A fully connected network topology with Neural Network

Abstract

We propose a novel "Completely Connected Networks" deep network structure to enhance the model's discriminability for small patches within the receptive field (CCN) , The conventional convolutional layer uses linear filters to examine the input followed by a nonlinear activation function  Instead we build more complex miniature neural networks to pool information from the receptive field  The tiny neural network is instantiated as a multilayer perceptron a robust function approximator Micro net- works much like CNN  are slid over the input to produce follow-up feature images, after which transmitted to the next layer for further processing , The architecture allows for the stacking of multiple instances to realise deep CNN , The micro network's enhanced local modelling allows us to employ categorization layer feature map pooling on a world scale which both increases interpretability and decreases the likelihood of overfitting in comparison to more traditional fully connected layers. We demonstrated that CNN yields state-of-the-art categorization results on the CIFAR-10 and CIFAR-100 datasets as well as reasonably good results on the SVHN and MNIST datasets.