Responsible AI in Industry: Practical Challenges and Lessons Learned

Editor's Notes

• #2: In this tutorial, we will focus primarily on three dimensions: P/F/E. Prior to this tutorial, we had presented tutorials on privacy, on fairness, and on explainability. In particular, I would like to thank Timnit Gebru and Meg Mitchell for the realization of the fairness in industry tutorial. I had reached out to Timnit Gebru to find out potential collaborators from other companies – Timnit connected me with Meg. Today, we will present case studies from Amazon, LinkedIn, and Microsoft. Please refer the longer version of our tutorial for additional case studies, especially case studies from Google (which we would not be able to present today).
• #3: First, we will motivate the need for bias detection and mitigation in ML systems. Challenges that have received a lot of attention in the media, and have really highlighted how important it is to get AI right – to make sure that AI does not discriminate or further disadvantage already disadvantaged groups. Many of these stories have focused on high-stakes decisions where machine learning systems are used allocate opportunities, resources, or information in ways that can have significant negative impacts on people’s lives.
• #4: Recently, policymakers, regulators, and advocates have raised awareness about the ethical, policy, and legal challenges posed by machine learning and data-driven systems. In particular, they have expressed concerns about the potentially discriminatory impact of such systems, for example, due to inadvertent encoding of bias into automated decisions. “Do Google’s ‘unprofessional hair’ results show it’s racist?” by Leigh Alexander
• #5: There have been several laws in countries such as the United States that prohibit discrimination based on “protected attributes” such as race, gender, age, disability status, and religion. Many of these laws have their origins in the Civil Rights Movement in the United States (e.g., US Civil Rights Act of 1964). When legal frameworks prohibit the use of such protected attributes in decision making, there are usually two competing approaches on how this is enforced in practice: Disparate Treatment vs. Disparate Impact. Avoiding disparate treatment requires that such attributes should not be actively used as a criterion for decision making and no group of people should be discriminated against because of their membership in some protected group. Avoiding disparate impact requires that the end result of any decision making should result in equal opportunities for members of all protected groups irrespective of how the decision is made. Please see NeurIPS’17 tutorial titled Fairness in machine learning by Solon Barocas and Moritz Hardt for a thorough discussion.
• #6: Renewed focus in light of privacy breaches observed over the last several years. Example: EU General Data Protection Regulation (GDPR), which came into effect in May, 2018. The focus is not only on privacy of users, but also related dimensions such as algorithmic bias (or ensuring fairness), transparency, and explainability of decisions. Image credit: https://pixabay.com/en/gdpr-symbol-privacy-icon-security-3499380/
• #8: Challenges with scaling fairness, explainability, (and privacy) mechanisms to cater to the needs of AWS customers from financial services, healthcare, HR, and other industries: Providing functionality that caters to different customer personas Providing functionality that caters to the needs in different stages of ML lifecycle Starting with SageMaker, and then scaling the functionality to AI services and to other ML products & services across Amazon
• #9: From https://news.linkedin.com/about-us#Statistics as of 2021-02-01 Largest professional network in the world. It is a platform for every professional to tell their story. Who they are, where they work, skills, etc. Once you have done that, the platform works infusing intelligence by harnessing the power of this data to help connect talent with opportunity at scale. * 740M members in more than 200 countries and territories worldwide | 1.5K fields of study | 600+ degrees | 24K titles ... Our vision is to develop a profile for every member of the global workforce, all 3 billion of them, every employer in the world, and every open job at each of these companies, to provide every member of the global workforce with transparency into the skills required to obtain those jobs.  We want to build a profile for every educational institution or training facility that enables people to acquire those skills, and a publishing platform that enables every individual, every company, and every university to share their professionally-relevant knowledge if they’re interested in doing so.
