Edge AI in practice: a survey and deployment framework for neural networks on embedded systems

Resumen

The growing demand for intelligent and autonomous devices has accelerated the integration of neural networks into embedded systems, a paradigm known as Edge AI. While this approach enables real-time, low-latency processing with improved privacy, it remains constrained by strict limitations in memory, computation and energy. This paper presents a systematic review, aligned with PRISMA principles, that examines the current landscape of deep learning deployment on embedded hardware. The review analyzes key optimization techniques-including pruning, quantization and inference-level improvements-together with lightweight architectures such as CNNs, RNNs and compact networks, as well as a diverse ecosystem of hardware platforms and software frameworks. From the recurring patterns identified in the literature, we derive a practical five-stage methodology that guides developers through requirement definition, model selection, optimization, hardware alignment and deployment. Unlike existing surveys that mainly provide descriptive taxonomies, this methodology offers a structured and reproducible workflow explicitly designed to support multi-objective trade-offs in resource-constrained environments. The review also identifies emerging trends such as TinyML and hybrid architectures and highlights persistent gaps, including limited support for ultra-low-precision inference, variability in hardware toolchains and the absence of standardized holistic benchmarking. By synthesizing these insights into a coherent framework, this work aims to facilitate more efficient, robust and scalable Edge AI implementations.