Description
Aim
Explainable AI (XAI) Program teaches methods to interpret ML models and communicate results clearly. Learn explainability tools, validation checks, and reporting practices for reliable, human-trustable AI.
Program Objectives
- XAI Basics: why explainability matters and common risks.
- Model Types: interpretable models vs black-box models.
- Global Explanations: feature importance, partial dependence (intro).
- Local Explanations: SHAP and LIME concepts + use cases.
- Debugging: leakage, bias, drift, spurious correlations.
- Fairness: bias checks and group-wise performance (intro).
- Reporting: explanation + limitations + monitoring plan.
- Capstone: explain and audit a real ML model.
Program Structure
Module 1: Why Explainable AI?
- Trust, safety, and compliance in AI systems.
- When explanations are required: high-stakes decisions.
- What explanations can/can’t prove.
- XAI workflow: model → explain → validate → report.
Module 2: Interpretable Models First
- Linear/logistic regression interpretation.
- Decision trees and rule-based models.
- Monotonic constraints (intro) and simple models that work well.
- Choosing interpretable baselines before complex models.
Module 3: Global Explainability
- Permutation importance and gain-based importance.
- Partial dependence and ICE plots (conceptual + practice).
- Interaction effects (intro).
- Stability of explanations across folds/samples.
Module 4: Local Explainability (Instance-Level)
- SHAP basics: additive explanations and common plots.
- LIME basics: local surrogate explanations (intro).
- Case-based explanations: similar examples (overview).
- When local explanations mislead and how to reduce risk.
Module 5: Explaining Text and Images (Intro)
- NLP explanations: token importance, perturbation concepts.
- Vision explanations: saliency/Grad-CAM concept (overview).
- Choosing explanation methods based on task.
- Limitations and failure modes.
Module 6: Debugging Models with XAI
- Spot leakage and shortcut learning.
- Detect spurious features and data artifacts.
- Bias checks: group metrics and error breakdown.
- Counterfactual thinking: “what needs to change?” (intro).
Module 7: Governance, Monitoring & Documentation
- Model cards: purpose, data, metrics, limits.
- Data drift and concept drift monitoring (overview).
- Human review loops for high-risk use cases.
- Audit-ready documentation checklist.
Module 8: Capstone Build Sprint
- Pick a model: churn, credit risk (non-regulated demo), healthcare (non-clinical), HR, demand.
- Run explainability + bias checks + error analysis.
- Create a final report with visuals and recommendations.
- Prepare a short presentation for stakeholders.
Final Project
- Deliverables: model + explanation notebook + dashboard/plots + audit report.
- Include: risks, limits, monitoring plan, and decision guidance.
Participant Eligibility
- Students and professionals in data science, AI/ML, analytics
- Basic Python + ML fundamentals recommended
- Anyone working on responsible AI or model governance
Program Outcomes
- Generate global and local explanations for ML models.
- Use XAI to debug leakage, bias, and spurious patterns.
- Create stakeholder-ready XAI reports and documentation.
- Deliver an explainable model audit as a portfolio project.
Program Deliverables
- e-LMS Access: lessons, notebooks, datasets.
- XAI Toolkit: SHAP/LIME templates, reporting checklist, model card template.
- Capstone Support: feedback and review.
- Assessment: certification after capstone submission.
- e-Certification and e-Marksheet: digital credentials on completion.
Future Career Prospects
- Responsible AI / Model Governance Analyst
- ML Engineer (Explainability Focus)
- AI Risk & Compliance Associate
- Data Scientist (Model Interpretability)
Job Opportunities
- Finance: credit scoring explainability and monitoring.
- Healthcare/Pharma: model validation and audit workflows (non-clinical analytics).
- HR/Operations: transparent decision models and fairness checks.
- Tech/IT: ML platforms, governance, and responsible AI teams.
