Computer Science · Artificial Intelligence
GPGPU based concurrent classification using trained model of handwritten digits
Syed Tahir Hussain Rizvi, Gianpiero Cabodi, Arslan Arif, Muhammad Yaqoob Javed, Muhammad Majid Gulzar
See on the mapAbstract
In this paper, General Purpose Graphical Processing Unit (GPGPU) based concurrent implementation of handwritten digit classifier is presented. Different styles of handwriting make it difficult to recognize a pattern but using neural network, it is not a difficult task to perform. Different softwares like torch and MATLAB provide the support of multiple training algorithms to train a network. By choosing an appropriate training algorithm for a specific application, speed of training can be increased. Furthermore, using computational power of GPUs, training and classification speed of neural network can be significantly improved. In this work, Modified National Institute of Standards and Technology (MNIST) database of handwritten digits is used to train the network. Accuracy and training time of digit classifier is evaluated for different algorithms and then concurrent training is performed by exploiting power of GPU. Trained parameters are imported and used for the concurrent classification with Compute Unified Device Architecture (CUDA) computing language which can be useful in numerous practical applications. Finally, the results of sequential and concurrent operations of training and classification are compared.