Stefan Kolek

Munich, Germany · kolek@math.lmu.de · CV · GitHub

Education

• 2021-now PhD Computer Science
  • Ludwig-Maximilians-Universität München, Germany
  • Title: Sparse Representation Methods for Explainable Image Classification
• 2018-2021 MS Mathematics
  • Technische Universität Berlin, Germany
• 2015-2018 BS Mathematics
  • Ludwig-Maximilians-Universität München, Germany

Open Source

• CartoonX
  • Understanding the decisions of deep image classifiers can be challenging. CartoonX is a explainability map method that learns sparsity dirven masks directly on the wavelet/shearlet coefficients to maximize target class probability. The visual explanation extracts the relevant piece-wise smooth (cartoon) part of an image.
• BigFile
  • Many large datasets do not fit into RAM and consist of millions of compressed files which greatly slow down the dataloader during training. I built a general Python utility to transform the dataset from RAM to a single binary file with merged gzip streams. It is fully compatible with torch.utils.data.Dataset and is up to 4X faster than standard ImageNet dataloaders.

Publications

• Learning Interpretable Queries for Explainable Image Classification with Information Pursuit.
  • S. Kolek, A. Chattopadhyay, K. Chan, H. Andrade-Loarca, G. Kutyniok, R. Vidal. ICCV 2025.
• Beyond the Calibration Point: Mechanism Comparison in Differential Privacy.
  • G. Kaissis*, S. Kolek*, B. Balle, J. Hayes, D. Rückert. ICML 2024.
• Optimal privacy guarantees for a relaxed threat model: Addressing sub-optimal adversaries in differentially private machine learning.
  • G. Kaissis, A. Ziller, S. Kolek, A. Riess, D. Rückert. NeurIPS 2023.
• Explaining Image Classifiers with Multiscale Directional Image Representation.
  • S. Kolek, R. Windesheim, H. Andrade-Loarca, G. Kutyniok, R. Levie. CVPR 2023.
• Cartoon Explanations of Image Classifiers.
  • S. Kolek, D. Nguyen, R. Levie, J. Bruna, G. Kutyniok. ECCV 2022 (Oral).

Personal Interests