- The paper systematically reviews Neuro-Symbolic AI research from 2020 to 2024 using the PRISMA methodology.
- It categorizes the field into five areas—knowledge representation, learning and inference, logic and reasoning, explainability, and meta-cognition—highlighting strengths and limitations.
- Findings reveal major progress in learning and inference while exposing gaps in explainability and meta-cognition to guide future research.
The paper "Neuro-Symbolic AI in 2024: A Systematic Review" presents a systematic literature review of Neuro-Symbolic AI, focusing on developments, methodologies, and applications within the 2020-2024 timeframe. The review employs the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) methodology, using databases such as IEEE Explore, Google Scholar, arXiv, ACM, and SpringerLink. The inclusion criteria target peer-reviewed papers published between 2020 and 2024, with a focus on papers with publicly available codebases to ensure reproducibility.
The paper defines Meta-Cognition within the context of Neuro-Symbolic AI and identifies key themes and gaps in the literature post the 2020 Neuro-Symbolic research surge. The authors take the stance that symbolic AI is essential and that Neuro-Symbolic AI represents the best way forward for the community, arguing against the idea that common sense reasoning can be addressed through big data alone.
The authors define Symbolic AI as a sub-field concerned with learning internal symbolic representations of the world, translating implicit human knowledge into formalized rules and logic. Examples include SHRDLU, ELIZA, DENDRAL, MYCIN, ConceptNet 5.5, CYC, and GOFAI (Good Old Fashioned AI) planning systems. Sub-Symbolic AI, conversely, encompasses systems that learn implicit data representations without requiring explicit rules or symbolic representations as inputs, including machine learning, deep learning, and generative AI. Examples include the Perceptron, Hopfield Networks, Backpropagation Algorithm, GPT (Generative Pre-trained Transformer) models, the YOLO family of CNNs (Convolutional Neural Networks), and the DALLE diffusion model transformer. Neuro-Symbolic AI combines Symbolic AI and Sub-Symbolic AI to create systems that leverage the strengths of both, drawing on Kahneman's concepts of System 1 and System 2 thinking.
The authors establish a taxonomy of five foundational research areas in Neuro-Symbolic AI, synthesized from a review of survey papers and seminal books:
- Knowledge Representation: Integrating symbolic and neural representations, and developing commonsense and domain-specific knowledge graphs.
- Learning and Inference: Combining learning and reasoning through end-to-end differentiable reasoning and dynamic multi-source knowledge reasoning.
- Explainability and Trustworthiness: Creating interpretable models and reasoning processes for trust and reliability.
- Logic and Reasoning: Integrating logic-based methods with neural networks, including logical and probabilistic reasoning, and the syntax and semantics of Neuro-Symbolic systems.
- Meta-Cognition: System's capacity to monitor, evaluate, and adjust its reasoning and learning processes, integrating neural networks and symbolic representations.
Meta-Cognition is defined as the processes that involve thinking about one's thinking, enabling self-awareness and self-regulation. The authors argue that neglecting Meta-Cognition limits system autonomy, adaptability, and reliability.
From an initial pool of 1,428 papers, 167 met the inclusion criteria and were analyzed. The majority of research is concentrated in learning and inference (63%), logic and reasoning (35%), and knowledge representation (44%). Explainability and trustworthiness are less represented (28%), and Meta-Cognition is the least explored area (5%). There is significant interdisciplinary potential, particularly in integrating explainability and trustworthiness with other research areas.
The findings indicate a well-integrated body of work in learning and inference, logic and reasoning, and knowledge representation. A notable gap exists in research on explainability and trustworthiness. The sparse representation of Meta-Cognition highlights the need for frameworks that enable AI systems to self-monitor, evaluate, and adjust their processes.
The paper highlights state-of-the-art technologies in each sub-field of Neuro-Symbolic AI since 2020. In Knowledge Representation, research has focused on semantic grounding, representing complex relationships, and improving data efficacy, including the development of commonsense knowledge bases and event-based representations [Mostafazadeh2020, Hwang2021, Ismayilzada2022]. NeuroQL captures complex, long-range relationships [Papoulias2023]. For Learning and Inference, Plan-SOFAI [Fabiano2023] and the ZeroC architecture [Wu2022a] enhance AI planning and zero-shot concept recognition. Semantic Enhancement and Model Trustworthiness were demonstrated by the introduction of a Pseudo-Semantic Loss for autoregressive models [Ahmed2023] and neural networks utilizing Logic Tensor Networks [Badreddine2022]. Research on Explainability and Trustworthiness has looked to advance NLP (Natural Language Processing) Techniques, Enhancing Logical Reasoning, and Refining Language Understanding and Summarization. Braid introduced a logical reasoner with probabilistic rules [Kalyanpur2020]. Structure-Aware Abstractive Conversation improved summarization by incorporating discourse relations [Chen2021]. FactPEGASUS ensures factuality in summarization by optimizing pre-training and fine-tuning methods [Huang2023]. In the research field of Neuro-Symbolic Logic and Reasoning, research has gravitated toward the Integration of Logical Reasoning and Probabilistic Models, Commonsense Knowledge and Language Understanding, and Enhanced Decision-Making.
AlphaGeometry [Trinh2024], a Neuro-Symbolic system for solving Euclidean plane geometry problems, stands out as a project at the intersection of all four research areas and is a groundbreaking example of how Neuro-Symbolic AI can achieve advanced problem-solving capabilities.
In Meta-Cognition, integrating symbolic features with RL (Reinforcement learning) algorithms has been demonstrated through meta-reinforcement learning combined with logical program induction to improve financial trading strategies [Harini2023]. Enhancing general intelligence by fusing cognitive architectures with LLMs has been investigated, creating embodied agents that leverage the strengths of both approaches [Joshi2024, Liu2024]. These projects align with the CMC (Common Model of Cognition), integrating cognitive architectures like ACT-R, Soar, and Sigma [Laird2017].