Improved Architecture of the Feedforward Neural Network for Image Recognition

Recently, the researchers have been focusing on the convolutional neural network due to its high reliability in image recognition. It is proposed that the feedforward neural network could compete equivalently with the convolutional neural network? In this paper, we have explored the possibility and proposed a feedforward neural network, namely the scaled conjugate gradient backpropagation feedforward neural network with random connections (SCGBP-FNN-RC) to learn big data through recognizing images from the widely known MNIST dataset which is applied with affine and elastic distortions. Based on our findings, SCGBP-FNN-RC has managed to achieve a state-of-the-art accuracy of 99.54 %. The proposed networks are then evaluated based on various parameters, e.g. the neuron size, connections and the K parameter.