Human-Informed Neural Networks (HINNs): Trust-Adaptive Learning with Expert Guidance

Abstract

Artificial intelligence (AI) is increasingly deployed in domains where decisions carry substantial ethical, financial, regulatory, and clinical consequences. Conventional neural networks are commonly trained using fixed data-driven objective functions that may fail to represent the complexity of human judgment, expert knowledge, behavioural preferences, and evolving policy constraints. As a result, such systems may exhibit limited interpretability, reduced adaptability, and insufficient trustworthiness in high-stakes decision environments. This paper introduces the Human-Informed Neural Network (HINNs), a dual-pathway architecture that embeds human supervisory intelligence directly into the optimization process. The first pathway is a deterministic Recurrent Neural Network (RNN) that learns temporal patterns from observed data through standard predictive objectives. The second pathway is a Bayesian Neural Network (BNN) that generates a probabilistic human-aligned supervisory signal calibrated from expert feedback, corrective decisions, or policy-consistent annotations. Both pathways are coupled through a time-sensitive trust function $$\lambda (t)$$ that dynamically regulates their relative influence according to the recency, density, and confidence of available human input, while enabling graceful fallback to autonomous learning when supervision is sparse or unreliable. A complete algorithm is developed, including corrected backpropagation-through-time dynamics for the recurrent pathway, a variational inference objective for the Bayesian pathway, and a joint uncertainty-aware training procedure that propagates semantic supervisory information into sequential hidden-state learning. Validation through both numerical simulations and an empirical case study demonstrates stable optimization trajectories, calibrated trust adaptation dynamics, and robust performance under noisy, drifting, or intermittent human supervision. The proposed framework provides a lightweight and interpretable alternative to computationally intensive generative systems, offering practical relevance for clinical decision support, financial risk assessment, governance analytics, and other policy-sensitive intelligent systems.