Project details

Explainable White
Theory-driven White

InDeep: Interpreting Deep Learning Models for Text and Sound

InDeep is a 6 year research project (2021-2027, part of the National Research Agenda programme), in which 5 universities and 7 companies collaborate. Goal of the research is to find ways to make popular Artificial Intelligence models for language, speech and music more explainable.

The project thus addresses the infamous “blackbox problem” with modern AI systems, that typically use machine learning to learn models from data. The project is part of a worldwide effort to develop techniques to explain the learned solutions, i.e. to open up the blackbox of such models. This effort has by now produced a whole toolbox of such techniques. Many of these techniques use small machine learning models to learn explanations for the large machine learning models.

Because the toolbox now contains many techniques, it is often not so easy to choose which technique to use. Moreover, which technique works best depends on the application area and how important ease of understanding and correctness are: for music recommendations or translation of safety critical documents, very different techniques might be chosen. An important goal of the project is to help create a “user manual” for the explanation toolbox.

Papers related to this project

Quantifying Context Mixing in Transformers