Computer Vision News - December 2022

42 A New Definition of Interpretability went away and performed an extensive literature review to find all the proposed taxonomies of interpretable AI. Even papers within the same technical domain could be contradictory. “ Some researchers in the technical sciences like us define interpretability in one way, and others define it in another way, ” Mara continues. “ This isn’t helping us get anywhere, especially when discussing interpretability with people from different backgrounds. For example, a lawyer writing the AI Act for the EU uses words and sentences that don’t interface very well with what we’re doing and vice versa. ” They worked for months with philosophers, social scientists, ethicists, cognitive scientists, and colleagues from the technical domain, looking at many papers and definitions to find agreement. They submitted their work to the Artificial Intelligence Review journal and received helpful feedback, allowing them to formalize the definition further. In April 2021, a virtual roundtable public meeting was held on ‘ A Global Taxonomy for Interpretable AI, ’ where a group of researchers from various backgrounds came together to work on a global definition of interpretability . The meeting was endorsed by AI4media, a European Union Horizon 2020 project on which Mara and Vincent both work. Its goal is to build a network of excellence for AI in the media sector and society . The multidisciplinary consortium has a wide range of partners, including lawyers, sociologists, ethicists, and human rights defenders. “ We realized we were all using a different terminology to address the goals of the project, ” Mara tells us. “ One of these goals is to develop explainable and robust AI tools for the media sector . We decided to call a bunch of experts and invite them to participate in a global event to express their opinion on the meaning of interpretability. ” After the meeting, Mara and Vincent Mara Graziani is a Postdoc Researcher at IBM Research Europe and the University of Applied Sciences Western Switzerland, where Vincent Andrearczyk is a Senior Researcher. They speak to us about their work with global stakeholders to converge on an overarching definition for the interpretability of AI systems.