Explainable AI: Building trustworthy AI models?
- 1. explainX.ai Explainable AI Building trustworthy AI models? Raheel Ahmad Co-Founder @ explainX.ai
- 2. Do you want AI to work in your company?
- 3. Are you confident about your AI performance? Do you have the required visibility into it? Biased? Accuracy? Trustworthy?
- 4. By 2022, according to Gartner, 85% of AI projects will deliver erroneous outcomes due to bias in data, algorithms and lack of interpretability & trust in most AI models.
- 5. Image Credit: Gorodenkoff / Shutterstock.com
- 8. - GDPR concerns around lack of explainability in AI (European Commission) - “Companies should commit to ensuring systems that could fall under GDPR, including AI, will be compliant. The threat of sizeable fines of 20 Million Euros or 4% of global turnover provides a sharp incentive.” (GDPR) - “Article 22 of GDPR empowers individuals with the right to demand an explanation of how an AI system made a decision that affects them.” (GDPR) - Growing Global AI Regulation - “Algorithmic Accountability Act 2019: Requires companies to provide an assessment of the risks posed by the automated decision system to the privacy or security and the risks that contribute to inaccurate, unfair, biased, or discriminatory decisions impacting consumers.” - “Washington Bill 1655: Establishes guidelines for the use of automated decision systems to protect consumers, improve transparency, and create more market predictability.”
- 9. MOST MODELS ARE BLACKBOXES! AI Model❖ No Visibility ❖ No Explanations ❖ No Monitoring Business User Can I trust our AI decisions? Customer Support How do I answer customer query? Data Scientist Is my model accurate & trustworthy? IT & Operations How do I monitor & debug my model? Regulators Is the AI model fair?
- 10. So in reality, AI teams are flying BLIND!
- 11. How can we make interpretability more accessible to data scientists, ML engineers & researchers?
- 12. explainX.ai OPEN SOURCE MODEL INTERPRETABILITY FRAMEWORK FOR MODEL DEVELOPERS GLOBAL MODEL EXPLAINABILITY LOCAL PREDICTION EXPLANATIONS FEATURES ANALYSIS DATA DISTRIBUTIONS
- 13. ATTRIBUTION ALGORITHMS + Integrated Gradients + DeepLift + SHAP + LIME FEATURE INTERACTION METHODS + ALEs & PDPs + Auto-Bias Detection + What-If Analysis + Data Distributions RULES-BASED ALGORITHMS + Prototypes + DICE (Counterfactuals) + Anchors (Scoped Rules) + Decision Trees explainX.ai Platform + API Call from Jupyter Notebook/IDE + Interactive Visualizations + Sharing dashboards & actionable insights PRODUCT - ARCHITECTURE AI MODELS MODEL EXPLANATIONS MODEL DEBUGGING MODEL MONITORING DATA EXPLAINABLE AI ENGINE USER LAYER RESULT LAYER
- 14. HOW CAN WE VISUALIZE & EXPLAIN COMPLEX MODELS?
- 15. SINGLE API CALL PRODUCT: explainX.ai v1.0 WEB APPLICATION ACCESSIBLE WITH A SINGLE API CALL WATCH MVP DEMO
- 16. SINGLE API CALL PRODUCT: explainX.ai v1.0 Any ML Model Interactive and detailed explanations One Line of Code In just a single line of code, data scientists can integrate our explainability module into their personal workspace EXPLAIN YOUR BLACKBOX MODEL NOW
- 17. HOW CAN WE VISUALIZE & EXPLAIN COMPLEX MODELS?
- 18. VISUALIZING ATTRIBUTIONS OF A BLACK-BOX MODEL e.g. XGBoost, KNN
- 19. SIMULATING MODEL BEHAVIOR BY TESTING VALUES IN A WHAT-IF FORM
- 20. EVALUATING MODEL PERFORMANCE ACROSS MULTIPLE COHORTS
- 21. FEATURE INTERACTIONS USING A PDP PLOT
- 22. CHECK DISTRIBUTIONS FOR STATISTICALLY TROUBLESHOOTING DATA
- 23. PROTOTYPICAL ANALYSIS Dan Deny Alex Deny 80% Similar Bob Deny 75% Similar Julia Deny 70% Similar
- 24. COUNTERFACTUAL ANALYSIS Dan Deny Credit Score: 500 → 600 Risk Estimate: 70 → 50 Job Years: 2 → 3 Dan Approved
- 25. How can users take action from explanations? MONITOR BIAS & PERFORMANCE EXPLAIN MODEL DECISIONS BUILD TRUST IN MODEL LOGIC MAKE INFORMED DECISIONS
- 26. CURRENT VERSION TRY IT OUT! FUTURE PIPELINE Feature attribution based algorithms. Prototypes & examples based algorithms Feature interaction plots An interactive platform for model understanding & debugging https://explainx.ai/register Adding more algorithms for counterfactuals, anchors and bias detection. Scaling for enterprise adoption Building support for complicated deep learning models Real-time monitoring
- 27. Thank you. Questions? Get in touch at raheel@explainx.ai Get started with explainX open-source