Explorable explanations for machine learning model behaviours

Recent advancements in AI have led to the creation of increasingly complex models. While these models offer high predictive accuracy, they often function as "black boxes", making their behaviour, particularly in edge cases, difficult to fully comprehend. This lack of transparency hinders their deployment in critical fields like healthcare, where establishing trust in model behaviour is essential. Current explanation methods, such as LIME, SHAP, counterfactuals, or mechanistic interpretations, provide static insights. But they fail to support deeper investigation into model behaviour or provide the flexibility needed to explore the full spectrum of candidate models.

In this talk, I will discuss how we can enhance the understanding of machine learning models through interactive visualization. I will describe the challenges and opportunities arising from the Rashomon effect (existence of multiple accurate yet different models). Finally, I will demonstrate how explorable interactive visualisations can facilitate a more comprehensive investigation of the model spectrum.