<p>Deep learning has shifted from single-model training toward outer-loop optimization, in which many architecture, hyperparameter, robustness, and compression decisions are non-differentiable, mixed-type, or black-box, beyond the scope of standard gradient-based training. Differential evolution (DE), with its scale-sensitive differential mutation and rapidly developing variant family, has become a recurring choice for such decisions. DE applications in deep learning have grown substantially over the past five years, but the literature has not yet been examined through a DE-specific, lifecycle-organized framework. This paper provides such a synthesis. It organizes DE applications across three lifecycle stages (data processing, model design and training, and model evaluation and deployment) and identifies, for each stage, which DE property carries the methodological burden, which variant families fit different evaluation-cost regimes, and how mechanism maturity is distributed across tasks. Direct empirical comparisons reported in the literature position DE against gradient-based methods, Bayesian optimization, reinforcement learning, and other metaheuristics, yielding a task-conditioned rather than universal account of DE’s competitiveness. Coverage extends across computer vision, biomedical applications, time-series forecasting, and industrial engineering. Emerging foundation models (large language, multimodal, diffusion) are discussed as promising directions where sub-problems align with DE’s structural strengths but direct evidence remains thin.</p>