Fernandes, Marcelo Augusto CostaGuimarães, Caio José Borba Vilar2020-12-042020-12-042020-10-22GUIMARÃES, Caio José Borba Vilar. Embedded artificial neural networks optimized for low-cost and low-size-memory devices. 2020. 63f. Dissertação (Mestrado em Engenharia Elétrica e de Computação) - Centro de Tecnologia, Universidade Federal do Rio Grande do Norte, Natal, 2020.https://repositorio.ufrn.br/handle/123456789/30829Artificial Neural Networks (ANNs) are bio-inspired systems with a high level of parallelization and almost infinite applications. However, due to the associated high computational power requirements, most application demands powerful processing characteristics and consequently, high-costs and not-so-small form-factors. This work presents an implementation of a Multilayer Perceptron (MLP) for 8-bit microcontrollers in two different scenarios, embedded training, and inference. Analysis of training convergence, inference time duration, and program code occupation into the internal memories and a technique to optimize this implementation to fit bigger MLP architectures. The aim of this work is to provide an overview of the feasibility of ANNs on this low-cost, low-size-memory devices, known as microcontrollers. This work shows a successful implementation of an MLP on a microcontroller with a linear behavior between the increase in hyperparameter values and the time-to-inference and code size. Also, an optimization to include more synaptic weights is presented for this same implementation, showing that even so the same behavior persists, validating further both implementations of the same solution proposal.Acesso AbertoMLPAI8-bitsMicrocontrollerArtificial neural networksEmbedded systemsmasterThesis