Methods of Interpretability of Deep Neural Networks in Decision-Making Tasks

Abstract

Interpretability is the main obstacle to deploying deep neural networks in high‑stakes settings. Despite high accuracy, regulators and users require explanations they can understand. This review summarizes evidence on which explanation methods truly support decision‑making. A PRISMA filter selected fifteen peer‑reviewed experiments from 2017 to 2025. Saliency, attribution, counterfactual, and inherently transparent techniques were reorganized. When publicly available, effect sizes from confusion matrices were recalculated; narrative synthesis filled in missing details. No new data were collected. Results were compared based on fidelity, stability, cognitive load, and privacy tolerance. Three consistent signals appeared. Gradient‑based visualizations excelled on medical images, while additive neural models led in credit‑risk scenarios, confirming task–method matching. Explanation stability dropped about 25% when differential privacy budgets fell below ε = 3, even though accuracy remained stable. The median fidelity gap for the top method per task was 2.3%. User‑trust scores increased when brief counterfactual explanations accompanied saliency heatmaps; combining these consistently improved audit‑ability. The synthesis creates a “Context–Layer–Fit” matrix that treats interpretability as a design requirement instead of an afterthought. Engineers and policymakers can use it to choose transparent pipelines that balance accuracy, privacy, and cognitive effort. Future research should test robustness under data shifts, establish combined quality–latency benchmarks, and promote open‑source explanation tools