Connectionist Structures Synthesis Acceleration on the basis of CUDA Parallel Computing Architecture

The paper scrutinizes the method of artificial neural networks on the basis of radial-basic functions using CUDA architecture which considerably allows accelerating calculations during parameter optimization of connectionist structures. Most resource-intensive calculating operations in the process of neural network synthesis are singled out. The algorithm of quick calculations of neural network output while analyzing the value variants of its parameters is proposed. The algorithm is based on the storage of interim calculations, and is characterized by the calculation speed independence from the dimension of data input vector. The software implementation of the examined quick calculations algorithm on the basis of graphic accelerators is described. The comparative analysis of calculation efficiency for various variants of synthesis algorithm implementation, including the use of CPU and GPU, is presented. The research shows that the efficiency of the examined algorithm grows at the increase of the training sample size. The paper presents the results achieved during the calculating experiment for various amount of neurons and sample elements. These results prove the efficiency of the described synthesis method. The proposed algorithm can be used in the systems that require the synthesis of connectionist models in the real time mode.