Neuro-Symbolic Artificial Intelligence: A Review

Abhishek Tiwari Rojalin Samantaroy

Computer Science Computer Science
Sri Aurobindo Institute of Technology Sri Aurobindo Institute of Technology
Indore, India Indore, India
[email protected] [email protected]

Abstract—Human intelligence deftly weaves sensory synergistic fusion designed to leverage the unique
perception and cognitive reasoning together, allowing advantages of each [2]. Neural networks, renowned for
us to transform raw information into meaningful their ability to learn complex patterns from raw data,
symbols that fuel our understanding and long-term provide NSAI with the capacity for sophisticated
planning. In the realm of Artificial Intelligence, we have perception and feature extraction. Subsequently, these
made significant strides with machine perception extracted features are channeled into symbolic systems,
through advanced neural networks, which excel at which excel at logical inference, knowledge representation,
recognizing patterns within vast datasets. However, the and reasoning. The primary objective of NSAI is to create
quest for machine cognition—capable of complex AI systems that can effectively bridge the divide between
reasoning, analogy, and strategic planning—remains a low-level perception and high-level cognition.[14][6][18]
formidable challenge. Unlike machines, humans By extracting meaningful features from data using deep
inherently manage and articulate their cognitive learning techniques, NSAI enables the representation of
processes, relying on well-defined symbolic connections complex information in a symbolic format. This symbolic
between sensory inputs and their accumulated representation facilitates the manipulation and reasoning
knowledge. This clarity is essential in high-stakes areas processes, allowing the AI system to make informed
such as healthcare, criminal justice, and autonomous decisions and generate explanations for its actions [4]. This
driving, where the ability to trace decision-making capability is particularly crucial in domains where
pathways is crucial for safety and accountability. While transparency, accountability, and reliability are paramount,
neural networks shine in perception, symbolic AI such as healthcare, finance, and autonomous systems.
emerges as a champion of cognition, delivering explicit Furthermore, NSAI holds the promise of developing AI
representations that enhance transparency and systems that can learn and reason in a manner more akin to
regulatory compliance. This research paper delves into human cognition. By combining the data-driven learning of
the innovative realm of Neuro-Symbolic AI, a cutting- neural networks with the knowledge-driven reasoning of
edge paradigm that harmoniously combines neural symbolic AI, NSAI aims to create systems that are not only
networks with symbolic reasoning. By investigating the capable of processing vast amounts of information but also
current landscape of Neuro-Symbolic AI applications, of understanding and interpreting it in a meaningful way.
this study aims to illuminate its potential for crafting AI This represents a significant step towards building AI that
systems that are not only more robust but also more can truly reason, learn, and adapt in complex and dynamic
transparent and explainable, ultimately steering us environments [4][13][15].
toward a safer and more trustworthy future in AI.
II. UNDERSTANDING NEURO-SYMBOLIC AI: BRIDGING
Keywords— Artificial Intelligence (AI), Neuro-symbolic NEURAL NETWORKS AND SYMBOLIC REASONING
AI, Symbolic AI, Neural Networks, Reasoning and Neuro-symbolic AI (NSAI) is a field of artificial
Problem-Solving. intelligence that combines traditional, rules-based AI
approaches with modern deep learning techniques [2]. It
I. OVERVIEW seeks to integrate neural network-based methods with
Neuro-symbolic artificial Intelligence (NSAI) represents a symbolic, knowledge-based approaches to create more
paradigm shift in AI research, strategically bridging the gap capable and flexible AI systems [3]. NSAI aims to leverage
between traditional, rule-based systems and contemporary the strengths of both neural (sub-symbolic) and symbolic
deep learning methodologies [1][3]. In essence, NSAI AI, addressing the limitations of relying on either approach
endeavors to harmonize the strengths of both worlds: the alone.
robust pattern recognition capabilities of neural networks A. Examples of Neuro-Symbolic AI Models:
with the logical reasoning and knowledge representation of
symbolic AI [3]. This fusion aims to overcome the inherent 1) Neuro-symbolic Concept Learner: This model
limitations of each approach, leading to the development of combines symbolic AI with deep learning by parsing input
AI systems that are not only powerful but also interpretable questions and scenes into symbolic programs [1][16].
and adaptable. At its core, NSAI is characterized by the 2) Visual Concept-Metaconcept Learning (VCML):
deliberate integration of neural network-based methods and This model uses embeddings to learn object properties with
symbolic knowledge-based approaches. This integration is less supervision [3][19].
not merely a superficial combination; rather, it is a deep,
 3) Neural Logic Machine (NLM): Applies both
inductive learning and logic reasoning to an object’s
relations, properties, quantifiers, and logic connectives to
accomplish the task of generalisation [3].
