Vivifying LIME: Visual Interactive Testbed for LIME Analysis

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.

💡 Research Summary

The paper introduces LIMEVis, an interactive visual analytics system that extends the capabilities of Local Interpretable Model‑agnostic Explanations (LIME) for image‑based explainable AI. Traditional LIME generates explanations for a single image at a time, and any change in LIME’s parameters (e.g., segmentation algorithm, “positive only”, number of features, hide‑rest) produces a new result only for that image, making it difficult to explore the broader behavior of a model. LIMEVis addresses these limitations by allowing simultaneous analysis of many images and by providing direct manipulation of superpixels to observe their effect on model predictions in real time.

The system is built around four coordinated views:

1. Config View – Users select a target class from the STL‑10 dataset (10 categories) and set LIME parameters. Upon execution, the system runs LIME on 100 images belonging to the chosen class.

2. Overview – The 100 original images and their corresponding LIME heat‑maps are displayed in a 10 × 10 grid. Each image’s border is colored blue if the underlying VGG‑16 classifier predicts the correct class, or red if it misclassifies, giving an immediate visual cue of overall accuracy.

3. Summary View – Features are extracted from each image using a pre‑trained VGG‑16 network, then reduced to two dimensions with PacMAP. The resulting points are plotted in a scatter plot, colored identically to the Overview (blue/red). Users can brush clusters of points; the brushed images are highlighted in the Overview, enabling rapid identification of common visual patterns among correctly or incorrectly classified samples.

4. Detail View – When a single image is selected, three panels appear: the original image, the LIME‑generated heat‑map, and a “Superpixel Image” where each superpixel is outlined in yellow. Users can click any superpixel to toggle its visibility (masking it black or revealing it). Each toggle instantly updates a bar chart at the bottom that shows the VGG‑16 prediction probabilities for the original image (orange bar) and for the current superpixel configuration (purple bar). This real‑time feedback reveals how individual regions contribute to the classifier’s decision.

The authors demonstrate the system with a case study on the “dog” class. By brushing clusters in the Summary View, they discover that many correctly classified images share salient dog heads or tails, indicating these regions are heavily weighted by the model. For misclassified examples (e.g., a dog image mistakenly labeled as “cat”), the Detail View reveals that certain background superpixels drive the wrong prediction. When those superpixels are masked, the model’s confidence shifts dramatically from 0.74 for “cat” to 0.90 for “dog”, confirming the causal role of the selected regions.

Technical contributions include:

• Batch LIME execution across a large set of images, enabling comparative analysis.
• Integration of high‑dimensional feature extraction (VGG‑16) with PacMAP dimensionality reduction, providing an intuitive 2‑D embedding of LIME results.
• Interactive superpixel toggling with immediate model‑prediction updates, turning a static explanation into an exploratory experiment.

Limitations noted by the authors are the focus on a single classifier (VGG‑16) and dataset (STL‑10), the manual nature of superpixel selection, and the fixed batch size of 100 images, which may hinder scalability to larger corpora. Future work aims to close the loop by feeding user‑guided superpixel modifications back into the model training pipeline, thereby allowing users not only to diagnose but also to improve model performance directly within the visual interface.

In summary, LIMEVis transforms LIME from a single‑image, static explanation tool into a multi‑image, interactive testbed that supports both high‑level pattern discovery and low‑level causal probing of image classifiers. This advances the state of XAI by giving practitioners a concrete, hands‑on way to understand and debug black‑box vision models.