Surrogate Model-Based Explainability Methods for Point Cloud NNs

Abstract

In the field of autonomous driving and robotics, point clouds are showing their excellent real-time performance as raw data from most of the mainstream 3D sensors. Therefore, point cloud neural networks have become a popular research direction in recent years. So far, however, there has been little discussion about the explainability of deep neural networks for point clouds. In this paper, we propose new explainability approaches for point cloud deep neural networks based on local surrogate model-based methods to show which components make the main contribution to the classification. Moreover, we propose a quantitative validation method for explainability methods of point clouds which enhances the persuasive power of explainability by dropping the most positive or negative contributing features and monitoring how the classification scores of specific categories change. To enable an intuitive explanation of misclassified instances, we display features with confounding contributions. Our new explainability approach provides a fairly accurate, more intuitive and widely applicable explanation for point cloud classification tasks. Our code is available at https://github.com/Explain3D/Expla inable3D

BubblEX: An Explainable Deep Learning Framework for Point-Cloud Classification

“…Another approach for explaining the decision of a DNN when it deals with 3D data is described in [36]. The authors proposed a point cloud-applicable explainability method based on local surrogate model-based approach to demonstrate which components are responsible to the classification.…”

“…Compared to [35], which performs tests by randomizing the weights of the network, we directly choose which layer to study (through visualization techniques) making our approach more punctual, precise and effective. Finally, we do not have to train surrogate models like [36], but we can directly use the original classification approach, without the need to retrain it.…”