Towards the Visualization of Aggregated Class Activation Maps to Analyse the Global Contribution of Class Features

AutorCherepanov, Igor; Sessler, David; Ulmer, Alex; Lücke-Tieke, Hendrik; Kohlhammer, Jörn
ArtConference Paper
AbstraktDeep learning (DL) models achieve remarkable performance in classification tasks. However, models with high complexity can not be used in many risk-sensitive applications unless a comprehensible explanation is presented. Explainable artificial intelligence (xAI) focuses on the research to explain the decision-making of AI systems like DL. We extend a recent method of Class Activation Maps (CAMs) which visualizes the importance of each feature of a data sample contributing to the classification. In this paper, we aggregate CAMs from multiple samples to show a global explanation of the classification for semantically structured data. The aggregation allows the analyst to make sophisticated assumptions and analyze them with further drill-down visualizations. Our visual representation for the global CAM illustrates the impact of each feature with a square glyph containing two indicators. The color of the square indicates the classification impact of this feature. The size of the filled square describes the variability of the impact between single samples. For interesting features that require further analysis, a detailed view is necessary that provides the distribution of these values. We propose an interactive histogram to filter samples and refine the CAM to show relevant samples only. Our approach allows an analyst to detect important features of high-dimensional data and derive adjustments to the AI model based on our global explanation visualization.
KonferenzWorld Conference on eXplainable Artificial Intelligence 2023