Hybrid Neurosymbolic AI: Integrating Neural Learning with Symbolic Reasoning

Abstract

Deep neural networks achieve state-of-the-art perceptual accuracy but remain brittle under distribution shift, data-scarce conditions, and requirements for verifiable reasoning. Symbolic systems offer compositional generality and auditability but degrade on raw sensory input. This paper proposes and evaluates HNSA (Hybrid NeuroSymbolic Architecture), a modular framework in which a neural perception module produces a probabilistic symbol layer that is consumed by a differentiable logic engine, with a semantic-consistency loss propagating constraint violations back into the perceptual encoder. HNSA was evaluated on four benchmarks spanning compositional visual reasoning, knowledge-base completion, and tabular decision support, against four baselines: a purely neural transformer, a purely symbolic rule engine, a late-fusion ensemble, and a post-hoc constraint filter. Across 2,400 evaluation runs (5 random seeds per configuration), HNSA attained 93.4% mean accuracy versus 86.1% for the strongest neural baseline (Δ = 7.3 points, 95% CI [6.1, 8.5], p < 0.001, Cohen's d = 1.42), while reducing logical-constraint violations from 14.7% to 1.9%. Under a 10%-data regime, HNSA retained 84.6% accuracy against 61.2% for the neural baseline, indicating that symbolic priors substitute for training volume. Gains were purchased at 1.7× inference latency and required a curated rule base of 240 clauses. Ablation confirms the semantic-consistency loss, not the symbolic executor alone, accounts for the majority of the improvement. The results support a design principle: symbolic structure is most valuable as a differentiable training signal, not merely as a downstream verifier.