Posts

Our latest work presents a new approach to disentangle image generation performance by decomposing cosine similarity into cluster-level contributions using central kernel alignment. This allows us to quantify how different pixel regions contribute to overall image quality, enabling more fine-grained evaluation and improved explainability of generative models across real-world use cases.

We introduce Explanatory Performance Estimation (XPE), a method that explains model behavior under feature shifts by linking performance changes to interpretable input characteristics using Optimal Transport and Shapley Values. This enables explanatory model monitoring across image, audio, and tabular data.

We show how integrating human gaze information aligns human and machine attention, improving model robustness and explainability. Demonstrated on real-world visual quality inspection, the approach highlights the value of explicit human knowledge in training trustworthy AI.

We improve the quality of feature attributions by optimally combining multiple explanation methods. Our convex-combination strategy enhances robustness and faithfulness, consistently outperforming individual methods and baselines across architectures. Published at the International Conference on Machine Learning.

This work compiles a domain-agnostic taxonomy of pitfalls in validation metrics, based on a multistage Delphi process and community feedback. It offers practical guidance to improve evaluation practices across application domains.

We present Metrics Reloaded, a comprehensive framework for problem-aware selection of evaluation metrics in biomedical image analysis. Developed through a large international Delphi process, it introduces the concept of a problem fingerprint to guide researchers toward meaningful and domain-relevant validation.

Our latest work introduces the first bias–variance–covariance decomposition for kernel scores, providing a unified framework for uncertainty estimation in generative models. We show how kernel-based entropy and variance capture uncertainty across image, audio, and language generation — even in closed-source models.

Paper accepted! Our latest work on developing a versatile framework for rapid prototyping and training of a wide range of factor analysis models for multi-omics data got accepted at this year’s ICML workshop on computational biology!

With 2 million Euros in funding from the European Research Council, we will be developing AI models to support doctors in the diagnosis and treatment of cancer. Read more!

Paper accepted! Our latest work on a general bias-variance decomposition for proper scores got accepted at this year’s AISTATS conference!