Research Seminar on "Teaching deep neural networks to localize sources in super-resolution microscopy via simulation-based and unsupervised learning"
Research Seminar on "Teaching deep neural networks to localize sources in super-resolution microscopy via simulation-based and unsupervised learning" by Artur Speiser, CNE – Computational Neuroengineering, TUM
Abstract: Single-molecule localization microscopy (SMLM) requires algorithms for inferring the locations of stochastically active fluorophores from noisy imaging data. Given their success in related computer vision tasks like segmentation and object localization deep neural networks (DNN) are an apparent alternative approach to iterative methods usually employed in SMLM. First successful applications use simulated data sampled from a generative model to train their networks in a supervised fashion as it is infeasible to obtain real labeled data for that task. We implemented and tested alternative training methods based on the variational autoencoder methods and show how they can make training more robust to model mismatch. Furthermore, we developed a network architecture that is able to use context from multiple images to benefit from the activation dynamics of fluorophores. Our method achieves state of the art performance on the popular SMLM challenge in all imaging conditions.
Bio: Diploma in Physics at the Karlsruher Institut für Technologie, Masters in Cognitive Sciences at the University of Osnabrueck and started his PhD at the caesar research center Bonn.
May 15, 16:30h, Karlstr. 45, 6th floor, room 6009.