Services / Explainable AI evaluation
Explainable AI evaluation
Accuracy alone does not show whether a model decides for the right reasons. Viki Labs evaluates AI decisions with established explainability methods and with custom questionnaire applications that collect structured feedback from your domain experts. The approach follows the evaluation levels described by Doshi-Velez and Kim: with experts on real tasks, with people on simplified tasks, and with proxy measures.
What it covers
- Review of how the model decides, using methods such as LIME, SHAP and Grad-CAM
- Custom questionnaire application for your experts to judge decisions and explanations
- Analysis of where the model is right for the wrong reasons, or wrong with high confidence
- Evidence on whether explanations help people trust the model appropriately
What you receive
- Evaluation report with examples of failures and their likely causes
- The questionnaire application, configured for your use case
- Recommendations to improve the model or the way it is used
Every engagement starts with a paid audit or proof of concept. NDA available.