4) DeepProbLog: Combines neural networks with
probabilistic logic. It retains all the essential components of
the probabilistic logic programming language: the
semantics, the inference mechanism, and the
implementation [1][16].

a) Symbolic AI: Symbolic AI, also known as Good fig. [1]. Symbolic AI (uses symbols and formal logic to
Old-Fashioned Artificial Intelligence (GOFAI), is an represent knowledge, reason, and solve problems)
approach to artificial intelligence that uses symbols and b) Neural Network: A neural network, or Artificial
rules to represent knowledge and perform reasoning [4]. It Neural Network (ANN), is a machine learning model
was the dominant paradigm in AI research from the 1950s inspired by the structure and functionality of the human
to the 1990s before machine learning and neural networks brain. It consists of interconnected nodes (artificial
gained prominence. neurons) organized in layers, which process and transmit
data to solve complex problems [3].

fig. [2]. Neural Network (It simulates how the human brain processes information, enabling machines to perform tasks like classification,
prediction, and decision-making with high accuracy).

c) Neuro Symbolic AI: Neuro-symbolic AI is a type

of artificial intelligence that integrates neural networks
(which excel at pattern recognition) and symbolic AI
(which is based on rule-based reasoning) [5]. The goal is to
combine the strengths of both approaches to create AI
systems that can reason, learn, and model cognitive
processes effectively. Neuro-symbolic AI systems aim to fig. [1]. Neuro-symbolic AI combines the strengths of
provide more accurate, explainable, and adaptable AI. neural networks (deep learning) with the precision of
They have potential applications in various fields, symbolic reasoning.
including healthcare, finance, autonomous systems, and
robotics [8][14].
III. THE IMPORTANCE OF NEURO-SYMBOLIC AI
Neuro-symbolic AI (NSAI) is used because it combines the
strengths of neural networks and symbolic AI, leading to
improvements at both the algorithmic and application
levels [2][19].
A. Algorithmic-level considerations:
1) Mimicking human intelligence: NSAI seeks to
replicate the way the human brain processes information by
integrating perception and cognition. It combines "System a) Decoupled Integration: Employs decoupled
1" (fast, intuitive processing, like neural networks) and integration between the neural and symbolic components
"System 2" (slow, deliberate reasoning, like symbolic AI) [2].
[2]. b) Intertwined Integration: Employs intertwined
2) Enables complex cognitive processes: NSAI integration between the neural and symbolic components
supports abstraction, analogy, and long-term planning [2]. [2].
3) Overcoming limitations of neural networks: While
neural networks excel at pattern recognition and processing V. ADVANTAGES OF NEURO-SYMBOLIC AI
raw data, they often cannot explicitly represent background The benefits of using Neuro-Symbolic AI (NSAI) stem
knowledge and struggle with reliable evaluation of from its ability to integrate neural networks with symbolic
cognitive capabilities [2]. AI, enhancing AI systems at both algorithmic and
application levels [8].
IV. UNDERSTANDING THE MECHANICS OF NEURO-
SYMBOLIC AI A. General Benefits:
Neuro-symbolic AI (NSAI) works by integrating neural 1) Superior Performance: NSAI models have shown
network-based methods with symbolic knowledge- the capacity to outperform state-of-the-art deep learning
based approaches [3][4]. It aims to blend the powerful models in areas such as image and video reasoning [8].
approximation abilities of neural networks with the 2) Reduced Data Needs: NSAI models can attain high
symbolic reasoning capabilities, enabling reasoning about accuracy even with significantly less training data
abstract concepts, extrapolation from limited data, and compared to traditional models. Some models achieve
generation of explainable results [4]. good accuracy with as little as 10% of the training data
A. Here's a breakdown of how it generally works: needed for purely deep learning models [6].
3) Addresses Generalisation: NSAI can tackle the
1) Feature Extraction with Deep Learning: Utilises
issue of generalisation in deep neural network systems. It
deep learning to extract complex features from data. This
can also promote compositional generalisation [7].
involves training neural networks on large datasets to
4) Improved problem-solving: It provides advanced
recognise patterns and convert raw data into meaningful
problem-solving and decision-making capabilities.
symbols [4].
