainstream AI. They frame background knowledge as a natural path for improving deep learning and view neural-symbolic combination as a possible route to stronger AI.
5. Conclusions
- The recent rise in NeSy AI publications, events, and public addresses indicates increasing awareness of the subfield’s importance in mainstream AI.The authors connect this moment with growing exploration of pure deep learning’s limitations.Source: §5 · p. 8
- The authors describe additional background knowledge as a natural path to attempt to improve deep learning systems, with much of this work falling under NeSy AI.They relate this interest to the exploration of limitations in pure deep learning approaches.Source: §5 · p. 8
- The broad consensus is that combining neural and symbolic approaches is at least a path toward stronger AI systems, and may become a stepping stone toward human-level AI.The human-level possibility is presented as an arguably prospective outcome.Source: §5 · p. 8