Fine Tuning Image Input Sets for Deep Neural Networks Using Hallucinations

Abstract

An artificial neural network is a system of software made up of neurons which work based on the neural structure of brain. While training any neural network, it is difficult to understand the features it has learnt at each layer. Deepdream is an algorithm which inverts the neural network using gradient ascent. It finds and enhances patterns of convolutional neural networks in input image based on the training images given to the network. Using this algorithm, we can find the pattern which each layer of a neural network recognizes from the image and enhance it. Often during training, the neural network learns the features which may or may not be the desired features. Having a knowledge of the features recognized in each layer will help us produce better results and increase the efficiency of the neural network. Deepdream has been used for the purpose of entertainment for the hallucinogenic pattern it produced in the input images. To our knowledge this is the first implementation of usage of deepdream in learning the features from each layer of network and thus helping us fine tune the training image set so that the neural network gains better understanding of features from the tuned input set.