Understandable AI-reasoning through mechanistic explanations

Neural networks are an essential part of everyday life: they are speech assistants, recommend us movies and music, and support us through medical diagnosis systems. How they arrive at a decision is, however, often opaque, as these complex networks come with millions of parameters. This is especially problematic in critical applications such as medicine. At the MPI for Informatics, we develop methods to better understand the reasoning process of neural networks through human-understandable mechanistic explanations.