Caracrachá: investigando eficiência e eficácia de uma rede neural portada para ponto fixo em ambiente embarcado

Abstract

This work aims to present the viability of creating and training a neural network for face recognition in floating point on a PC environment and later to convert it into a fixed point neural network and port it to an embedded system environment with the same behavior. It was used a neural network structured in a Deep Feedforward architecture composed of three hidden layers and an input layer being fed by grayscale images of faces at a 48x48px resolution, where each pixel communicates with each neuron of the first layer. The question of behavior, performance, consistency and efficiency of this neural network after its conversion from floating point to fixed point was also addressed. In addition, it was also verified the degree to which these network characteristics remain after porting the network to an embedded system and which embedded systems have been able to use it. The training was performed on a regular PC and the performance and viability tests were done on a Raspberry Pi. Gains of 50% were observed in the processing time of neural networks when using the proposed technique without considerable loss of reliability.