Visual Analytics for Deep Learning

This project addresses the research problem of visualization for machine learning (Vis4ML). We aim to open the black box of machine learning for the case of deep learning (DL), making DL models more transparent and controllable. To achieve this, we develop and evaluate visual analytics methods that allow users to better understand, improve, and control DL models. We plan to show both the features learned and make the learning and decision process more transparent and controllable by giving humans access to internal information of DL models in such a way that they can understand the model decisions and improve the final performance. A challenge is the complexity of the DL models, where direct visualization of the models is only partially possible and useful. Therefore, we leverage the power of visual analytics: relying on automatic data analysis as much as possible and using interactive visualization. For automatic analysis, we investigate both unsupervised machine learning (e.g., dimension reduction) and supervised machine learning methods (not necessarily DL). In this sense, we follow the approach of machine learning (as data analysis approach) for (interactive) visualization for machine learning (as the machine learning problem in the application areas): ML4Vis4ML. We investigate a representative number of application domains, including fluid mechanics and cognition-inspired learning related to visual attention and natural language processing.