Moritz Böhle

Moritz Böhle is a researcher at the Max Planck Institute for Informatics, affiliated with the Computer Vision and Machine Learning department. He completed his PhD in 2024 at Saarland University, titled Towards Designing Inherently Interpretable Deep Neural Networks for Image Classification. His work focuses on interpretable machine learning, particularly in developing architectures like B-cos Networks to align models for faithful explanations. He has contributed to foundational research in model interpretability, including transforming pre-trained models (B-cosification), knowledge distillation with explanations, and automated concept discovery.

His research spans computer vision tasks, vision transformers, and convolutional neural networks, with a strong emphasis on aligning model components for transparency. He has co-authored papers in top venues like ICLR, NeurIPS, CVPR, and ECCV, addressing challenges in post-hoc explanations, alignment mechanisms, and systematic evaluation of attribution methods. His work bridges theoretical advancements and practical applications in explainable AI.

Key contributions include B-cos alignment for CNNs and transformers, explanation-enhanced knowledge distillation, and frameworks for task-agnostic concept discovery. These efforts aim to make AI systems more transparent and trustworthy while maintaining competitive performance.