Machine Learning Researcher from Denmark. My research interests are interpretability, robustness, and compositionality. My work has been published at ACL, EMNLP, and at workshops at NeurIPS.
Preprint, 2025
A novel form of self-repair in the softmax operation leads to flaws in pertubation and gradient based attribution methods. GIM (gradient interaction modification) is a new method that alleviates many existing issues with gradient-based attribution methods and achieves state-of-the-art on multiple benchmarks.
Empirical Methods in Natural Language Processing (EMNLP), 2025
A multi-agent system that performs clinical coding with human-like reasoning steps; demonstrates competitive accuracy on medical coding benchmarks and outperforms existing approaches on rare codes with no fine-tuning.
Preprint, 2025
Introduces the Bi-Axial Transformer (BAT) that attends across both clinical-variable and time axes in EHRs. BAT achieves SOTA on sepsis prediction, is competitive for mortality classification, and shows robustness to missingness; baselines were re-implemented in PyTorch for reproducibility.
Association for Computational Linguistics (ACL), 2025
AOPC is frequently used to assess the quality of explanations generated by attribution methods. However, the method suffers from a fundamental flaw when features interact that renders the finding of many previous studies invalid. We identify the cause of the problem and propose a solution.
IEEE Access, 2021
CNN-based method that crops images directly—no explicit saliency modeling or candidate ranking—trained on editor-curated crops and producing boxes for multiple aspect ratios.