Simulation of artificial neural networks using general purpose graphics processing unit

Abstract

The relevance of the discussed issue is caused by the high computational load generating by an artificial neural network simulation, while the latter is the most successful solution for several AI tasks. In most cases, the high computational load of artificial neural network simulation causes a decline of its functionality and restricts its applicability. The main aim of the study is to improve the efficiency of resolving the AI tasks using artificial neural networks by improving simulation performance applying parallel computations on general purpose graphics processing unit. The methods used in the study. The theoretical researches were carried out using concurrency theory, graph theory, vector algebra and methods of systems analysis. During the experimental study the authors tested an image analysis system software complex that uses the proposed approaches. The results. The authors proposed an approach to simulate the variety of artificial neural networks with high degree of parallelism, which is based on specific precomputation of the groups of compute-time parallel connections between neurons. This group defines explicitly what parts of overall computational task can be performed in parallel. The approach allows transferring as well a computational load to graphics processing unit and performing a batch processing on central processing unit. The achieved performance speed-up ratio reaches the ratio of GPU peak theoretical performance to that of CPU indicating the high efficiency of the proposed approach.

Keywords

• artificial neural networks
• massive parallelism
• general purpose graphics processing unit
• CUDA
• image recognition