Exploring internal representations of deep neural networks

This paper introduces a method for the generation of images that activate any target neuron or group of neurons of a trained convolutional neural network (CNN). These images are created in such a way that they contain attributes of natural images such as color patterns or textures. The main idea of the method is to pre-train a deep generative network on a dataset of natural images and then use this network to generate images for the target CNN. The analysis of the generated images allows for a better understanding of the CNN internal representations, the detection of otherwise unseen biases, or the creation of explanations through feature localization and description