Fast-DENSER: Fast Deep Evolutionary Network Structured Representation

• Autor(a) principal: Assunção, Filipe
• Data de Publicação: 2021
• Outros Autores: Lourenço, Nuno, Ribeiro, Bernardete, Machado, Penousal
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Texto Completo: https://hdl.handle.net/10316/100856; https://doi.org/10.1016/j.softx.2021.100694
• topic: Artificial Neural Networks; Automated machine learning; NeuroEvolution

This paper introduces a grammar-based general purpose framework for the automatic search and deployment of potentially Deep Artificial Neural Networks (DANNs). The approach is known as Fast Deep Evolutionary Network Structured Representation (Fast-DENSER) and is capable of simultaneously optimising the topology, learning strategy and any other required hyper-parameters (e.g., data pre-processing or augmentation). Fast-DENSER has been successfully applied to numerous object recognition tasks, with the generation of Convolutional Neural Networks (CNNs). The code is developed and tested in Python3, and made available as a library. A simple and easy to follow example is described for the automatic search of CNNs for the Fashion-MNIST benchmark