5) Enhanced Learning: It enhances learning by fusing
2) Symbolic Manipulation: Manipulates the extracted
symbolic reasoning with neural learning mechanisms,
features using symbolic AI techniques, which are good for
adapting common-sense knowledge for few-shot settings
formalising human-like reasoning. This involves using
and transforming observations into logical facts [8].
rules, logic, and knowledge representation to perform
reasoning, planning, and decision-making [4]. B. Algorithmic Level Benefits:
3) Knowledge Representation: Integrates symbolic and 1) Mimicking human intelligence: NSAI replicates
neural representations, developing commonsense and how the human brain processes information by integrating
domain-specific knowledge graphs [4]. perception and cognition. It combines "System 1" (fast,
4) Learning and Inference: Combines learning and intuitive processing, like neural networks) and "System 2"
reasoning processes through end-to-end differentiable (slow, deliberate reasoning, like symbolic AI).
reasoning and dynamic multi-source knowledge reasoning 2) Enables complex cognitive processes: It supports
[5]. abstraction, analogy, and long-term planning.
5) Integration of Reasoning Methods: This method 3) Overcoming limitations of neural networks: While
Integrates neural networks with logic and probability [5]. neural networks excel at pattern recognition and processing
B. NSAI methods can be classified into two main raw data, they often cannot explicitly represent background
categories: knowledge and struggle with reliable evaluation of
cognitive capabilities.
1) Methods that compress structured symbolic
knowledge to integrate with neural patterns and reason C. Application Level Benefits:
using the integrated neural patterns: 1) Enhancing Explainability and Interpretability:
a) Compressed Knowledge Graph Representation- NSAI provides mechanisms for explainability,
Based Methods: These methods utilise compressed interpretability, safety, and trust in AI systems. Symbolic
knowledge graph representations for integration with knowledge structures are used to impose domain
neural patterns [2]. constraints for safety and provide explicit reasoning traces
b) Compressed Logic Representation-Based for explainability, leading to more trustworthy and
Methods: These methods utilise compressed formal logic- dependable AI systems [8].
based representations for integration with neural patterns 2) Addressing Neural Network Limitations: Neural
[4]. networks can be vulnerable to data artefacts, adversarial
2) Methods that extract information from neural attacks, and distribution shifts. NSAI incorporates
patterns to allow for mapping to structured symbolic background knowledge to set guardrails during training.
knowledge (lifting) and perform symbolic reasoning In summary, the benefits of NSAI lie in its ability to blend
the pattern recognition of neural networks with the
reasoning and knowledge representation of symbolic VII. THE FUTURE OF NEURO-SYMBOLIC AI
AI. This combination results in AI systems that are more The future of Neuro-Symbolic AI (NSAI) involves
capable, reliable, and trustworthy, with broad applications advancements on algorithmic and application levels, with
across various domains [8]. considerable potential to enhance AI systems [2].
A. Algorithmic Level Advancements:
VI. TARGETED USES:
1) Combining Language Models and Knowledge
The benefits of using Neuro-Symbolic AI (NSAI) stem Graphs: Current implementations show the effectiveness
from its ability to integrate neural networks with symbolic of combining language models and knowledge graphs.
AI, enhancing AI systems at both algorithmic and Future knowledge graphs could model heterogeneous types
application levels [4]. of application and domain-level knowledge beyond
schemas, including workflows, constraint specifications
1) Drug Discovery and Repurposing: Facilitates and process structures [2].
chemical structure analysis and enhances drug repurposing 2) Enhanced Knowledge Graphs: Knowledge graphs
efforts through relationship prediction between molecules are suitable for symbolic structures that bridge cognition
and targets [8][20]. and perception because they support real-world dynamism
2) Autonomous Vehicles: Improves decision-making and are easy to update, unlike static and brittle symbolic
by translating sensor data into 3D representations for logic. They also have established standards for portability
analysis with symbolic techniques [8][20]. [2].
