Merge pull request #186399 from lgayhardt/azml-r-ai-0122

 Responsible AI docs  Arcolinx

Author: PRMerger5

articles/machine-learning/concept-differential-privacy.md

@@ -36,15 +36,15 @@ Differential privacy tries to protect against the possibility that a user can pr
 
 Epsilon values are non-negative. Values below 1 provide full plausible deniability. Anything above 1 comes with a higher risk of exposure of the actual data. As you implement machine learning solutions with differential privacy, you want to data with epsilon values between 0 and 1.
 
-Another value directly correlated to epsilon is **delta**. Delta is a measure of the probability that a report is not fully private. The higher the delta, the higher the epsilon. Because these values are correlated, epsilon is used more often.
+Another value directly correlated to epsilon is **delta**. Delta is a measure of the probability that a report isn’t fully private. The higher the delta, the higher the epsilon. Because these values are correlated, epsilon is used more often.
 
 ## Limit queries with a privacy budget
 
 To ensure privacy in systems where multiple queries are allowed, differential privacy defines a rate limit. This limit is known as a **privacy budget**. Privacy budgets prevent data from being recreated through multiple queries. Privacy budgets are allocated an epsilon amount, typically between 1 and 3 to limit the risk of reidentification. As reports are generated, privacy budgets keep track of the epsilon value of individual reports as well as the aggregate for all reports. After a privacy budget is spent or depleted, users can no longer access data. 
 
 ## Reliability of data
 
-Although the preservation of privacy should be the goal, there is a tradeoff when it comes to usability and reliability of the data. In data analytics, accuracy can be thought of as a measure of uncertainty introduced by sampling errors. This uncertainty tends to fall within certain bounds. **Accuracy** from a differential privacy perspective instead measures the reliability of the data, which is affected by the uncertainty introduced by the privacy mechanisms. In short, a higher level of noise or privacy translates to data that has a lower epsilon, accuracy, and reliability. 
+Although the preservation of privacy should be the goal, there’s a tradeoff when it comes to usability and reliability of the data. In data analytics, accuracy can be thought of as a measure of uncertainty introduced by sampling errors. This uncertainty tends to fall within certain bounds. **Accuracy** from a differential privacy perspective instead measures the reliability of the data, which is affected by the uncertainty introduced by the privacy mechanisms. In short, a higher level of noise or privacy translates to data that has a lower epsilon, accuracy, and reliability. 
 
 ## Open-source differential privacy libraries
 
@@ -72,7 +72,7 @@ The system library provides the following tools and services for working with ta
 |---------|---------|
 |Data Access     | Library that intercepts and processes SQL queries and produces reports. This library is implemented in Python and supports the following ODBC and DBAPI data sources:<ul><li>PostgreSQL</li><li>SQL Server</li><li>Spark</li><li>Preston</li><li>Pandas</li></ul>|
 |Service     | Execution service that provides a REST endpoint to serve requests or queries against shared data sources. The service is designed to allow composition of differential privacy modules that operate on requests containing different delta and epsilon values, also known as heterogeneous requests. This reference implementation accounts for additional impact from queries on correlated data. |
-|Evaluator     | Stochastic evaluator that checks for privacy violations, accuracy, and bias. The evaluator supports the following tests: <ul><li>Privacy Test -  Determines whether a report adheres to the conditions of differential privacy.</li><li>Accuracy Test - Measures whether the reliability of reports falls within the upper and lower bounds given a 95% confidence level.</li><li>Utility Test - Determines whether the confidence bounds of a report are close enough to the data while still maximizing privacy.</li><li>Bias Test - Measures the distribution of reports for repeated queries to ensure they are not unbalanced</li></ul> |
+|Evaluator     | Stochastic evaluator that checks for privacy violations, accuracy, and bias. The evaluator supports the following tests: <ul><li>Privacy Test -  Determines whether a report adheres to the conditions of differential privacy.</li><li>Accuracy Test - Measures whether the reliability of reports falls within the upper and lower bounds given a 95% confidence level.</li><li>Utility Test - Determines whether the confidence bounds of a report are close enough to the data while still maximizing privacy.</li><li>Bias Test - Measures the distribution of reports for repeated queries to ensure they aren’t unbalanced</li></ul> |
 
 ## Next steps
 

articles/machine-learning/concept-fairness-ml.md

@@ -25,7 +25,7 @@ Two common types of AI-caused harms are:
 
 - Harm of allocation: An AI system extends or withholds opportunities, resources, or information for certain groups. Examples include hiring, school admissions, and lending where a model might be much better at picking good candidates among a specific group of people than among other groups.
 
