Making sense of data and models
Abstract:
Extracting actionable knowledge from scientific data and gaining insight into machine learning models are major goals in modern science. In this talk, I will argue that these are essentially the same problem and show how they can be addressed with essentially the same tools. To start with, I will present novel methods for extracting trait-specific patterns from high-dimensional genomics data and for discovering and characterizing exceptional subpopulations in material science data. I will then show how we can use the very same methods to understand when machine learning models are prone to make errors. Finally, I will show how we can use these techniques to gain direct insight into the decision process of modern deep neural networks.
About Nils:
I am a second-year Ph.D. student at CISPA Helmholtz Center for Information Security (Protected link to cispa.de), supervised by Jilles Vreeken (Protected link to vreeken.eu). I am broadly interested in robust and explainable machine learning for large-scale real-world applications. In my Ph.D, I intend to develop new approaches that are at the same time descriptive and predictive. That is the models not only offer predictive capabilities but also facilitate practitioners to gain deeper insights into the problems they are addressing. Currently, I mostly work on understanding how neural networks process information and building methods to describe when and how models make errors. Before joining CISPA, I was a research assistant in the goup of Bernt Schiele (Protected link to mpi-inf.mpg.de)at the Max-Planck-Institut for Informatics (Protected link to mpi-inf.mpg.de), supervised by David Stutz (Protected link to davidstutz.de). My research focused on adversarial and out-of-distribution robustness of Quantized Neural Networks. I also worked on the influence of Batch Normalization on the vulnerability and generalization capabilities of neural networks.