A Framework for Counterfactual Explanation of Predictive Uncertainty in Multimodal Models.

Abstract

Both predictive uncertainty estimation and visual explanation are crucial elements in helping humans understand the artificial intelligence (AI) decision-making process and in building trustworthy AI. However, there has been comparatively limited investigation into the intersection of these two domains in multimodal scenarios. In this article, we propose a universal explanation framework to evaluate counterfactual samples of predictive uncertainty in multimodal models. Inspired by multimodal representation learning, our framework leverages a shared latent space of multimodal variational autoencoders (MVAEs) to generate counterfactual explanations (CEs) of predictive uncertainty, enabling us to identify the input features contributing to high predictive uncertainty. To further evaluate the quality of counterfactual samples, we propose a Bayesian local linear approximation (BLLA) method. This method models the overall linear space as an inverse chi-square distribution while representing feature importance and the error term as normal distributions. By doing so, it captures the uncertainty and feature importance of each modality. Through a comprehensive suite of experiments conducted on multimodal classification and regression tasks, we demonstrate that our framework successfully generates accurate CEs of predictive uncertainty, establishes the consistency of feature importance, and comprehensively facilitates users' comprehension of multimodal model behavior.