Mlp neural network and on-line backpropagation learning implementation in a low-cost fpga
Ernesto Ordoñez-Cardenas, René de Jesús Romero-Troncoso · 2008
This paper presents an implementation of a multilayer perceptronneural network and the backpropagation learning algorithm in an FPGA. The resulting implementation, in contrast to others, is a low-cost system with effective resource utilization, capable of training the neural network for any given task. The system is based on a modular scheme conforming to a system-on-a-chip (SoC), where modules can be replaced or scaled to suit a specific application. The system uses fixed-point arithmetic and it was carried out using generic hardware description language. A pipeline architecture is used in order to build a time-efficient system. The efficacy of the systems was tested in a pattern recognition application, tests were done in a low-cost Xilinx Spartan-3E FPGA.