Explainable AI in Time-Sensitive Scenarios: Prefetched Offline Explanation Model

TL;DR

This work tackles the need for trustworthy, time-sensitive explainability in image classification by introducing Poem, a model-agnostic, offline–online explanation framework that produces exemplars, counterexemplars, and saliency maps. Poem builds a precomputed explanation base using abele and an Adversarial Autoencoder to enable fast, local explanations for new instances by matching latent representations and surrogate trees. Empirical results show substantial speedups over the prior approach and richer, more diverse explanations, validated on MNIST, FASHION, and EMNIST with RF and DNN back-ends, including qualitative saliency-map assessments. The work suggests promising directions for extending to other data types and real-world settings, with potential for user studies and drift-detection-driven retraining triggers.

Abstract

As predictive machine learning models become increasingly adopted and advanced, their role has evolved from merely predicting outcomes to actively shaping them. This evolution has underscored the importance of Trustworthy AI, highlighting the necessity to extend our focus beyond mere accuracy and toward a comprehensive understanding of these models' behaviors within the specific contexts of their applications. To further progress in explainability, we introduce Poem, Prefetched Offline Explanation Model, a model-agnostic, local explainability algorithm for image data. The algorithm generates exemplars, counterexemplars and saliency maps to provide quick and effective explanations suitable for time-sensitive scenarios. Leveraging an existing local algorithm, \poem{} infers factual and counterfactual rules from data to create illustrative examples and opposite scenarios with an enhanced stability by design. A novel mechanism then matches incoming test points with an explanation base and produces diverse exemplars, informative saliency maps and believable counterexemplars. Experimental results indicate that Poem outperforms its predecessor Abele in speed and ability to generate more nuanced and varied exemplars alongside more insightful saliency maps and valuable counterexemplars.

Figures (9)

• Figure 1: Components of the Adversarial AutoEncoder. The encoder maps an input instance $x$ to a point $z$ into a latent space. Any point in this space is first filtered by the discriminator before the decoder reconstructs the corresponding image.
• Figure 2: abele workflow, starting from the mapping of the instance $x$ to the extraction of the transparent model: (a) instance $x$ encoded within the latent space; (b) neighborhood generation around $x$; (c) discriminator filtering the neighborhood; (d) decoder transforming the points into images (e) annotated by the black box; (f) supervised data forming a local training set to (g) learn a decision tree.
• Figure 3: Extraction of rules and counter-rules from the surrogate model learned by abele in the neighborhood $H$
• Figure 4: Representation of the explanation base, where each point $t \in D_t$ is linked to the decision tree and the neighborhood computed during the offline step.
• Figure 5: Generation and selection of exemplars for a point $z_x$. Dashed lines represent the half-planes determined by the predicates of the rule. A set of points is generated in the latent space and filtered by the rule predicate (gray points). The top $k$ distant points are selected as exemplars (green dots)