Contemporary artificial intelligence systems, particularly large language models and deep neural networks, achieve remarkable performance through statistical pattern recognition trained on massive datasets often exceeding trillions of tokens. However, this approach fundamentally diverges from human cognition, which demonstrates extraordinary capacity for generalization from minimal examples through understanding of relational structures and causal principles. This paper examines the theoretical foundations and empirical implementations of relational learning paradigms that prioritize structural knowledge over volumetric data, potentially reducing training data requirements by several orders of magnitude. We analyze three complementary approaches:(1) neuro-symbolic architectures that integrate logical reasoning with neural computation,(2) causal inference frameworks that distinguish correlation from causation, and(3) meta-learning systems that acquire transferable learning strategies across task distributions. Drawing on recent advances in differentiable logic programming, structural causal models, and few-shot learning, we demonstrate that AI systems trained to understand relationships, principles, and abstractions—analogous to how mathematics education provides tools for infinite problem-solving rather than memorization of solutions—can achieve superior generalization with dramatically reduced data dependencies. The findings suggest a fundamental paradigm shift from data-centric to knowledge-centric AI development, with implications for model efficiency, interpretability, and alignment with human cognitive architectures.
Revista et al. (Thu,) studied this question.