• #21: TODO split slide into 3
• #32: TODO split slide into 3
• #39: Amazon SageMaker helps data scientists and developers to prepare, build, train, and deploy high-quality machine learning (ML) models quickly by bringing together a broad set of capabilities purpose-built for ML. Ignore notes below (I won’t go into details) Predictive Maintenance Predict if a component will fail before failure based on sensor data. Example applications include predicting failure and remaining useful life (RUL) of automotive fleets, manufacturing equipment, and IoT sensors. The key value is increased vehicle and equipment up-time and cost savings. This use case is widely used in automotive and manufacturing industries. Industries: Automotive, Manufacturing Georgia Pacific uses SageMaker to detect machine issues early. To learn more, read the case study.   Demand Forecasting Use historical data to forecast key demand metrics faster and make more accurate business decisions around production, pricing, inventory management, and purchasing/re-stocking. The key value is meeting customer demand, reducing inventory carrying costs by reducing surplus inventory, and reducing waste. This use case is used mainly in financial services, manufacturing, retail, and consumer packaged goods (CPG) industries. Industries: Financial Services (FSI), Manufacturing, Retail, Consumer Packaged Goods (CPG) Advanced Microgrid Solutions has built an ML model with SageMaker to forecast energy prices in near real time. Watch the re:Invent session. Fraud Detection Automate the detection of potentially fraudulent activity and flag it for review. The key value is reducing costs associated with fraud and maintaining customer trust. This use case is used mainly in financial services and online retail industries. Industries: FSI, Retail Euler Hermes uses SageMaker to catch suspicious domains. Learn more from the blog post. Credit Risk Prediction Explain individual predictions from a credit application to predict whether the credit will be paid back or not (often called a credit default). The key value is identifying bias and satisfying regulatory requirements. This use case is used mainly in financial services and online retail industries. Industries: FSI We have a Explaining Credit Decisions customized solution using SageMaker that can be used to explain individual predictions from machine learning models, including applications for credit decisions, churn prediction, medical diagnosis, and fraud detection. Extract & Analyze Data from Documents Understand text in written and digital documents and forms, extract information, and use it to classify items and make decisions. Industries: Healthcare, FSI, Legal, M&E, Education Computer Vision (image analysis) Main sub-use cases are: 1) Automatically medical diagnosis from X-ray and other imaging data; 2) Manufacturing quality control automation to detect defective parts; 3) Drug discovery; 4) Social distancing and tracking concentration of people for COVID-19 in public places. Industries: Healthcare/Pharma, Manufacturing, Public Sector Autonomous Driving Reinforcement learning and object detection algorithms. Industries: Automotive Personalized Recommendations Make personalized recommendations based on historical trends. Industries: M&E, Retail, Education (most likely classes to ensure graduation) Churn Prediction Predict customer likelihood to churn. Industries: Retail, Education, Software & Internet (SaaS)
• #40: Biases are imbalances in the training data or the prediction behavior of the model across different groups, such as age or income bracket. Biases can result from the data or algorithm used to train your model. For instance, if an ML model is trained primarily on data from middle-aged individuals, it may be less accurate when making predictions involving younger and older people. ----- Regulatory Compliance Regulations may require companies to be able to explain financial decisions and take steps around model risk management. Amazon SageMaker Clarify can help flag any potential bias present in the initial data or in the model after training and can also help explain which model features contributed the most to an ML model’s prediction. Internal Reporting & Compliance Data science teams are often required to justify or explain ML models to internal stakeholders, such as internal auditors or executives. Amazon SageMaker Clarify can provide data science teams with a graph of feature importance when requested and can help quantify potential bias in an ML model or the data used to train it in order to provide additional information needed to support internal requirements. Customer Service Customer-facing employees, such as financial advisors or loan officers, may review a prediction made by an ML model as part of the course of their work. Working with the data science team, these employees can get a visual report via API directly from Amazon SageMaker Clarify with details on which features were most important to a given prediction in order to review it before making decisions that may impact customers. ----- Ignore notes below (I will use description at https://aws.amazon.com/sagemaker/clarify) ML models, especially those that make predictions which serve end customers, are at risk of being biased and producing incorrect or harmful outcomes if proper precautions are not taken, making the ability to detect bias across the ML lifecycle critical. Let’s look at some of the ways bias may become present in a model: The initial data set you use for your model might contain imbalances, such as not having enough examples of members of a certain class which then cause the model to become biased against that class. Your model might develop biased behavior during the training process. For example, the model might use bank location as a positive indicator to approve a loan as opposed to actual financial data if one particular bank location approves more loans than others. In this case, the model has bias towards applicants that apply at that location and against applicants that do not, regardless of their financial standing. Finally, bias may develop over time if data in the real world begins to diverge from the data used to train your deployed model. For example, if your model has been trained on an outdated set of mortgage rates it may start to become biased against certain home loan applicants. But to understand WHY bias is present, we need explainability. And explainability is useful for more than just bias. Let me explain. Many regulators need to understand why the ML model made a given prediction and whether the prediction was free from bias, both in training and at inference You may need to provide explanations to internal teams (loan officers, customer service reps, compliance officers) in addition to end users / customers. For example a loan officer may need help explaining to a customer what factors caused their application to be denied. Finally, data science teams can improve models given a deeper understanding on whether a model is making the right inferences for the right reasons, or if perhaps irrelevant data points are being used that are altering model behavior.
