On the conditions for integrating deep learning into the study of visual politics

Abstract

Traditional methods to study visual politics have been limited in geographical, media and temporal coverage. Recent advances in deep learning have the potential to dramatically extend the scope of the field especially with respect to making sense of the contemporary social and political developments in digital media. While some early adopters may be tempted to take the new computational tools at face value, others see their black-box character as cause for concern. This paper argues that the integration of deep learning into the study of visual politics must be approached still more critically and boldly. On the one hand, the complexity of visual political themes requires a more substantial human involvement if compared with other applications of deep neural networks. Therefore, a question is how the scientist and the network should best interact. On the other hand, it is important to acknowledge that a deep learning tool will never simply replace specific tasks inside a research process: its adoption has implications for the broader process from the delineation of the object of analysis, to data collection, to the interpretation and communication of results. We examine the conditions of integrating a deep learning tool for image classification into the large-scale study of visual politics in digital and social media along these two dimensions.