3) Healthcare: Enhances the accuracy of disease 3) Scalability: Knowledge graphs have a mature
diagnosis and prognosis and improves diagnostic accuracy ecosystem of algorithms that enable efficient graph
in brain-computer interfaces by interpreting neural signals management and querying. This scalability allows for
and applying logical rules for patient-specific treatments modelling large and complex datasets [2].
[8]. 4) Meta-Cognition: Further research is required in the
4) Financial Fraud Detection: Identifies subtle fraud area of Meta-Cognition in Neuro-Symbolic AI, which is the
patterns using neural networks and uses them to train system’s capacity to monitor, evaluate, and adjust its own
symbolic algorithms [8][20]. reasoning and learning processes. Neglecting meta-
5) Recommender Systems: Structures domain cognition in NSAI research limits system autonomy,
knowledge to tune neural networks for more accurate adaptability, and reliability, hindering error correction and
recommendations and learns interpretable recommendation reducing trustworthiness in dynamic environments[5].
rules based on user and item attributes [8][20].
6) Manufacturing and Quality Control: It forecasts B. Application Level Developments:
machine failures using sensor data and predicts process 1) User-Level Utility: Combining enhanced knowledge
instability based on historical trends. It adjusts order lists in graphs with high-capacity neural networks would provide
real-time, considering machine efficiency and potential the end-user with a high degree of algorithmic and
delays for production planning optimisation [8][20]. application-level utility [4].
7) Education: NSAI combines neural networks to 2) Safety and Trustworthiness: Encoding guidelines,
analyse student performance with symbolic AI for policy, and regulations via extended forms of knowledge
personalised educational recommendations in learning graphs can provide explainability, accountability, rigorous
platforms [8][20]. auditing capabilities, and safety [4][13].
8) Regulatory Compliance Monitoring: It combines 3) Specific Domains:
neural networks with symbolic AI to detect violations and a) Healthcare: A key challenge in the health domain
ensure compliance with complex regulations in industries is to build transparent artificial intelligence. Neuro-
like finance and healthcare [8]. symbolic algorithms are being used in Drug Repurposing
9) Cyber Security: A genetic algorithm is used to find (DR), which involves finding a new indication for an
combinatorial optimisation of logic-programmed existing drug [8].
constraints and deep learning from given rule-based and b) Recommender Systems: Neuro-Symbolic
neural components [8]. Interpretable Collaborative Filtering learns interpretable
10) Other Applications: NSAI is also used in image recommendation rules based on user and item attributes
processing, business management, brain modelling, using neural networks [8].
information retrieval, language generation, dialogue
systems, education systems, robotics, smart cities, and safe C. Overcoming Current Limitations:
machine learning [8]. • Scalable Frameworks: Development of scalable
frameworks that integrate traditional logic
NSAI is particularly useful in scenarios where training data programming with neural networks for complex
is limited or when training time is prohibitive. It offers reasoning tasks [5].
advantages in terms of requiring less data to train and • Commonsense Reasoning: The Incorporation of
demonstrating common sense. commonsense knowledge and advanced language
understanding to enhance multi-hop reasoning
capabilities [5].
 • Trustworthy Decision-Making: Combining [5]. Brandon C. Colelough, William Regli, “Neuro-Symbolic AI in
2024: A Systematic Review [arXiv:2501.05435v1 [cs.AI] 9 Jan
symbolic logic with neural networks to ensure 2025]” published.
reliable and trustworthy decision-making [5]. [6]. A. d. Garcez and L. C. Lamb, “Neurosymbolic AI: The 3rd wave,”
• Adaptive Symbolic Representations: Adapting Artificial Intelligence Review, pp. 1–20, 2023 published.
and evolving symbolic representations in real- [7]. D. Bouneffouf, R. Feraud, S. Upadhyay, M. Agarwal, Y.
Khazaeni, and I. Rish, “Toward skills dialog orchestration with
time to maintain transparency [5]. online learning,” in ICASSP 2021-2021 IEEE International
• Incremental Learning: Developing incremental Conference on Acoustics, Speech and Signal Processing
learning that allows symbolic systems to evolve (ICASSP), pp. 3600– 3604, IEEE, 2021, published.
with new experiences [5]. [8]. Djallel Bouneffouf and Charu C. Aggarwal “Survey on
Applications of Neurosymbolic Artificial
• Context-Awareness: Creating context-aware Intelligence[arXiv:2209.12618v1 [cs.AI] 8 Sep 2022]” published.
inference mechanisms that adjust reasoning based [9]. Arnab Kumar, Punit Shukla, Aalekh Sharan and Tanay Mahindru,
on situational cues [5]. “National Strategy for Artificial Intelligence” published.
