Domain 4: Responsible AI Practices

Contents


Fainess measured by the bias and variance of outcomes across groups

Effects of bias and variance

  • Demographic disparities: Unequal outcomes or treatment of different groups because of biased models
  • Inaccuracy: Reduced model performance and reliability because of bias or variance issues
  • Overfitting: Models that are too complex and fail to generalize well to new data
  • Underfitting: Models that are too simple and fail to capture underlying patterns in the data
  • User trust: Eroding confidence in Al systems because of biased or inconsistent outputs
     
  • Bias-variance trade-off

Challenges of Generative AI

  • Toxicity (generate inappropriate content)
  • Hallucinations (inaccurate responses)
  • Intellectual property
  • Plagiarism and cheating
  • Disruption of the nature of work

Core Dimensions of Responsible AI (features of responsible AI)

  • Fairness - How a system impacts different subpopulations of users (for example, by gender, ethnicity)
  • Explainability - Mechanisms to understand and evaluate the outputs of an Al system
  • Privacy and security - Data that is used in accordance with privacy considerations and protected from theft and exposure
  • Transparency - Information about an Al system so stakeholders can make informed choices about their use of the system
  • Veracity and robustness - Mechanisms to ensure that an Al system operates reliably
  • Governance - Processes to define, implement, and enforce responsible Al practices within an organization
  • Safety
  • Controllability

Developing Responsible AI Systems

Reviewing Amazon service tools for responsible AI

  • Foundation model evaluation
    • Model evaluation on Amazon Bedrock - evaluate, compare, and select between FMs
    • SageMaker Clarify - evaluate FMs for metrics like accuracy, robustness, and toxicity
  • Safeguards for generative AI
    • Amazon Bedrock Guardrails - implement safeguards
  • Bias detection
    • Amazon SageMaker Clarify - identify potential bias in ML models and dataset without extensive coding
    • Amazon SageMaker Data Wrangler - import, prepare, transform, balance data if imbalance
  • Model prediction explanation
    • SageMaker Clarify in integrated with Amazon SageMaker Experiments
  • Monitoring and human reviews
    • Amazon SageMaker Model Monitor - monitors models in production
    • Amazon Augmented AI (Amazon A2I) - workflow for human review on ML predictions, low confidence/random predictions sent for human review
  • Governance improvement
    • Amazon SageMaker Role Manager - define minimum permissions
    • Amazon SageMaker Model Cards - model behavior in production, all in one place
      • Catalog details include information such as the intended use and risk rating of a model, training details and metrics, evaluation results and observations, and additional callouts such as considerations, recommendations, and custom information.
  • Providing transparency
    • AWS AI Service Cards - responsible AI documentation for AWS AI services
      • Basic concepts to help customers better understand the service or service features
      • Intended use cases and limitations
      • Responsible AI design considerations
      • Guidance on deployment and performance optimization

Responsible datasets are the foundation of Responsible AI

Characteristics of datasets

  • Inclusivity: Representing diverse populations, perspectives, and experiences in training data
  • Diversity: Incorporating a wide range of attributes, features, and variables to avoid bias
  • Balanced datasets: Ensuring equal representation of different groups and avoiding skewed distributions
  • Privacy protection: Safeguarding sensitive information and adhering to data protection regulations
  • Consent and transparency: Obtaining informed consent from data subjects and providing clear information about data usage
  • Regular audits: Conducting periodic reviews of datasets to identify and address potential issues or biases

Responsible Practices to Select a Model

  • use Model evaluation on Amazon Bedrock or SageMaker Clarify to evaluate models for accuracy, robustness, toxicity, or nuanced content that requires human judgement.
     
  • Environmental considerations: Assessing the carbon footprint and energy consumption of Al models
  • Sustainability: Prioritizing models with minimal environmental impact and long-term viability
  • Transparency: Providing clear information about model capabilites, limitations, and potential risks
  • Accountability: Establishing clear lines of responsibility for Al model outcomes and decision making
  • Stakeholder engagement: Involving diverse perspectives in model selection and deployment processes
     
  • Define application use case narrowly
  • Choosing a model based on performance
    • Level of customization – The ability to change a model’s output with new data ranging from prompt-based approaches to full model retraining
    • Model size – The amount of information the model has learned as defined by parameter count
    • Inference options – From self-managed deployment to API calls
    • Licensing agreements – Some agreements can restrict or prohibit commercial use
    • Context windows – The amount of information that can fit in a single prompt
    • Latency – The amount of time it takes for a model to generate an output
  • Choosing a model based on sustainability concerns (socially, environmentally, and economically sustainable over the long term.)
    • Responsible agency considerations for selecting a model
      • Value alignment
      • Responsible reasoning skills
      • Appropriate level of autonomy
      • Transparency and accountability
    • Environmental considerations for selecting a model
      • Energy consumption
      • Resources utilization
      • Environmental impact assessment
    • Economic considerations for selecting a model (impact on jobs)

Responsible Preparation for Datasets

  • use SageMaker Clarify and SageMaker Data Wrangler to help balance your datasets.
     
  • Balancing datasets
    • Inclusive and diverse data collection
    • Data curation
      • Data preprocessing
      • Data augmentation
      • Regular auditing

Transparent and Explainable Models

Advantages over black box models

  • Increased trust
  • Easier to debug and optimize for improvements
  • Better understanding of the data and the model’s decision-making process

AWS tools for transparency

  • AWS AI Service Cards - Amazon provides transparent documentation on Amazon services that help you build your AI services.
    • Intended use cases and limitations
    • Responsible AI design considerations
    • Guidance on deployment and perforamnce optimization
  • Amazon SageMaker Model Cards - you can catalog and provide documentation on models that you create or develop yourself.
    • Model details automatically populated.

AWS tools for explainability

  • SageMaker Clarify
    • Feature attributions - how much each feature contributed for the model predictions.
    • Partial dependence plots - plot graph on models predications change for different values of feature
  • SageMaker Autopilot (how ML models make predictions)

Model Trade-Offs

  • Interpretability trade-offs
    • Interpretability
      • More transparent
      • Deep level understanding of internal mechanics
      • Uses interpretable algorithms
      • Performance and security tradeoffs
    • Explainability
      • Less transparent
      • High-level understanding
      • Model agnostic (black box) approach
  • Safety and transparency trade-offs
  • Model controllability
     
  • A model that provides transparency into a system so a human can explain the model’s output based on the weights and features is an example of interpretability in a model.
  • A model that uses model agnostic methods to explain the behavior of the model in human terms is an example of explainability in a model.
  • A model that avoids causing harm in its interactions with the world is an example of safety in a model.
  • A model that you can influence the predictions and behavior by changing aspects of the training data is an example of controllability in a model.

Principles of Human-Centered Design for Explainable AI

  • Design for amplified decision-making.
    • Clarity
    • Simplicity
    • Usability
    • Reflexivity
    • Accountability
  • Design for unbiased decision-making.
    • Transparency
    • Fairness
    • Training
  • Design for human and AI learning.
    • Cognitive apprenticeship
    • Personalization
    • User-centered design
       

  • Amazon Augmented AI (Amazon A2I) - workflow for human review on ML predictions, low confidence/random predictions sent for human review

  • Reinforcement learning from human feedback (RLHF)
    • incorporated human feedback in the rewards function
    • Amazon SageMaker Ground Truth - humans involved for making high value data sets, incorporating human feedback across the ML lifecycle