Dynamic Neural Networks for Adaptive Edge AI: Techniques, Tools, and Development Directions
Year: 2025
Type: conference-paper
Abstract: Edge artificial intelligence (Edge AI) is increasingly vital in smart surveillance, industrial IoT, and autonomous systems, where real-time processing and energy efficiency are critical. Traditional deep neural networks are accurate but static and resource-hungry, limiting deployment on constrained devices. Dynamic neural networks enable input-adaptive computation, adjusting workload at runtime to context, input difficulty, and hardware limits. This survey synthesizes methods that support dynamic adaptability alongside offline compression (e.g., pruning, quantization) and runtime strategies (e.g., early exiting, dynamic resolution control). We review supporting software frameworks, including Slimmable Networks, Once-for-All (OFA), and PyTorch-based modules for adaptive execution. Finally, we outline a research agenda to co-optimize pre-deployment compression with runtime adaptation, targeting accuracy, latency, energy trade offs for robust, real-world edge deployment.
Authors: Tamara Milovanović, Bratislav Predić, Davide Bertozzi, Zoran Perić, Sofija Perić, Jelena Nikolić
Institutions: University of Nis, University of Manchester
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4416923872