Flow AM: Generating Point Cloud Global Explanations by Latent Alignment

TL;DR

This work tackles the lack of reliable global explanations for 3D point cloud models by introducing Flow AM, an activation-flow based Activation Maximization method that does not rely on generative priors. By regularizing intermediate-layer activations to align with real object contours (latent alignment) and enforcing point continuity while respecting dataset bounds, Flow AM yields perceptible, faithful global explanations. Extensive experiments on ModelNet40 (with PointNet) and ShapeNet demonstrate that Flow AM achieves superior perceptibility and maintains fidelity to the classifier, avoiding the fidelity pitfalls observed in generative-model based AM. The study also includes sanity checks, ablations, and limitations, highlighting the trade-off between perceptibility and priors and outlining future directions for diversity and adaptive weighting.

Abstract

Although point cloud models have gained significant improvements in prediction accuracy over recent years, their trustworthiness is still not sufficiently investigated. In terms of global explainability, Activation Maximization (AM) techniques in the image domain are not directly transplantable due to the special structure of the point cloud models. Existing studies exploit generative models to yield global explanations that can be perceived by humans. However, the opacity of the generative models themselves and the introduction of additional priors call into question the plausibility and fidelity of the explanations. In this work, we demonstrate that when the classifier predicts different types of instances, the intermediate layer activations are differently activated, known as activation flows. Based on this property, we propose an activation flow-based AM method that generates global explanations that can be perceived without incorporating any generative model. Furthermore, we reveal that AM based on generative models fails the sanity checks and thus lack of fidelity. Extensive experiments show that our approach dramatically enhances the perceptibility of explanations compared to other AM methods that are not based on generative models. Our code is available at: https://github.com/Explain3D/FlowAM

Figures (15)

• Figure 1: Visualization comparison of non-generative model-based point cloud AM approaches with the target label "Table". From left to right : Vanilla AM, regularized by L2-norm, with Gaussian blur applied, regularized by Total variance and our flow-based AM.
• Figure 2: Upper part: similarities in the activation of neurons within the model when predicting real objects of the same (red) and different classes (blue), respectively. The purple curve indicates the activation similarity between non-generative AMs and real objects. Lower part: similarities between the activation of neurons within the model when predicting real objects and when predicting explanations generated by non-generative model-based (purple), generative model-based AM (green) and our flow-based AM (black), respectively. The y-axis indicates the cosine similarity and the x-axis is the name of each layer of PointNet. We utilize cosine similarity to calculate the activation likelihood of each layer, which yields similar results with other metrics such as Pearson and Spearman's coefficients.
• Figure 3: Overview of the structure of Flow AM. Raw AM represents the explanations generated by AM without any regularization.
• Figure 4: The intuition of the point continuity loss. Point A is a critical point that tends to expand during AM, while points B and C are common points that are ignored by the gradient. The outer dashed curve is the boundary of a legal point cloud instance. The green and blue arrows indicate the direction of two different terms in the continuity loss, i.e., the distance to the origin and to the nearest neighboring point, respectively.
• Figure 5: Qualitative comparison of non-generative model-based global explanations. From left to right are, no regularization, L2, Gaussian blur, total variance, flow regularizations (ours), AE tan2023visualizing, AED tan2023visualizing and AED tan2023visualizing. Note that the first five approaches are based on non-generative models while the last three are based on generative models. More visualizations can be seen in Fig. \ref{['Fig:more_visu_compare']}.