Abstract

Abstract: Deep Neural Networks (DNNs) show excellent performance incomplex machine learning tasks such as image classification and speechrecognition. However, due to their multilayered non-linear structure, they arenot transparent, that is, they can be classified into certain categories, whichare given new, invisible data samples. To solve this problem the authorproposes an RNN (Recurrent Neural Network) algorithm for classifying thetrained data given by the CNN (Convolution Neural Network) algorithm. Thesystem uses SUN 397 dataset, heatmap method, and de-convolution method. Insystem first, perform heat mapping method on data. Heatmap method extracts theregions of the particular image. Then the de-convolution method is apropagation method for calculating the gradient of a function. LRP (layer-wise relevancepropagation) algorithm learns the features of an image layer by layer from theinput vector of convolution. After that classification performed on thedataset.  Hybrid Network model shows howwe will use CNN-RNN (LSTM) to the learned dataset. So this way get the requireddata from CNN local space while training image dataset of the image and provideit to RNN algorithms part Long Short-Term Memory (LSTM) to provide a predictionabout what NN has learned in CNN model. Finally, the result shows that theproposed RNN algorithm is more accurate than CNN algorithm.

Keywords:Deep Neural Networks, RNN (Recurrent NeuralNetwork) algorithm, CNN (Convolution Neural Network) algorithm, Heat mapmethod, LRP (layer-wise relevance propagation) algorithm.