A Highly Parallel Multi-class Pattern Classification on GPU

Multi-class pattern classification has a variety of applications and could be achieved using artificial neural networks (ANN). There are two major system architectures for using ANNs in multi-class pattern classification: using a single ANN and using multiple ANNs. Independent of what architecture is used, one of the main concerns of using ANNs is that with increasing number of pattern classes and training datasets, the training time will increase dramatically which renders the ANN unfeasible. In this paper, the vast computational power of Graphics Processing Units (GPUs) is utilized to mitigate this problem. Different architectures and different methods of feeding pattern classes are implemented in a GPU platform. Different methods have been proposed to achieve maximum parallelism and subsequently maximize throughput. Our implementation exceeds the state-of-the-art in literature in terms of speed and the accurate use of GPU resources. As a result, the proposed approach´s run time is about 75% shorter than the previous approaches. In multi-ANN architecture, due to the inherent parallelism in the proposed implementation, the execution time of a system for a digit recognition application is reduced from seven hours in CPU to about 4 seconds in GPU.