Vivifying LIME: Visual Interactive Testbed for LIME Analysis

LIME has limitations in current practice: While LIME is widely used to explain AI model decisions, it traditionally analyzes only one image at a time and doesn't allow users to interact with or modify the explanations, restricting how thoroughly people can understand what a model is actually doing.

Interactive tools can improve AI transparency: LIMEVis demonstrates that adding visualization and interactive features—such as viewing multiple explanations simultaneously and directly manipulating which image segments are analyzed—makes it easier for people to identify patterns in how AI models make decisions.

Visual explanations reveal potential model biases: By using LIMEVis to examine multiple images together, users can discover common features that models rely on for classification, which may help identify whether models are focusing on relevant factors or potentially problematic patterns.

Direct manipulation enables deeper investigation: The ability to interactively modify which image segments are tested allows users to determine cause-and-effect relationships between specific visual elements and model predictions, moving beyond passive observation to active testing.

Better tools support responsible AI deployment: Improved explainability tools like LIMEVis can help organizations and stakeholders more effectively audit AI systems before deployment, potentially catching errors or biases that single-image analysis might miss.