• #41: Detect bias in your data and model Identify imbalances in data SageMaker Clarify is integrated with Amazon SageMaker Data Wrangler, making it easier to identify bias during data preparation. You specify attributes of interest, such as gender or age, and SageMaker Clarify runs a set of algorithms to detect any presence of bias in those attributes. After the algorithm runs, SageMaker Clarify provides a visual report with a description of the sources and measurements of possible bias so that you can identify steps to remediate the bias. For example, in a financial dataset that contains only a few examples of business loans to one age group as compared to others, SageMaker will flag the imbalance so that you can avoid a model that disfavors that age group. Check your trained model for bias You can also check your trained model for bias, such as predictions that produce a negative result more frequently for one group than they do for another. SageMaker Clarify is integrated with SageMaker Experiments so that after a model has been trained, you can identify attributes you would like to check for bias, such as age. SageMaker runs a set of algorithms to check the trained model and provides you with a visual report that identifies the different types of bias for each attribute, such as whether older groups receive more positive predictions compared to younger groups. Monitor your model for bias Although your initial data or model may not have been biased, changes in the world may introduce bias to a model that has already been trained. For example, a substantial change in home buyer demographics could cause a home loan application model to become biased if certain groups were not present or accurately represented in the original training data. SageMaker Clarify is integrated with SageMaker Model Monitor, enabling you to configure alerting systems like Amazon CloudWatch to notify you if your model exceeds certain bias metric thresholds.  Explain model behavior Understand your model Trained models may consider some model inputs more strongly than others when generating predictions. For example, a loan application model may weigh credit history more heavily than other factors. SageMaker Clarify is integrated with SageMaker Experiments to provide a graph detailing which features contributed most to your model’s overall prediction-making process after the model has been trained. These details may be useful for compliance requirements or can help determine if a particular model input has more influence than it should on overall model behavior. Explain individual model predictions Customers and internal stakeholders both want transparency into how models make their predictions. SageMaker Clarify integrates with SageMaker Experiments to show you the importance of each model input for a specific prediction. Results can be made available to customer-facing employees so that they have an understanding of the model’s behavior when making decisions based on model predictions. Monitor your model for changes in behavior Changes in real-world data can cause your model to give different weights to model inputs, changing its behavior over time. For example, a decline in home prices could cause a model to weigh income less heavily when making loan predictions. Amazon SageMaker Clarify is integrated with SageMaker Model Monitor to alert you if the importance of model inputs shift, causing model behavior to change.
• #42: Amazon SageMaker Clarify works across the entire ML workflow to implement bias detection and explainability. - It can look for bias in your initial dataset as part of SageMaker Data Wrangler - It can check for bias in your trained model as part of SageMaker Experiments, and also explain the behavior of your model overall - It extends SageMaker Model Monitor to check for changes in bias or explainability over time in your deployed model - It can provide explanations for individual inferences made by your deployed model
• #43: So to recap: - You can check your data for bias during data prep - You can check for bias in your trained model, and explain overall model behavior - You can provide explanations for individual predictions made by your deployed model - You can monitor and alert on any changes to model bias or behavior over time
• #44: Let’s take a quick product tour SageMaker Clarify is integrated with Amazon SageMaker Data Wrangler, making it easier to identify bias during data preparation. You specify attributes of interest, such as gender or age, and SageMaker Clarify runs a set of algorithms to detect any presence of bias in those attributes. After the algorithm runs, SageMaker Clarify provides a visual report with a description of the sources and measurements of possible bias so that you can identify steps to remediate the bias. For example, in a financial dataset that contains only a few examples of business loans to one age group as compared to others, SageMaker will flag the imbalance so that you can avoid a model that disfavors that age group.
• #45: You can also check your trained model for bias, such as predictions that produce a negative result more frequently for one group than they do for another. SageMaker Clarify is integrated with SageMaker Experiments so that after a model has been trained, you can identify attributes you would like to check for bias, such as age. SageMaker runs a set of algorithms to check the trained model and provides you with a visual report that identifies the different types of bias for each attribute, such as whether older groups receive more positive predictions compared to younger groups.
