LLEXICORP: End-user Explainability of Convolutional Neural Networks

CNNs now have automated explanation tools: LLEXICORP combines neural network analysis with AI language models to automatically explain what computer vision systems "see" and why they make decisions, eliminating the need for manual expert interpretation of thousands of activation images.

Explanations can be customized for different audiences: The system generates tailored descriptions ranging from technical details for AI experts to simple summaries for non-technical stakeholders, making AI decision-making more accessible to people who need to understand or oversee these systems.

Automation improves scalability of AI transparency: By automating the naming of visual concepts and explanation generation, LLEXICORP addresses a major bottleneck in explainable AI—the manual work required to make neural networks interpretable—allowing these explanations to be applied more broadly across different applications.

Faithfulness safeguards prevent misleading explanations: The researchers designed specific prompts and separated naming from explanation tasks to ensure the language model's descriptions accurately reflect what the neural network actually learned, rather than generating plausible-sounding but inaccurate explanations.

Explainability matters for high-stakes applications: The paper emphasizes that as CNNs are deployed in "critical areas" (like medical imaging or autonomous vehicles), understanding why these systems make decisions becomes essential for responsible deployment and user trust.