Responsible AI Practices

Question 1

What is Responsible AI?

Answer 1

Refers to practices and principles
Ensures AI systems are transparent and trustworthy
Mitigates potential risks and negative outcomes
Applicable throughout the lifecycle of an AI application - Design, Development, Deployment, Monitoring, Evaluation etc.
Applicable to both traditional and generative AI.

Question 2

What are the challenges of Responsible AI?

Answer 2

Bias - predictions that are biased against historically unfavored groups.
This can arise from:
* Data Bias - training data is biased or underrepresents certain groups
* Algorithm Bias - assumptions and simplifications of the model to optimize performance may lead to bias
* Interaction bias - arise from the way humans interact with the system and the context of deployment - e.g. face recognition tested on certain groups may perform poorly for other groups.
* Bias Amplification - existing societal biases amplified

Question 3

How can bias be mitigated?

Answer 3

1. Ensure training data is diverse and representative
2. Audit algorithms for potential bias
3. Incorporating fairness metrics and constraints in the development process
4. Promoting transparency and explainability
5. Involving diverse stakeholders in AI development

Question 4

What are the unique challenges of Generative AI?

Answer 4

1. Toxicity - content that is offensive, disturbing, and inappropriate
2. Hallucinations - assertions or claims that sound plausible but are verifiably incorrect. - e.g. non existing scientific citations
3. Intellectual property - LLMs can sometimes produce text or code verbatim from training data. Even when it content original, there may be copyright issues -e.g. a cat in the style of Picasso
4. Plagiarism and Cheating - e.g. college essays written by Gen AI
5. Disruption of the nature of work - worry that some professions may become obsolete.

Question 5

What are the core dimensions of responsible AI?

Answer 5

Mnemonic: Friendly Elephants Prefer Vast Tropical Grasslands, Staying Cool.

F - Fairness
E - Explainability
P - Privacy and Security
V - Veracity and Robustness
T - Transparency
G - Governance
S - Safety
C - Controllability

Question 6

What is Fairness?

Answer 6

Promotes inclusion, minimizes discriminatory output

Question 7

What is Explainability?

Answer 7

Humans must be able to understand how the model makes decisions.

Question 8

What does Privacy and Security involve?

Answer 8

Data is protected from theft and disclosure
Individuals control if and when their data is used.
No unauthorized users can have access to an individual’s data

Question 9

What is transparency?

Answer 9

How information about an AI system is communicated to its stakeholders.
This helps stakeholders make informed choices about how they use the system.
e.g. Information on development process, capabilities, and limitations of the system, types of testing performed.
E.g. AI model cards

Question 10

What is veracity and robustness?

Answer 10

Reliability even in unexpected situations, uncertainty, and errors.
Resilience to changes in input parameters, data distributions

Question 11

What is Governance?

Answer 11

Set of processes that define, implement, and enforce responsible AI practices within an organization.
Governance policies used to enforce compliance with regulatory obligations.

Question 12

What is Safety?

Answer 12

Develop AI in a way that is safe and beneficial for individuals and society as a whole.
Use of guardrails

Question 13

What is controllability?

Answer 13

Monitor and guide an AI system’s behavior to align with human values.

Question 14

Business benefits of Responsible AI?

Answer 14

Increased trust and reputation
Regulatory compliance
Mitigate risks such as bias, privacy violations, security breaches
Competitive edge
Improved decision making
Improved products and businesses

Question 15

What are Amazon services and tools that help with responsible AI?

Answer 15

1. Model Evaluation on Bedrock - automatic and human evaluation
   * Automatic Evaluation - predefined metrics for accuracy, robustness and toxicity
   * Human evaluation - evaluates friendliness, style, and brand alignment. Either customer employees or AWS-managed team
2. SageMaker Clarify - does model evaluation; can detect bias (e.g. gender and age); use of workforce to review model responses.
3. Guardrails for Amazon Bedrock - filter out undesirable content, PII
4. SageMaker Data Wrangler - corrects data imbalances (random undersampling/oversampling etc. to rebalance data)
5. SageMaker Clarify - provides explainability metrics -e.g. which feature contributed most to the prediction
6. Monitoring and human reviews - SageMaker Model Monitor - monitors the quality of SM ML models. You can set up SM to run jobs to detect deviations in model performance.
7. Amazon Augmented AI (A2I) - coordinates human review workflows of ML predictions; managing large number of human reviewers.
8. Amazon SageMaker tools for Governance include Role Manager, Model Cards, and Model Dashboard (to monitor model behavior in production)
9. AWS AI Service Cards - helps understand AI services

Question 16

What is the structure of AWS AI Service Cards?