-- Harm of quality-of-service: An AI system does not work as well for one group of people as it does for another. As an example, a voice recognition system might fail to work as well for women as it does for men.
+- Harm of quality-of-service: An AI system doesn’t work as well for one group of people as it does for another. As an example, a voice recognition system might fail to work as well for women as it does for men.
 
 To reduce unfair behavior in AI systems, you have to assess and mitigate these harms.
 
@@ -45,12 +45,12 @@ Together, these components enable data scientists and business leaders t
 
 ## Assess fairness in machine learning models
 
-In the Fairlearn open-source package, fairness is conceptualized though an approach known as **group fairness**, which asks: Which groups of individuals are at risk for experiencing harms? The relevant groups, also known as subpopulations, are defined through **sensitive features** or sensitive attributes. Sensitive features are passed to an estimator in the Fairlearn open-source package as a vector or a matrix called  `sensitive_features`. The term suggests that the system designer should be sensitive to these features when assessing group fairness. 
+In the Fairlearn open-source package, fairness is conceptualized through an approach known as **group fairness**, which asks: Which groups of individuals are at risk for experiencing harms? The relevant groups, also known as subpopulations, are defined through **sensitive features** or sensitive attributes. Sensitive features are passed to an estimator in the Fairlearn open-source package as a vector or a matrix called  `sensitive_features`. The term suggests that the system designer should be sensitive to these features when assessing group fairness. 
 
 Something to be mindful of is whether these features contain privacy implications due to private data. But the word "sensitive" doesn't imply that these features shouldn't be used to make predictions.
 
 >[!NOTE]
-> A fairness assessment is not a purely technical exercise.  The Fairlearn open-source package can help you assess the fairness of a model, but it will not perform the assessment for you.  The Fairlearn open-source package helps identify quantitative metrics to assess fairness, but developers must also perform a qualitative analysis to evaluate the fairness of their own models.  The sensitive features noted above is an example of this kind of qualitative analysis.     
+> A fairness assessment is not a purely technical exercise.  The Fairlearn open-source package can help you assess the fairness of a model, but it will not perform the assessment for you.  The Fairlearn open-source package helps identify quantitative metrics to assess fairness, but developers must also perform a qualitative analysis to evaluate the fairness of their own models.  The sensitive features noted above is an example of this kind of qualitative analysis.
 
 During assessment phase, fairness is quantified through disparity metrics. **Disparity metrics** can evaluate and compare model's behavior across different groups either as ratios or as differences. The Fairlearn open-source package supports two classes of disparity metrics:
 
@@ -70,7 +70,7 @@ During assessment phase, fairness is quantified through disparity metrics. **Dis
 
 ### Parity constraints
 
-The Fairlearn open-source package includes a variety of unfairness mitigation algorithms. These algorithms support a set of constraints on the predictor's behavior called **parity constraints** or criteria. Parity constraints require some aspects of the predictor behavior to be comparable across the groups that sensitive features define (e.g., different races). The mitigation algorithms in the Fairlearn open-source package use such parity constraints to mitigate the observed fairness issues.
+The Fairlearn open-source package includes a variety of unfairness mitigation algorithms. These algorithms support a set of constraints on the predictor's behavior called **parity constraints** or criteria. Parity constraints require some aspects of the predictor behavior to be comparable across the groups that sensitive features define (for example, different races). The mitigation algorithms in the Fairlearn open-source package use such parity constraints to mitigate the observed fairness issues.
 
