Leila Arras

Leila Arras is currently a doctoral researcher and research associate at the Fraunhofer Heinrich Hertz Institute, where she is part of the Explainable Artificial Intelligence research group within the Department of Artificial Intelligence and a member of the BIFOLD Graduate School. Among her qualifications, she previously earned an M.Sc. in Computer Science with distinction from Technische Universität Berlin, specializing in machine learning and scalable data science. Her research interests include machine learning, neural networks, interpretability, and their applications to natural language processing and computer vision. As a researcher, she is particularly interested in understanding the inner workings of neural networks and developing new methods to make artificial intelligence models more transparent.

Her research has broadened the scope of explainable artificial intelligence (XAI) to novel tasks and model architectures, in particular by extending XAI methods to recurrent neural networks for the processing of textual and sequential data. In addition, she has introduced new quantitative evaluation approaches for the objective assessment and comparison of XAI methods. Through an international collaboration, she advanced causal inference by introducing a neural network based framework for decomposing probabilistic predictions from an assumed causal model to identify combinations of causes increasing the risk of an outcome. More recently, her research has focused on examining and extending decomposition-based attribution methods for Transformer-based large language models.