Helmholtz Zentrum München & Technical University Munich
Good Gradients and How To Find Them: Towards Multi-Scale Representation Learning
Modern machine learning techniques are based on the concept of having dynamic representations of a data set that can be updated continuously during training. While this perspective has led to astounding successes in image analysis, for instance, the analysis of high-dimensional data sets in the life sciences is still plagued by the occurrence of spurious features in embeddings, leading to specious visualisations. Topological methods, which focus on more fundamental aspects of a data set, can help alleviate this problem. Being of a discrete, i.e. non-differentiable, nature, their integration into modern machine learning techniques proves to be challenging. Fortunately, this gap started to be bridged recently! In this talk, I will discuss recent advances in topological data analysis that enable the development of multi-scale aware representation learning techniques. I will primarily focus on building intuition for such methods, making this talk accessible to a wide audience of machine learning practitioners.
Hensel et al. :A Survey of Topological Machine Learning Methods (Frontiers in Artificial Intelligence, Volume 4, 2021); Moor et al.: Topological Autoencoders (Proceedings of ICML 2020); Rieck et al.: Uncovering the Topology of Time-Varying fMRI Data using Cubical Persistence (Proceedings of NeurIPS, 2020)
Bastian is the principal investigator of the AIDOS Lab at the Institute of AI for Health at the Helmholtz Centre Munich, Germany. His main research interests are algorithms for graph representation learning and time series analysis, with a focus on biomedical applications and healthcare topics. Bastian is also enticed by finding new ways to explain neural networks using concepts from algebraic and differential topology. He is a big proponent of scientific outreach and enjoys blogging about his research, academia, supervision, and software development. Bastian received his M.Sc. degree in mathematics, as well as his Ph.D. in computer science, from Heidelberg University in Germany.