Vivifying LIME: Visual Interactive Testbed for LIME Analysis

TL;DR

This paper addresses the limitations of LIME's single-image, non-interactive explanations by introducing LIMEVis, an interactive visualization tool that enables simultaneous exploration of multiple LIME results and direct manipulation of superpixels. The approach analyzes a VGG16 image classifier trained on STL-10 and uses PacMAP-based dimensionality reduction to visualize LIME results across many images, while enabling per-image, real-time prediction updates when superpixels are toggled. The main contributions are the multi-image LIME analysis workflow, interactive superpixel manipulation, and a demonstration showing how users can uncover shared influential features and pinpoint regions driving misclassifications. This work enhances model interpretation in practice by providing a tangible, interactive means to inspect and adjust explanations, potentially guiding model refinement and more reliable decisions in computer vision tasks.

Abstract

Explainable Artificial Intelligence (XAI) has gained importance in interpreting model predictions. Among leading techniques for XAI, Local Interpretable Model-agnostic Explanations (LIME) is most frequently utilized as it notably helps people's understanding of complex models. However, LIME's analysis is constrained to a single image at a time. Besides, it lacks interaction mechanisms for observing the LIME's results and direct manipulations of factors affecting the results. To address these issues, we introduce an interactive visualization tool, LIMEVis, which improves the analysis workflow of LIME by enabling users to explore multiple LIME results simultaneously and modify them directly. With LIMEVis, we could conveniently identify common features in images that a model seems to mainly consider for category classification. Additionally, by interactively modifying the LIME results, we could determine which segments in an image influence the model's classification.

Figures (1)

• Figure 1: Visual interface of LIMEVis. (A) In the Config View, users can set the desired image category for which they want to interpret the model's prediction and the LIME parameters (i.e. segmentation algorithm, positive_only, num_features, and hide_rest). (B) The Summary View shows the dimensionality-reduced features of LIME results on a 2D space. (C) The Overview shows the original image or the LIME result images. (D) If users select one image they want to see in detail, three related images appear: Original Image, LIME Image, and Superpixel Image. In Superpixel Image panel, users can select superpixels by left-clicking on it. (E) shows the model's prediction probabilities. The purple bar represents the prediction value for the image created from user-selected superpixels. This can be compared to the orange bar that is the prediction for the original image.