DAM: Diffusion Activation Maximization for 3D Global Explanations

TL;DR

This work addresses the opacity of 3D point-cloud models by proposing Diffusion Activation Maximization (DAM), a DDPM-based framework that yields global explanations. DAM introduces a symmetric Point Diffusion Transformer (PDT) and dual-classifier guidance (from $F$ and a noised $F'$) to generate perceptible, diverse global saliency maps, while Integrated Gradients for Diffusion (IGD) provides diffusion-path attributions with reduced bias. The approach achieves superior perceptibility, representativeness, and diversity compared with Autoencoder-based global explainers, and delivers faster generation times. Practically, DAM enables robust, globally interpretable explanations for point-cloud models across multiple architectures, potentially improving safety-critical deployments in robotics and healthcare.

Abstract

In recent years, the performance of point cloud models has been rapidly improved. However, due to the limited amount of relevant explainability studies, the unreliability and opacity of these black-box models may lead to potential risks in applications where human lives are at stake, e.g. autonomous driving or healthcare. This work proposes a DDPM-based point cloud global explainability method (DAM) that leverages Point Diffusion Transformer (PDT), a novel point-wise symmetric model, with dual-classifier guidance to generate high-quality global explanations. In addition, an adapted path gradient integration method for DAM is proposed, which not only provides a global overview of the saliency maps for point cloud categories, but also sheds light on how the attributions of the explanations vary during the generation process. Extensive experiments indicate that our method outperforms existing ones in terms of perceptibility, representativeness, and diversity, with a significant reduction in generation time. Our code is available at: https://github.com/Explain3D/DAM

Figures (18)

• Figure 1: Overview of DAM. A. Feeding the explanation into the classifier. B. Obtaining the target activation value and gradients. C. Guiding the DDPM model with the gradients. D. Generating a better explanation and E. Global saliency map.
• Figure 2: Overview of the DAM structure. There are two main explanations, one for the globally explainable sample $x_0$ (gray block on the right), and the other for the saliency map of the diffusion process (yellow block below).
• Figure 3: Visual comparison of path methods IGD and typical IG. The gradient path in the diffusion process is the integration from $x_0$ to $x_T$ (the black curve). Typical IG paths for $x_t$ are the linear integration of $x_0$ to $x_t$, which may lead to bias, while the path of IGD for $x_t$ is the integration from $x_{t-1}$ to $x_t$, which better approximates the real path.
• Figure 4: Global explanations of 5 classes generated by DAM. For comparison, we present the identical amount of explanations generated by AE, AED and NAED tan2023visualizing. More visualizations are shown in Fig. \ref{['More_DAM']}.
• Figure 5: Diversity examples. We randomly generated 5 explanations for the category "vase". For intuition, we also show 5 randomly chosen objects of the same class from the dataset (Random-5), and 5 samples that most highly activate the neuron "vase" (Top-5). More diversity is displayed in Fig. \ref{['More_DAM']}.