FAME: Formal Abstract Minimal Explanation for Neural Networks
arXiv cs.AI / 3/12/2026
📰 NewsIdeas & Deep AnalysisModels & Research
Key Points
- FAME (Formal Abstract Minimal Explanations) is proposed as a new class of abductive explanations for neural networks, grounded in abstract interpretation.
- It scales to large networks and reduces explanation size by using dedicated perturbation domains that eliminate the need for traversal order, with progressively shrinking domains and LiRPA-based bounds to discard irrelevant features.
- The method converges to a formal abstract minimal explanation and introduces a procedure to measure the worst-case distance between an abstract minimal explanation and a true minimal explanation, combining adversarial attacks with an optional VERIX+ refinement step.
- Empirical benchmarks show consistent gains in explanation size and runtime on medium- to large-scale networks compared to VERIX+.
- The work contributes to explainable AI by providing a scalable, formal framework for generating minimal explanations for neural networks.
Related Articles
The Honest Guide to AI Writing Tools in 2026 (What Actually Works)
Dev.to
Next-Generation LLM Inference Technology: From Flash-MoE to Gemini Flash-Lite, and Local GPU Utilization
Dev.to
The Wave of Open-Source AI and Investment in Security: Trends from Qwen, MS, and Google
Dev.to
How I built a 4-product AI income stack in 4 months (the honest version)
Dev.to
I stopped writing AI prompts from scratch. Here is the system I built instead.
Dev.to