Answer 16

AWS AI Service Cards have four sections:
a) Basic Concepts to help understand the service and the features
b) Intended use cases and limitations
c) Responsible AI design considerations
d) Guidance on deployment and performance optimization

Question 17

What are responsible considerations to select a model?

Answer 17

• Model selection is very important. Evaluate the model (Amazon Bedrock Evaluation or SM Clarify)
• Define application use case narrowly - helps tune the model for the specific use case (e.g. Favoring precision over recall for certain use cases with Amazon Rekognition).
• AI applications for cataloging a product vs. persuading to buy
• e.g. Persuading to buy - e.g. targeting people on the coast for beach products.

Question 18

What are some of the factors to consider when choosing a model?

Answer 18

1. Level of customization - ability to change a model’s output using prompts and fine tuning.
2. Model size - how many parameters it can support
3. Inference options - self deployment or API calls?
4. Licensing agreements
5. Context window - how big is the prompt window.
6. Latency - time to generate output

Generally, a model’s performance is not only a function of its algorithm but also the training data set.

Question 19

What are responsible agency considerations when choosing a model?

Answer 19

Responsible Agency means a AI’s systems capacity to make good judgement and act in a socially responsible manner.

1. Value alignment - AI system’s goals should be inline with responsible human values.
2. Responsible reasoning skills - should be able to logically think through moral dilemmas, understand moral concepts; should be able to apply these principles to new situations.
3. Right level of autonomy - with clear boundaries and human oversight
4. Transparency and accountability - in decision making

These are all desirable targets which current AI systems cannot yet meet.

Question 20

What are the RAI capabilities of SageMaker ?

Answer 20

1. SM Clarify - Bias Detection in models and datasets - e.g. specify input features such as age or gender and SMC can identify any bias. Can also check for accuracy, robustness, and toxicity.
   2.SM Data Wrangler - balance data in case of imbalances - undersampling, oversampling and SMOTE.
2. SM Clarify Explainability - when combined with SM Experiments, for tabular, NLP, and computer vision model, SMC can identify features of the data which most contributed to the decision. This helps explainability.
3. SM Model Monitor - monitor the quality of SM ML models; setup notifications in case of model deviations
4. SM Amazon Augmented AI (Amazon A2I) - workflows required for human review of ML predictions.
5. SM Autopilot - uses tools provided by SMC to provide insights into how Model make predictions. Can determine the contribution of individual features or inputs to the model’s output and provides insights into the relevance of different features. provide per-instance explanation during inference
6. SM Groundtruth - human-in-the-loop capabilities for incorporating human feedback across the ML lifecycle to improve model accuracy and relevancy.

SM Governance Features include:
8. SM Role Manager - define permissions
9. SM Model Cards - build model cards (intended use, risk ratings, training details)
10. SM Model Dashboards (keep team informed of model behavior)

Question 21

What safeguards are available in Amazon Bedrock?

Answer 21

1. Guardrails for Bedrock - controls interaction between users and FMs.
   * Filters undesirable inputs and outputs - i.e. content filters (hate speech, sexual, violence, insults etc.)
   * Redacts PII
   * Guardrails are FM agnostic.
   * Customers can configure multiple GRs.

Question 22

What is random undersampling and ovesampling?

Answer 22

Undersampling - Random undersampling is a technique used in data analysis to balance datasets by randomly removing samples from the majority class.

Random oversampling is a sampling method that increases the number of minority class data points in a dataset by randomly selecting and duplicating existing samples from that class

Question 23

What does Responsible AI in dataset preparation mean?

Answer 23

1. Data is balanced and representative of diverse groups - diverse and inclusive
2. Data Curation - labeling, organizing, and preprocessing the data; ensures that the AI model is trained on high quality data. It involves:
   a) Data preprocessing - ensuring it is accurate, complete, and unbiased.
   b) Data augmentation - to compensate for under representation of groups
   c) Regular auditing - to ensure it is balanced and fair

Question 24

What’s the difference between Transparency and Explainability?

Answer 24

Transparency is HOW a model makes a decision
* Helps with accountability and trust.

