Interpretation of artificial neural networks for image recognition

This thesis studies the problem of neural networks interpretation in image classification tasks. It is inspired by the publication by professor Michaela Wänke (2015), who was trying to find the answer to the question if people are capable of correctly identifying, whether politician in the photo is Republican or Democrat. The first part covers the methodology needed in image recognition tasks. The second part briefly goes through the main ideas in Wänke,2015, but it also covers deep Taylor decomposition method proposed by Montavon et al., 2017. Deep Taylor decomposition is used in deep learning models (and neural networks in general) interpretation. Data used in the analysis are official photos of American and British politicians, downloaded from www.wikipedia.org with the use of web scraping technologies. In the part of analysis, deep Taylor decomposition was used during neural network optimization and for experiments' results evaluation. The main goal of this thesis is, with the help of neural networks and methodology proposed by Montavon et al., 2017, to define what are the signs, that could possibly give out the political ideology of the person in the photo. The main goal consists of three intermediate goals: review literature on Artificial Neural Networks (ANNs) and the interpretability of ANNs, such as those based on Deep Taylor decomposition; train a classificational neural network for the selected problem; perform the analysis of decision boundaries of the trained ANN using the selected algorithmic approach for explaining ANN decisions. This research can contribute to better understanding of biases applied in interpersonal interactions.