Feature Attribution Stability Suite: How Stable Are Post-Hoc Attributions?

1. 1Identify the limitations of existing metrics for evaluating feature attribution stability
2. 2Develop a suite of metrics that condition on prediction preservation and capture explanation fragility separately from model sensitivity
3. 3Apply the Feature Attribution Stability Suite to various post-hoc feature attribution methods and evaluate their performance under realistic input perturbations
4. 4Analyze the results to determine the stability of different feature attribution methods and identify areas for improvement

Machine learning researchers and engineers working on safety-critical vision systems can benefit from this research to improve the reliability of their models, while data scientists and AI engineers can apply these methods to evaluate the stability of their own feature attribution methods

💡 The stability of post-hoc feature attribution methods is crucial for safety-critical vision systems and can be evaluated using a suite of metrics that condition on prediction preservation

🚨 Improve reliability of safety-critical vision systems with the Feature Attribution Stability Suite! 🚨