Robin Hesse

I am an incoming postdoctoral researcher (starting January 2026) in the Department for Computer Vision and Machine Learning at the Max Planck Institute for Informatics. My research focuses on narrowing the gap between human and machine intelligence, especially in the context of visual perception. To this end, I study methods of explainable artificial intelligence.

• 

  Activation Subspaces for Out-of-Distribution Detection

  Barış Zöngür, R. Hesse, S. Roth, in International Conference on Computer Vision (ICCV), 2025.

  Separating representations into important and unimportant components significantly improves out-of-distribution detection.
• 

  Beyond Accuracy: What Matters in Designing Well-Behaved Models?

  R. Hesse, D. Bagci, B. Schiele, S. Schaub-Meyer, S. Roth, in arXiv:2503.17110 [cs.CV], 2025.

  A large-scale study that evaluates how different design choices and training strategies affect nine different quality dimensions of over 300 backbone models for image classification.
• 

  Disentangling Polysemantic Channels in Convolutional Neural Networks

  R. Hesse, S. Schaub-Meyer, J. Fischer, S. Roth, in CVPR 2025 Workshop on Mechanistic Interpretability for Vision (MIV), oral presentation.

  A method to disentangle polysemantic channels in a CNN by manually restrcuturing the weights.
• 

  Benchmarking the Attribution Quality of Vision Models

  R. Hesse, S. Schaub-Meyer, S. Roth, in Conference on Neural Information Processing Systems (NeurIPS), 2024.

  An evaluation protocol for attribution methods that overcomes the limitations of lacking inter-model comparisons and unaligned training and testing domains.
• 

  FunnyBirds: A Synthetic Vision Dataset for a Part-Based Analysis of Explainable AI Methods

  R. Hesse, S. Schaub-Meyer, S. Roth, in International Conference on Computer Vision (ICCV), 2023, oral presentation.

  A novel dataset that allows for image space interventions to evaluate explainable artificial intelligence (XAI) methods.
• 

  Content-Adaptive Downsampling in Convolutional Neural Networks

  R. Hesse, S. Schaub-Meyer, S. Roth, in CVPR 2023 Workshop on Efficient Deep Learning for Computer Vision (ECV).

  We propose an adaptive downsampling scheme that allows CNNs to process informative regions at a higher resolution than less informative ones.
• 

  Fast Axiomatic Attribution for Neural Networks

  R. Hesse, S. Schaub-Meyer, S. Roth, in Advances in Neural Information Processing Systems (NeurIPS), vol. 34, 2021.

  For ReLU DNNs without bias-terms, we show that Input×Gradient is a closed-form solution of Integrated Gradients that is therefore significantly more efficient to compute than the usual numerical approximation.
• 

  Homography Estimation in the Realm of Deep Learning

  Robin Hesse - Master's Thesis (2020)

  Even though deep learning has caused a paradigm shift in most computer vision applications, traditional handcrafted methods are still used extensively for image registration. In this thesis, we show that deep learning also caused a paradigm shift in image registration...
• 

  Development and Evaluation of 3D Autoencoders for Feature Extraction

  Robin Hesse - Bachelor's Thesis (2017)

  This thesis addresses the problem of aligning connectome data. The advancing scale in connectome reconstruction introduces new problems. One particular problem are spatial shifts and distortions in the image data that make it hard to come to a global alignment...
• 

  Efficient Verification of Program Fragments: Eager POR

  P. Metzler, H. Saissi, P. Bokor, R. Hesse, and N. Suri, in Automated Technology for Verification and Analysis, 2016.

  Software verification of concurrent programs is hampered by an exponentially growing state space due to non-deterministic process scheduling. Partial order reduction (POR)-based verification has proven to be a powerful technique to handle large state spaces...