TRAINING OF A DEEP NEURAL NETWORK USING CUDA PROGRAMMING

• DOI: https://doi.org/10.22533/at.ed.3174292419121

• Palavras-chave: CUDA programming, graphics processing unit, training phase, neural network, spatial convolution, algorithm.

• Keywords: CUDA programming, graphics processing unit, training phase, neural network, spatial convolution, algorithm.

• Abstract:

  Deep neural networks have been successfully applied in the fields of computer vision, automatic image recognition and speech, among others. An important part of their architecture is the use of convolution operations that perform feature filtering at different levels of abstraction during the network training phase. This paper proposes the use of GPU graphics acceleration units to reduce the computational load based on its SIMT (Single Instruction Multiple Thread) architecture that exploits the intrinsic data parallelism of these applications.