A Pruning Method for Multi‐Layer Perceptron Optimisation in IoT
ABSTRACT The increasing adoption of deep learning in edge computing environments, such as Internet of Things (IoT) devices, demands efficient neural network optimisation techniques to mitigate computational and power constraints. This study presents a novel pruning methodology for multi‐layer percep...
| Autores: | , , , , , , |
|---|---|
| Tipo de recurso: | artículo |
| Fecha de publicación: | 2025 |
| País: | España |
| Institución: | Conselleria de Salut i Consum del Govern de les Illes Balears |
| Repositorio: | Docusalut |
| Idioma: | inglés |
| OAI Identifier: | oai:docusalut.com:20.500.13003/26131 |
| Acceso en línea: | https://hdl.handle.net/20.500.13003/26131 |
| Access Level: | acceso abierto |
| Palabra clave: | Artificial Intelligence Internet of Things Inteligencia Artificial Internet de las Cosas artificial intelligence learning (artificial intelligence) microcontrollers |
| Sumario: | ABSTRACT The increasing adoption of deep learning in edge computing environments, such as Internet of Things (IoT) devices, demands efficient neural network optimisation techniques to mitigate computational and power constraints. This study presents a novel pruning methodology for multi‐layer perceptrons based on auxiliary morphological neural networks (MNNs). These MNNs generate structured pruning masks for each hidden layer, significantly reducing model complexity while preserving performance. The proposed technique is validated using the Modified National Institute of Standards and Technology database (MNIST) and Fashion‐MNIST datasets, demonstrating superior compression rates compared to conventional pruning methods. The experimental results highlight the effectiveness of the method in reducing computational demands, memory footprint and inference latency on IoT platforms while maintaining competitive accuracy levels. These findings underscore the potential of morphological‐based pruning for enabling scalable and energy‐efficient deep learning models in resource‐constrained environments. |
|---|