Study Notes

Appearance

Domain 4: Guidelines for Responsible AI (14%) ​

← Domain 3 · Next Domain →

4.1: Responsible AI Principles ​

Responsible AI Principles

1 / 3
❓

What is Fairness in AI?

(Click to reveal)
💡
Avoid bias and discrimination
Ensure AI treats all groups equitably
Test across demographic groups, use diverse training data.

1. Fairness ​

Avoid bias and discrimination

Types of Bias:

  • Training Data Bias: Unrepresentative data
  • Selection Bias: Biased sampling
  • Algorithmic Bias: Model amplifies existing bias

Mitigation:

  • Diverse training data
  • Test across demographic groups
  • Regular bias audits
  • Use SageMaker Clarify

2. Explainability ​

Understand how models make decisions

Why It Matters:

  • Build trust
  • Debug errors
  • Meet regulatory requirements
  • Identify bias

Techniques:

  • Feature importance
  • SHAP values
  • Attention visualization
  • Model cards

3. Privacy ​

Protect sensitive data

Best Practices:

  • Data encryption (at rest, in transit)
  • Access controls (IAM)
  • Data anonymization
  • Differential privacy
  • Secure data deletion

4. Safety ​

Prevent harmful outputs

Risks:

  • Toxic content generation
  • Misinformation
  • Dangerous instructions
  • Manipulation

Guardrails:

  • Content filtering
  • Human review (Amazon A2I)
  • Output validation
  • Rate limiting

5. Transparency ​

Document capabilities and limitations

Model Cards Should Include:

  • Intended use cases
  • Training data description
  • Known limitations
  • Performance metrics
  • Bias testing results

4.2: AWS Tools for Responsible AI ​

Amazon SageMaker Clarify ​

Detect bias in data and models

Responsible AI Tools

1 / 3
❓

What does SageMaker Clarify do?

(Click to reveal)
💡
Detect bias in data and models
Pre-training and post-training bias detection
Feature importance and explainability reports.

Features:

  • Pre-training bias detection
  • Post-training bias detection
  • Feature importance
  • Explainability reports

Bias Metrics:

  • Class imbalance
  • Disparate impact
  • Difference in positive proportions

Amazon SageMaker Model Monitor ​

Continuous monitoring

Monitors:

  • Data quality
  • Model quality (drift)
  • Bias drift
  • Feature attribution drift

Alerts:

  • CloudWatch alarms
  • SNS notifications

Amazon Augmented AI (A2I) ​

Human review workflows

Use Cases:

  • Review low-confidence predictions
  • Audit high-stakes decisions
  • Compliance requirements
  • Continuous improvement

← Domain 3 · Next Domain →