Neuro-Symbolic Autonomous Learning Architectures for Explainable General Intelligence Systems

Abstract

ABSTRACT Deep neural networks have achieved empirical success across visual perception, natural language processing, and complex decision-making. However, connectionist models remain constrained by opacity, high sample complexity, vulnerability to adversarial perturbations, and an inability to execute formal logical reasoning. Conversely, classical symbolic artificial intelligence delivers transparent, verifiable, and sample-efficient logical inference but suffers from brittleness when handling continuous sensory inputs. Neuro-symbolic autonomous learning architectures have emerged as a foundational paradigm toward Artificial General Intelligence (AGI), unifying connectionist sub-symbolic perception with symbolic reasoning and formal knowledge structures. This review paper presents a systematic synthesis of modern neuro-symbolic paradigms, categorizing neural-to-symbolic, symbolic-over-neural, and deeply integrated dual-process systems. We evaluate mechanisms for neural concept grounding, differentiable logic solver integration, dynamic knowledge graph extraction, and causal reasoning engines. Furthermore, we examine the role of neuro-symbolic integration in providing intrinsic explainability, out-of-distribution safety guarantees, and sample-efficient continual learning without catastrophic forgetting. Through a meta-analysis of standard benchmarks, we highlight empirical performance trade-offs across safety-critical domains including autonomous driving, medical diagnostics, automated robotics, and financial risk assessment. Finally, we establish a research roadmap outlining technical bottlenecks, scalability challenges, and key future directions required to realize trustworthy neuro-symbolic general intelligence.

KEYWORDS: Neuro-Symbolic AI, Artificial General Intelligence (AGI), Explainable AI (XAI), Differentiable Logic, Concept Grounding, Continual Learning, Autonomous Systems.