2 results
Search Results
Now showing 1 - 2 of 2
Item Assessing the Reliability of Integrated Gradients-Based Saliency Maps for 3D Point Cloud Semantic Segmentation Models(The Eurographics Association, 2024) Ciprián-Sánchez, Jorge F.; Burmeister, Josafat-Mattias; Cech, Tim; Richter, Rico; Döllner, Jürgen; Hunter, David; Slingsby, AidanDeep learning models achieve high accuracy in the semantic segmentation of 3D point clouds; however, it is challenging to discern which patterns a model has learned and how it derives its output from the input. Recently, the Integrated Gradients method has been adopted to explain semantic segmentation models for 3D point clouds. This method can be used to generate saliency maps that visualize the contribution of input points to a particular model output. However, there is a lack of quantitative evaluation of the reliability of the generated saliency maps and the influence of the baseline selection (a central component of Integrated Gradients) on the method's results. In this paper, we quantitatively evaluate the reliability of saliency maps generated by the Integrated Gradients method for a 3D point cloud semantic segmentation model through well-known sanity checks from the image domain that we adapt to 3D point cloud segmentation. We perform these sanity checks for three different baselines to further evaluate the stability of the generated saliency maps concerning the baseline choice. Our results indicate that the Integrated Gradients method is sensitive to both the parameters of the model and training labels, unstable concerning the choice of baseline, and that, although it can identify points with high contributions to the model output, it fails to identify correctly if such contributions are positive or negative. Finally, we propose an averaging approach to aggregate the results of points that receive multiple scores from Integrated Gradients during the segmentation process and show that it produces saliency maps that better reflect high-contribution input points than previous approaches.Item Exploring High-Dimensional Data by Pointwise Filtering of Low-Dimensional Embeddings(The Eurographics Association, 2024) Atzberger, Daniel; Jobst, Adrian; Scheibel, Willy; Döllner, Jürgen; Hunter, David; Slingsby, AidanDimensionality reductions are a class of unsupervised learning algorithms that aim to find a lower-dimensional embedding for a high-dimensional dataset while preserving local and global structures. By representing a high-dimensional dataset as a twodimensional scatterplot, a user can explore structures within the dataset. However, dimensionality reductions inherit distortions that might result in false deductions. This work presents a visualization approach that combines a two-dimensional scatterplot derived from a dimensionality reduction with two pointwise filtering possibilities. Each point is associated with two pointwise metrics that quantify the correctness of its neighborhood and similarity to surrounding data points. By setting threshold for these two metrics, the user is supported in several scatterplot analytics tasks, e.g., class separation and outlier detection. We apply our visualization to a text corpus to detect interesting data points visually and discuss the findings.