Explainability is WHY the model made the decision it made.
* Insight into model limitation
* Informs users into how to use the model
* Helps with debug/troubleshooting

Question 25

What are the advantages of a transparent and explainable model over opaque ones?

Answer 25

1. Increased trust
2. Easier to debug and optimize for improvements.
3. Better understanding of the data and decision making process - in some use cases, this is more important than performance (e.g. healthcare).

Question 26

What are some potential solutions for creating transparent and explainable models?

Answer 26

1. Explainability Frameworks - e.g. SHAP, LIME and Counterfactual - these help summarize and interpret decision. They provide insight into factors that influenced a particular decision. They help assess fairness and consistence of AI systems.
2. Transparent documentation - that talk about AI architecture, data sources, training processes, and assumptions.
3. Monitoring and auditing - to ensure that there is no bias or discriminatory behavior.
4. Human oversight and involvement - human review decision in high stake situations
5. Counterfactual explanations - show how outputs can change if inputs were different - helps users understand how AI functions
6. User Interface explanations - UIs that provide information on how the AI system works, inputs, outputs, rationale, etc.

Question 27

What are the risks of transparency and explainability?

Answer 27

1. Increase in complexity and cost
2. Creating vulnerabilities that threat actors exploit
3. Presenting unrealistic expectations that the model is perfectly transparent and explainable
4. Too much info can give away IPR, create privacy and security concerns.

Question 28

What is variance?

Answer 28

Variance refers to the model’s sensitivity to fluctuations or noise in the training data.
Ultimately, you want variance to be low.

Question 29

What is overfitting and underfitting?

Answer 29

Overfitting is when model performs well on the training data but does not perform well on the evaluation data. This is because the model is memorizing the data it has seen and is unable to generalize to unseen examples. It has been over trained.

Underfitting is when the model does not capture enough features of the dataset.

Question 30

What is the bias-variance trade-off?

Answer 30

Bias-variance tradeoff is when you optimize your model with the right balance between bias and variance. This means that you need to optimize your model so that it is not underfitted or overfitted. The goal is to achieve a trained model with the lowest bias and lowest variance tradeoff for a given data set.

In an under-fitted model, bias is high, but variance is low. In an overfitted model, bias is low, but variance is high.

Question 31

What techniques can you use to overcome the bias-variance errors?

Answer 31

1. Cross-validation - train multiple models on a subset of data and test it on a different subset - detects overfitting
2. Increase data in the training set.
3. Regularization - penalization of extreme weight values
4. Use simpler models
5. Dimension reduction - reduce the dimensionality (number of features) within a dataset while still retaining as much information as possible
6. Stop training early - end training early so model does not memorize the data.

Question 32

What is Interpretability?

Answer 32

1. Interpretability is the degree to which a human can understand the cause of a decision.
2. A subset of model transparency.
3. Interpretability is the access into a system so that a human can interpret the model’s output based on the weights and features

Explainability is how to take an ML model and explain the behavior in human terms.

Model’s interpretability can affect performance - high performance model (e.g. Neural Networks) have poor interpretability, whereas lower ones like Linear Regression have high interpretability.

If performance is important then you may have to rely on explainability more than interpretability.

Question 33

What is model safety?

Answer 33

Model safety is the ability of an AI system to avoid causing harm in its interactions with the world. This includes avoiding social harm, such as bias in decision-making algorithms, and avoiding privacy and security vulnerability exposures.

Safety and transparency is a trade-off. Making a model more safe may make it less transparent.

Question 34

What is model controllability?

Answer 34

A controllable model is one where you can influence the model’s predictions and behavior by changing aspects of the training data. Higher controllability provides more transparency into the model and allows correcting undesired biases and outputs.

More complex models (link NNs) are harder to control than simpler ones like Linear Reg.

Question 35

In AI, what is human-centered design?

Answer 35

Human-centered design (HCD) is an approach to creating products and services that are intuitive, easy to use, and meet the needs of the people who will be using them.

In AI, this means, explanations and interfaces are clear.

Question 36

What are the key principles of HCD?

Answer 36

1. Design for amplified decision-making - support decision makers in high-stake situations; designing for clarity, simplicity, usability, reflexivity, and accountability.
2. Design for unbiased decision-making - decision-making processes, systems, and tools is free from biases that can influence the outcomes. Be transparent, minimize unfairness and discrimination,
3. Design for human and AI learning - focuses on creating better AI systems. e.g. AI learns from humans (e.g. RLHF where humans reward AI systems for correct predictions).