An optimal parameter analysis and GPU acceleration of the image receptive fields neural network approach
Alexander Woodward, Patrice Jean Delmas, Takashi Ikegami · 2012
The Image Receptive Fields Neural Networks (IRFNN) algorithm is a recent approach for image classification that is as accurate and an order of magnitude faster than using a traditional feed-forward neural network (multi-layer perceptron), with a linear input layer, non-linear hidden layer and linear output layer. This paper investigates the algorithm's optimal parameter configuration along with a GPU implementation, further extending the performance of the algorithm. Optimization of classification is achieved through a deep search of potential configurations with respect to the number of neurons in the hidden layer and receptive field placement within the image plane. Second stage refinement is achieved through a search for optimal Gaussian receptive field size and shape in 2D. These processes guarantee an optimal network configuration. Secondly, a GPU acceleration of the feed-forward processing of images into the network is implemented. Receptive fields are uploaded to the GPU and all computations take place on the GPU resulting in a large performance increase. Analysis of both improvements are described in the paper.