AI summary

This paper introduces a CUDA-based high-performance library for training and classification of CNNs on the GPU, demonstrating performance and scalability improvements compared to CPU implementations on the MNIST dataset. The GPU version achieves speedups of 2 to 24 times compared to the CPU, with better scaling for larger networks.

Abstract

In this paper we present the implementation of a framework for accelerating training and classification of arbitrary Convolutional Neural Networks (CNNs) on the GPU. CNNs are a derivative of standard Multilayer Perceptron (MLP) neural networks optimized for two-dimensional pattern recognition problems such as Optical Character Recognition (OCR) or face detection. We describe the basic parts of a CNN and demonstrate the performance and scalability improvement that can be achieved by shifting the computation-intensive tasks of a CNN to the GPU. Depending on the network topology training and classification on the GPU performs 2 to 24 times faster than on the CPU. Furthermore, the GPU version scales much better than the CPU implementation with respect to the network size.