[10]. Dr. Munish Sabharwal, “The use of Artificial Intelligence (AI)
• Fine-Grained Explainability: Achieving fine- based technological applications by Indian Banks”
grained explainability for complex inference [11]. Zachary Susskind, Bryce Arden, Lizy K. John, Patrick Stockton
chains [5]. and Eugene B. John,“Neuro-Symbolic AI: An Emerging Class of
AI Workloads and their Characterization [arXiv:2109.06133v1]
D. Obstacles and Considerations: ” published.
1) Addressing Safety Concerns: There is a concern that [12]. “IBM Research” [Online]. Available:
current systems could significantly harm individuals and https://research.ibm.com/topics/neuro-symbolic-ai
society without additional guardrails. [13]. “Wikipedia.org” [Online]. Available:
https://en.wikipedia.org/wiki/Neuro-symbolic_AI
2) Explainability and Trustworthiness: Significant [14]. “Techtarget.com” [Online]. Available :
gaps remain in explainability and trustworthiness. https://www.techtarget.com/searchenterpriseai/definition/neuro-
symbolic-AI.
Overall, the future of NSAI is geared towards creating [15]. Artur d'Avila Garcez, Luis C. Lamb, "Neurosymbolic AI: the 3rd
wave", 2023.
AI systems that are more intelligent, reliable, and [16]. Md Kamruzzaman Sarker, Lu Zhou, Aaron Eberhart, Pascal
context-aware by integrating neural networks with Hitzler, "Neuro-symbolic artificial intelligence: a survey", 2024.
knowledge-guided symbolic approaches. This involves [17]. Zachary Susskind, et al., "Neuro-Symbolic AI: An Emerging
enhancing both the algorithmic capabilities and Class of AI Workloads and their Characterization", 2021.
[18]. Xin Zhang, Victor S. Sheng, "Bridging the Gap: Representation
application-level features of AI systems [5][8]. Spaces in Neuro-Symbolic AI", 2024.
[19]. Jonathan Feldstein, Paulius Dilkas, Vaishak Belle, Efthymia
Tsamoura, "Mapping the Neuro-Symbolic AI Landscape by
VIII. CONCLUSION Architectures: A Handbook on Augmenting Deep Learning
Through Symbolic Reasoning", 2024.
Neuro-symbolic AI (NSAI) integrates rules-based AI with [20]. Oualid Bougzime, Samir Jabbar, Christophe Cruz, Frédéric
deep learning, outperforming traditional models in areas Demoly, "Unlocking the Potential of Generative AI through
like reasoning and achieving high accuracy with less data. Neuro-Symbolic Architectures: Benefits and Limitations", 2025
While combining statistical and symbolic AI strengths,
NSAI's performance characteristics resemble existing
workloads, with symbolic components unlikely to
bottleneck execution. Research since 2020 highlights
advancements in learning, inference, logic, and knowledge
representation but reveals gaps in explainability,
trustworthiness, and meta-cognition, which need
addressing for reliable, context-aware AI systems. The
field should focus on integrating explainability,
trustworthiness, and meta-cognition to enhance AI
systems' reliability, adaptability, and safety, especially
in critical applications.

REFERENCES
[1]. Luc De Raedt, Robin Manhaeve, Sebastijan Dumanciˇ c´, Thomas
Demeester2 and Angelika Kimmig “Neuro-Symbolic = Neural +
Logical + Probabilistic” published.
[2]. Amit Sheth, Kaushik Roy, and Manas Gaur “Neurosymbolic AI -
Why, What, and How [arXiv:2305.00813v1 [cs.AI] 1 May 2023]”
published.
[3]. Zachary Susskind, Bryce Arden, Lizy K. John, Patrick Stockton
and Eugene B. John “Neuro-Symbolic AI: An Emerging Class of
AI Workloads and their Characterization [arXiv:2109.06133v1
[cs.AI] 13 Sep 2021]” published.
[4]. Rabinandan Kishor “Neuro-Symbolic AI: Bringing a new era of
Machine Learning [International Journal of Research Publication
and Reviews, Vol 3, no 12, pp 2326-2336, December 2022]”
published.