 >[!NOTE]
 > Mitigating unfairness in a model means reducing the unfairness, but this technical mitigation cannot eliminate this unfairness completely.  The unfairness mitigation algorithms in the Fairlearn open-source package can provide suggested mitigation strategies to help reduce unfairness in a machine learning model, but they are not solutions to eliminate unfairness completely.  There may be other parity constraints or criteria that should be considered for each particular developer's machine learning model. Developers using Azure Machine Learning must determine for themselves if the mitigation sufficiently eliminates any unfairness in their intended use and deployment of machine learning models.  
@@ -88,8 +88,8 @@ The Fairlearn open-source package supports the following types of parity constra
 
 The Fairlearn open-source package provides postprocessing and reduction unfairness mitigation algorithms:
 
-- Reduction: These algorithms take a standard black-box machine learning estimator (e.g., a LightGBM model) and generate a set of retrained models using a sequence of re-weighted training datasets. For example, applicants of a certain gender might be up-weighted or down-weighted to retrain models and reduce disparities across different gender groups. Users can then pick a model that provides the best trade-off between accuracy (or other performance metric) and disparity, which generally would need to be based on business rules and cost calculations.  
-- Post-processing: These algorithms take an existing classifier and the sensitive feature as input. Then, they derive a transformation of the classifier's prediction to enforce the specified fairness constraints. The biggest advantage of threshold optimization is its simplicity and flexibility as it does not need to retrain the model. 
+- Reduction: These algorithms take a standard black-box machine learning estimator (for example, a LightGBM model) and generate a set of retrained models using a sequence of re-weighted training datasets. For example, applicants of a certain gender might be up-weighted or down-weighted to retrain models and reduce disparities across different gender groups. Users can then pick a model that provides the best trade-off between accuracy (or other performance metric) and disparity, which generally would need to be based on business rules and cost calculations.  
+- Post-processing: These algorithms take an existing classifier and the sensitive feature as input. Then, they derive a transformation of the classifier's prediction to enforce the specified fairness constraints. The biggest advantage of threshold optimization is its simplicity and flexibility as it doesn’t need to retrain the model.
 
 | Algorithm | Description | Machine learning task | Sensitive features | Supported parity constraints | Algorithm Type |
 | --- | --- | --- | --- | --- | --- |

articles/machine-learning/concept-responsible-ml.md

@@ -19,7 +19,7 @@ In this article, you'll learn what responsible AI is and ways you can put it int
 
 ## Responsible AI principles
 
-Throughout the development and use of AI systems, trust must be at the core. Trust in the platform, process, and models. At Microsoft, responsible AI with regards tomachine learning encompasses the following values and principles:
+Throughout the development and use of AI systems, trust must be at the core. Trust in the platform, process, and models. At Microsoft, responsible AI with regard to machine learning encompasses the following values and principles:
 
 - Understand machine learning models
   - Interpret and explain model behavior
@@ -49,7 +49,7 @@ Unfairness in AI systems can result in the following unintended consequences:
 - Withholding opportunities, resources or information from individuals.
 - Reinforcing biases and stereotypes.
 
-Many aspects of fairness cannot be captured or represented by metrics. There are tools and practices that can improve fairness in the design and development of AI systems.
+Many aspects of fairness can’t be captured or represented by metrics. There are tools and practices that can improve fairness in the design and development of AI systems.
 
 Two key steps in reducing unfairness in AI systems are assessment and mitigation. We recommend [FairLearn](https://github.com/fairlearn/fairlearn), an open-source package that can assess and mitigate the potential unfairness of AI systems. To learn more about fairness and the FairLearn package, see the [Fairness in ML article](./concept-fairness-ml.md).
 
@@ -75,7 +75,7 @@ See the following sample to learn [how to deploy an encrypted inferencing web se
 
 Documenting the right information in the machine learning process is key to making responsible decisions at each stage. Datasheets are a way to document machine learning assets that are used and created as part of the machine learning lifecycle.
 
-Models tend to be thought of as "opaque boxes" and often there is little information about them. Because machine learning systems are becoming more pervasive and are used for decision making, using datasheets is a step towards developing more responsible machine learning systems.
+Models tend to be thought of as "opaque boxes" and often there’s little information about them. Because machine learning systems are becoming more pervasive and are used for decision making, using datasheets is a step towards developing more responsible machine learning systems.
 
 Some model information you might want to document as part of a datasheet:
 
@@ -84,7 +84,7 @@ Some model information you might want to document as part of a datasheet:
 - Training data used
 - Evaluation data used
 - Training model performance metrics
-- Fairness information.
+- Fairness information
 
 See the following sample to learn how to use the Azure Machine Learning SDK to implement [datasheets for models](https://github.com/microsoft/MLOps/blob/master/pytorch_with_datasheet/model_with_datasheet.ipynb).