Jonas Mirlach

Reference-Guided Machine Unlearning

Jonas Mirlach, Sonia Laguna, Julia E. Vogt

We proposed a machine unlearning method that uses predictions on unseen data to guide forgetting while preserving accuracy on retained data.

PyTorch vs. JAX: Automatic Differentiation Performance on Scientific Codes

Baraq Lipshitz, Jonas Mirlach, Amin Oudrhiri, Gerald Prendi, Leyla Yaayladere

We compared reverse-mode automatic differentiation in PyTorch and JAX, examining how computational patterns affect performance on scientific workloads across CPUs and GPUs.

R-LiViT: A LiDAR-Visual-Thermal Dataset Enabling Vulnerable Road User Focused Roadside Perception

Jonas Mirlach, Lei Wan, Andreas Wiedholz, Hannan Ejaz Keen, Andreas Eich

We created a roadside dataset combining LiDAR, RGB and thermal recordings, with a focus on detecting pedestrians and cyclists during day and night.

Evaluating Apertus Models for Multilingual Legal Headnote Generation

Kaushik Karthikeyan, Jonas Mirlach, Max Neuwinger

We evaluated fine-tuning approaches for Apertus models to generate headnotes for Swiss court decisions, examining summary quality across German, French and Italian.

Feed-forward Dynamic Scene Reconstruction with Humans Using Gaussian Splatting

İrem Demir, Emircan Gündoğdu, Jonas Mirlach

We developed a Gaussian splatting approach to reconstruct people and their surroundings from monocular video, with a human representation that supports reposing.

Uncertainty Guided Knowledge Distillation for Monocular Depth Estimation

Tengerleg Enkhtuvshin, Jonas Mirlach, Iaroslava Novoselova, Piotr Wilczyński

We tested whether filtering uncertain teacher predictions helps smaller models learn to estimate depth from a single image through knowledge distillation.

Leveraging Explainable AI for Informed Building Retrofit Decisions: Insights from a Survey

Daniel Leuthe, Jonas Mirlach, Simon Wenninger, Christian Wiethe

We examined how people understand explanations of building energy predictions, comparing machine learning models and explanation methods through a survey of 137 participants.

Adversarial Robustness as a Predictor of Catastrophic Forgetting

Arthur Kirchgessner, Stefan Künzli, Jonas Mirlach, Alexandre Zwahlen

We studied whether an image’s robustness to adversarial perturbations predicts how likely a model is to forget it when learning new classes.

Towards Personalised Prediction of PTSD from Speech

Jonas Mirlach

We explored predicting PTSD from speech, comparing acoustic features and machine learning models and testing personalisation through groups of similar speakers.

Enabling the Smart Factory: A Digital Platform Concept for Standardized Data Integration

Julia Donnelly, Andreas John, Jonas Mirlach, Kilian Osberghaus, Silvia Rother, Christian Schmidt, Hannes Voucko-Glockner, Simon Wenninger

We designed a platform concept for providing manufacturers with standardised data integration modules, drawing on literature and interviews with industry experts.