More Faster Self-Organizing Maps by General Purpose on Graphics Processing Units

Abstract

Recently, GPUs have been attempted to process for general computation, it is called “General Purpose Graphic on Processing Units(GPGPU)” and is the focus of attention. The feature of GPUs is that small and limited capable processors are parallelized on a massive scale, and those processors operate synchronously. From this feature, GPGPU have capability of acceleration of the specific use computation. The capability of GPGPU is specifically suitable for acceleration of simply calculation of loop iteration because processing is in parallel.

In this paper, we adopt GPGPU to accelerate learning of Self- Organizing Maps(SOM) because there are many loops which are simple calculation in learning algorithm of SOM. Some approaches which are acceleration of learning of SOM by GPGPU have been proposed. In contrast to those proposals, our proposals are that some graphic processors calculate concertedly using “NVIDIA Scalable Link Interface technology(NVIDIA SLI).” In the experiments, we compare the learning speed and recognition rate of character recognition using SOMs which are calculated by MPU, a GPU(CUDA-SOM) and multi-GPU(SLI-SOM). As a result of experiments, the recognition rates for test patterns are almost identical in 3 approaches of SOM. In comparison of speed of learning, CUDA-SOM is about 7 times as fast as that in MPU in case of small size of competitive layer, and SLI-SOM about 80 times faster than MPU in large size of competitive layer. We succeeded in accelerating the learning performance from those results.