Generative Adversarial Nets for Social Scientists

2020-11-04 | Input Talk | Marcel Neunhoeffer (Mannheim)

Abstract
In this talk I introduce Generative Adversarial Networks (GANs) for Social Scientists. GANs are an innovative neural network architecture where two neural networks adversarially learn arbitrary target distributions. A Generator network learns to produce simulated samples that mimic real data. At the same time, a Discriminator network learns to distinguish between real and simulated data. A GAN is successful in producing simulated data if a Discriminator is maximally uncertain about the origins of the data (real or simulated). GANs achieve impressive results in producing synthetic samples from complex data like images (e.g. cats, faces) or audio data (e.g. voices, songs). In this talk, I introduce current applications of GANs and present my work on their use for Social Science research. In particular, I will cover applications to Multiple Imputation, Small Area Estimation and the Generation of fully Synthetic Data. All applications will be accompanied by hands-on code examples.

Presenter
Marcel Neunhoeffer is a PhD Candidate and Research Associate at the chair of Political Science, Quantitative Methods in the Social Sciences, at the University of Mannheim. His research focuses on political methodology, specifically on the application of deep learning algorithms to social science problems. His substantive interests include data privacy, political campaigns, and forecasting elections.

[phnk]

Excellent talk, thanks a lot to the speaker!