Neural Network Simulation The recognition application

This paper presents the GPU mapping of the recognition algo-rithm of a Convolution Neural Network (CNN). This work is based on a C-implementation of the application. The mapping to GPU was performed through different approaches which are explained in detail. The improvements achieved by each ap-proach are presented as well as the overall speed up of the appli-cation. The GPU implementation demonstrates the feasibility to integrate the algorithm in a real-time system as the frame rate accomplished is 30fps. In order to examine the scalability of the application, the use of multiples GPUs was explored. The CPU implementation is executed on a 800MHz AMD Phenom. The GPU platform used for the experiment is an Nvidia GeForce GTX570. 1