Triple Pool Net: A novel robust Convolution neural network for image/content classification

Abstract

With the rise of artificial intelligence and machine learning, it is highly desired to find a more efficient neural network architecture for real-life applications. In this paper we propose a novel convolution neural network (CNN) architecture known as triple-pool network (TP-Net), to achieve light-weight training and classification processes. We will firstly provide a comprehensive review on the state-of-the-art, and give a detailed description on the proposed TP-Net. To verify its efficiency, extensive experiments are conducted to compare its performance in terms of training time, error rate (or accuracy), and CPU load in flops, to a number of recently reported CNN architectures, where a well-known publicly available datesets, including CIFAR 10/100, German traffic signs, and SVHN. The network is designed for a convolution neural network(CNN) and can be readily used for image classification or even light weight authentication. We have carefully designed the network to address the gradient vanishing problem that persists in several larger neural network architectures and also addressed the problem of feature loss while reducing dimensions in the pooling layer.