Fast and Simple Explainability for Point Cloud Networks

TL;DR

This work introduces Feature Based Interpretability (FBI), where the features' norm is computed, per point, before the bottleneck, and demonstrates how the proposed measure is helpful in analyzing and characterizing various aspects of 3D learning, such as rotation invariance, robustness to out-of-distribution outliers or domain shift and dataset bias.

Abstract

We propose a fast and simple explainable AI (XAI) method for point cloud data. It computes pointwise importance with respect to a trained network downstream task. This allows better understanding of the network properties, which is imperative for safety-critical applications. In addition to debugging and visualization, our low computational complexity facilitates online feedback to the network at inference. This can be used to reduce uncertainty and to increase robustness. In this work, we introduce \emph{Feature Based Interpretability} (FBI), where we compute the features' norm, per point, before the bottleneck. We analyze the use of gradients and post- and pre-bottleneck strategies, showing pre-bottleneck is preferred, in terms of smoothness and ranking. We obtain at least three orders of magnitude speedup, compared to current XAI methods, thus, scalable for big point clouds or large-scale architectures. Our approach achieves SOTA results, in terms of classification explainability. We demonstrate how the proposed measure is helpful in analyzing and characterizing various aspects of 3D learning, such as rotation invariance, robustness to out-of-distribution (OOD) outliers or domain shift and dataset bias.

Figures (13)

• Figure 1: Typical data flow of a point cloud classification architecture.
• Figure 1: Critical Points, Gradients and FBI. Additional samples are provided to illustrate the smoothness of our method. Both critical points and gradients exhibit a substantial number of points with low or zero influence.
• Figure 2: Non-Smooth Gradients. Gradients in PointNet pointnet are zero outside the critical set (e.g., wing's base), and exhibit a non-smooth characteristic (e.g., wing's edge). This trend is similarly observed in DGCNN dgcnn. Our approach results in a smoother influence map, predicting the potential influence even for points with zero gradients.
• Figure 2: OOD evaluation on additional samples. Validations of the observed trend that GDANet exhibits a higher capability in attending to semantic meaning rather than outliers, whereas DGCNN performs comparatively worse.
• Figure 3: FBI (ours) vs. Critical Points. FBI provides rankings based on semantic meaning across the entire shape. Notably, elements like the cup handle or the top of the monitor exhibit high influence, while other parts receive smooth ranking. In contrast, critical points predominantly highlight prominent regions but, in other areas, the selection of points appears nearly random.

Theorems & Definitions (7)

• definition thmcounterdefinition: FBI
• proposition thmcounterproposition: Existence of zero gradients
• proof
• proposition thmcounterproposition: Smoothness
• proof
• lemma thmcounterlemma: Auxiliary
• proof