Deep learning methods achieve their remarkable performance from a very large amount of annotated training data. However, in many applications in the sciences, e.g. in medicine, the annotation by the user/expert is very expensive or completely impossible. This core project therefore deals with modern, unsupervised learning methods and their application in the sciences. Specifically, methods for learning generative models, such as e.g. Generative Adversarial Networks (Goodfellow et al. 2014), are examined and expanded so that, for example, unsupervised methods for anomaly detection can be developed. In addition, such models might also be used for data augmentation, i. e. to generate additional training data, or be the basis for estimating and subsequently analysing the underlying data distribution. This core project supports two pilot projects, P2: Detection of causal relationships through deep learning, and P3: Data-driven virus diagnostics at multiple levels I (Methods).
K2: Knowledge gain – Generative modelsBettina Färber