Feature selection is a critical aspect of eXplainable Artificial Intelligence (XAI), and it has implications for model interpretability and predictive performance. CoFE (Consistency-driven Feature Elimination) framework was introduced recently using a frequentist approach. CoFE eliminates features with inconsistent coefficient signs in Linear Regression models by estimating the Sign Entropy (variability of the sign) of the coefficients using Bootstrapping. However, the uncertainty associated with estimating Sign Entropy using bootstrapping leads to slower convergence and inconsistency in feature subset selection in CoFE. In this paper, we present Bay-CoFE, 1 a Bayesian reformulation of CoFE, to solve the slow convergence and inconsistency issues of CoFE while retaining the benefits of selecting features with lower Sign Entropy in a Bayesian framework. We provide theoretical justifications and empirical evidence to prove Bay-CoFE’s superior convergence properties. Across all datasets, Bay-CoFE achieves significantly superior sign stability compared to traditional feature selection methods (Mann-Whitney U test, p-value < = 1.63e-03 and mean Cliff’s delta ≈ 0.83), with minimal predictive performance differences (Mann-Whitney U test, p-value > 0.1 and mean Cliff’s delta ≈ 0.32), demonstrating a highly favorable trade-off for interpretable modeling.

Saliency maps are widely used as a post-hoc approach to explain the decision-making process of Deep Learning (DL) based image classification models, but evaluating their fidelity remains a complex problem. While saliency metrics have been introduced to evaluate the fidelity of saliency maps, existing saliency metrics, such as perturbation-based saliency metrics, have been previously reported to demonstrate statistical inconsistency. Although inconsistencies have been noted in different works, there exists no mechanism for estimating the same, i.e., Inconsistency Estimation (IE). Our primary objective is to address this limitation, and therefore, we propose a framework to estimate the inconsistency of saliency metrics for any given DL model. The framework enables building IE models for estimating the inconsistency by employing a set of perturbation types and schemes. The framework’s modular architecture provides flexibility across (i) perturbation types (Inpainting, Uniform, and Gaussian blur), (ii) perturbation schemes (pixel-wise and patch-wise), (iii) learning mechanisms (Convolutional Neural Networks and Vision Transformers) and (iv) IE modeling techniques (bagging and boosting). Extensive experimental results are shown on three well-known DL architectures (Inception-V3, Xception, and ResNet-50) on three different public datasets, including the Imagenette, Oxford-IIIT Pets Dataset, and PASCAL VOC 2007, along with results on ViTs for Oxford-IIIT Pets Dataset, and PASCAL VOC 2007. With a comprehensive evaluation of seven different perturbation types that include two inpainting, two Gaussian blur (with kernel widths of 0.9 and 1.5), and three uniform perturbations, our work shows the effectiveness of the proposed approach in estimating inconsistency. Statistically founded tests such as repeated cross-validation and the Permutation Test further validate the idea of the proposed framework for estimating the inconsistency of saliency metrics across unseen perturbations, making it useful in real-world scenarios.

Explanations of Local Interpretable Model-agnostic Explanations (LIME) are often inconsistent across different runs making them unreliable for eXplainable AI (XAI). The inconsistency stems from sign flips and variability in ranks of the segments for each different run. We propose a Bayesian Regularization approach to reduce sign flips, which in turn stabilizes feature rankings and ensures significantly higher consistency in explanations. The proposed approach enforces sparsity by incorporating a Sign Entropy prior on the coefficient distribution and dynamically eliminates features during optimization. Our results demonstrate that the explanations from the proposed method exhibit significantly better consistency and fidelity than LIME (and its earlier variants). Further, our approach exhibits comparable consistency and fidelity with a significantly lower execution time than the latest LIME variant, i.e., SLICE (CVPR 2024).

Class Activation Maps (CAMs) are one of the important methods for visualizing regions used by deep learning models. Yet their robustness to different noise remains underexplored. In this work, we evaluate and report the resilience of various CAM methods for different noise perturbations across multiple architectures and datasets. By analyzing the influence of different noise types on CAM explanations, we assess the susceptibility to noise and the extent to which dataset characteristics may impact explanation stability. The findings highlight considerable variability in noise sensitivity for various CAMs. We propose a robustness metric for CAMs that captures two key properties: consistency and responsiveness. Consistency reflects the ability of CAMs to remain stable under input perturbations that do not alter the predicted class, while responsiveness measures the sensitivity of CAMs to changes in the prediction caused by such perturbations. The metric is evaluated empirically across models, different perturbations, and datasets along with complementary statistical tests to exemplify the applicability of our proposed approach.

Local Interpretable Model-agnostic Explanations (LIME) - a widely used post-ad-hoc model agnostic ex-plainable AI (XAI) technique. It works by training a simple transparent (surrogate) model using random samples drawn around the neighborhood of the instance (image) to be explained (IE). Explanations are then extracted for a black-box model and a given IE, using the surrogate model. However, the explanations of LIME suffer from inconsistency across different runs for the same model and the same IE. We identify two main types of inconsistencies: variance in the sign and importance ranks of the segments (superpixels). These factors hinder LIME from obtaining consistent explanations. We analyze these inconsistencies and propose a new method, Stabilized LIME for Consistent Explanations (SLICE). The proposed method handles the stabilization problem in two aspects: using a novel feature selection technique to eliminate spurious superpixels and an adaptive perturbation technique to generate perturbed images in the neighborhood of IE. Our results demonstrate that the explanations from SLICE exhibit significantly better consistency and fidelity than LIME (and its variant BayLime).