hide
Free keywords:
Poster
Abstract:
In geosciences, machine learning (ML) has become essential for solving complex problems, such as predicting natural disasters or analysing the impact of extreme temperatures on mortality rates. However, the integration of ML into geoscience scenarios faces significant challenges, especially in explaining the influence of hyperparameters (HP) on model performance and model behaviour in specific scenarios. The Explainable Artificial Intelligence (XAI) system ClarifAI developed at GFZ addresses these challenges by combining XAI concepts with interactive visualisation.
ClarifAI currently provides users with two interactive XAI methods: HyperParameter Explorer (HPExplorer) and Hypothetical Scenario Explorer (HSExplorer).
HPExplorer allows interactive exploration of the HP space by computing an interactive tour through stable regions of the HP space. We define a stable region in HP space as a subspace of HP space in which ML models show similar model performance. We also employ HP importance analysis to deepen the understanding of the impact of separate HPs on model performance.The Hypothetical Scenarios Explorer (HSExplorer) helps users explore model behaviour by allowing them to test how changes in input data affect the model's response.
In our presentation, we will demonstrate how HSExplorer helps users understand the impact of individual HPs on model performance. As ClarifAI is an important research area in our lab, we are interested in discussing relevant XAI challenges with the XAI community in ESSI.
Our goal is to create a comprehensive set of tools that explain the mechanics of ML models and allow practitioners to apply ML to a wide range of geoscience applications.
