TRAINING AND APPLYING A FEEDFORWARD MULTILAYER NEURAL NETWORK IN GPU

Abstract. This work presents the implementation of Feedforward Multi-Layer Perceptron (FFMLP) Neural Networks on General-purpose computing graphics processing units (GPU) by using CUDA programming language (Compute Unified Device Architecture). The paper presents an introduction to how data is decomposed and processed in a GPU in order to take advantage of its powerful parallel processing capabilities. A feedforward multi-layer perceptron is briefly presented and explained. Since a neural network must be trained in order to solve a particular problem, this work implements a variation of the gradient descent optimization method called backpropagation. The results for both the feedforward implementation and backpropagation training are then compared with their CPU versions