REFRAMING DEEP LEARNING OPTIMIZATION AS A STOCHASTIC-ROBUST INDUSTRIAL PROCESS: A BIBLIOMETRIC AND SYSTEMATIC REVIEW OF VARIANCE GOVERNANCE LIFECYCLES

The most commonly used paradigms for deep learning optimization are based on localized summary point-estimates, e.g., validation accuracy, precision and recall, and F₁-score for convolutional neural networks (CNNs). These measures are good indicators of the "central" component of model predictive ability, but are systemically unresponsive to the dispersion of the process and high-frequency variability in its predictive power associated with stochastic initialization noise and the concurrent variability of the hardware. Therefore, models trained with mistaken high accuracy in training/test sets are very susceptible to operational degradation if deployed in a distributed edge setting. This paper tries to fill this gap, offering a comprehensive bibliometric and systematic literature review (SLR) to evaluate the structural integration of quality engineering (QE) principles into the machine learning lifecycle. A bibliometric corpus of 1,142 publications identified from the Scopus database (2015–2026) was clustered and classified using keyword co-occurrence and thematic mapping, indicating that computer science and quality engineering studies have traditionally existed in parallel communities with very little convergence. Moreover, a separate PRISMA-screened SLR corpus of 1,892 raw candidate records was subjected to stringent exclusion criteria. This reduces the sample to just 55 high-fidelity studies (2.9% retained). In essence, this demonstrates that transferring a process capability level to the active training loops using a statistical process capability (Cpk)-level is still extremely rare [19]. Finally, this paper builds on these quantitative empirical gaps to suggest a conceptual synthesis, which is implemented in a structural way for AI development. We translate algorithmic misclassifications into process defect rates, and set Individuals and Moving Range (I-MR) process control limits to offer a structured roadmap to re-optimizing deep learning as a continuous process that is statistically controlled, like other industrial manufacturing processes. [18].