• Contents:

Clarify (part of SageMaker Studio)

• 🔧 Evaluate and compare FMs, explain model outputs, detect bias
  • Built-in or custom datasets
  • Can evaluate human factors such as friendliness or humor
• Features
  • Built-in metrics and algorithms
  • Leverage an AWS-managed team or bring your own employees
• Screenshot

Model Explainability

• Set of tools to explain how ML models make predictions
  • Can help understand model prior to deployment
  • Can help debug model predictions after deployment
  • Increases trust and understanding of the model
• Examples
  • “Why did the model predict a negative outcome such as a loan rejection for a given applicant?”
  • “Why did the model make an incorrect prediction?”
• Screenshot

Detect (Human) Bias

• Detect and explain biases in your datasets and models
• Measure bias using statistical metrics
• Specify input features and bias will be automatically detected
• Kinds of biases
  1. Sampling bias
     • Training data does not represent the full population fairly
     • Leads to model that over-represents or disproportionately affects certain groups
     • 💡 Example: an algorithm only flags people from specific ethnic groups. This is probably a sampling bias, and you need to perform data augmentation for imbalanced classes
  2. Measurement bias
     • Tools or measurements used in data collection are flawed or skewed
  3. Observer bias
     • Person collecting or interpreting the data has personal biases that affect the results
  4. Confirmation bias
     • Individuals interpret or favor information that confirms their preconceptions
     • ❗More applicable to human decision-making rather than automated model outputs
• Screenshot

Ground Truth

• 🔧 RLHF – Reinforcement Learning from Human Feedback
  • Model review, customization and evaluation by humans
  • Aligns model to human preferences
  • Reinforcement learning where human feedback is included in the “reward” function