Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Sep 7·5 min read
TreeFI: Value-Aware Statistical Fault Injection for Deep Neural Networks
TL;DR
TreeFI introduces a value-aware statistical fault injection method for deep neural networks, significantly reducing the injection budget needed for reliability evaluation. It achieves up to 72.1x budget reduction while maintaining accuracy in failure-rate estimation.
✦ Why It Matters
Implement TreeFI to optimize fault injection processes and reduce testing costs in your DNN reliability evaluations.
Key Takeaways
- Utilize TreeFI for efficient fault injection in DNNs.
- Reduce testing budgets significantly while maintaining accuracy.
- Focus on high-impact value intervals for fault injection.
- Leverage regression trees to understand bit-flip behaviors.
How It Works
TreeFI partitions the value distribution of DNN layers into intervals based on similar expected bit-flip behavior. It uses regression trees to learn these behaviors, allowing for targeted fault injections that prioritize high-impact areas, thus optimizing the testing process.