• #46: Although your initial data or model may not have been biased, changes in the world may introduce bias to a model that has already been trained. For example, a substantial change in home buyer demographics could cause a home loan application model to become biased if certain groups were not present or accurately represented in the original training data. SageMaker Clarify is integrated with SageMaker Model Monitor, enabling you to configure alerting systems like Amazon CloudWatch to notify you if your model exceeds certain bias metric thresholds. 
• #47: Trained models may consider some model inputs more strongly than others when generating predictions. For example, a loan application model may weigh credit history more heavily than other factors. SageMaker Clarify is integrated with SageMaker Experiments to provide a graph detailing which features contributed most to your model’s overall prediction-making process after the model has been trained. These details may be useful for compliance requirements or can help determine if a particular model input has more influence than it should on overall model behavior.
• #48: Changes in real-world data can cause your model to give different weights to model inputs, changing its behavior over time. For example, a decline in home prices could cause a model to weigh income less heavily when making loan predictions. Amazon SageMaker Clarify is integrated with SageMaker Model Monitor to alert you if the importance of model inputs shift, causing model behavior to change.
• #49: This concludes the overview of Amazon SageMaker Clarify, with a focus on the explainability functionality. Please refer the SageMaker Clarify webpage and the AWS blog post for additional information, including best practices for evaluating fairness and explainability in the ML lifecycle. Next, we will present the demo of SageMaker Clarify. Demo: https://youtu.be/cQo2ew0DQw0
• #50: Regulatory Compliance Regulations may require companies to be able to explain financial decisions and take steps around model risk management. Amazon SageMaker Clarify can help flag any potential bias present in the initial data or in the model after training and can also help explain which model features contributed the most to an ML model’s prediction. Internal Reporting & Compliance Data science teams are often required to justify or explain ML models to internal stakeholders, such as internal auditors or executives. Amazon SageMaker Clarify can provide data science teams with a graph of feature importance when requested and can help quantify potential bias in an ML model or the data used to train it in order to provide additional information needed to support internal requirements. Customer Service Customer-facing employees, such as financial advisors or loan officers, may review a prediction made by an ML model as part of the course of their work. Working with the data science team, these employees can get a visual report via API directly from Amazon SageMaker Clarify with details on which features were most important to a given prediction in order to review it before making decisions that may impact customers. ----- Ignore below: As we mentioned earlier, there are a few use cases where bias detection and explainability are key – this is by no means an exhaustive list: Compliance: Regulations often require companies to remain unbiased and to be able to explain financial decisions. Internal Reporting: Data science teams are often required to justify or explain ML models to internal stakeholders, such as internal auditors or executives who would like more transparency. Operational Excellence: ML is often applied in operational scenarios, such as predictive maintenance and application users may want insight into why a given machine needs to be repaired. Customer Service: Customer-facing employees such as healthcare workers, financial advisors, or loan officers often need to field questions around the result of a decision made by an ML model, such as a denied loan.
• #51: Here are some best practices for evaluating fairness and explainability in the ML lifecycle. Fairness and explainability should be taken into account during each stage of the ML lifecycle, for example, Problem Formation, Dataset Construction, Algorithm Selection, Model Training Process, Testing Process, Deployment, and Monitoring/Feedback. It is important to have the right tools to do this analysis. To encourage engaging with these considerations, here are a few example questions worth asking during each of these stages. Fairness as a Process: We recognize that the notions of bias and fairness are highly application dependent and that the choice of the attribute(s) for which bias is to be measured, as well as the choice of the bias metrics, may need to be guided by social, legal, and other non-technical considerations. Building consensus and achieving collaboration across key stakeholders (such as product, policy, legal, engineering, and AI/ML teams, as well as end users and communities) is a prerequisite for the successful adoption of fairness-aware ML approaches in practice. Wide spectrum of customers, ranging from highly sophisticated (e.g., data scientists / applied ML researchers who would like granular options) to less technical (e.g., visualize the bias measures)
• #52: Here are some best practices for evaluating fairness and explainability in the ML lifecycle. Fairness and explainability should be taken into account during each stage of the ML lifecycle, for example, Problem Formation, Dataset Construction, Algorithm Selection, Model Training Process, Testing Process, Deployment, and Monitoring/Feedback. It is important to have the right tools to do this analysis. To encourage engaging with these considerations, here are a few example questions worth asking during each of these stages. Fairness as a Process: We recognize that the notions of bias and fairness are highly application dependent and that the choice of the attribute(s) for which bias is to be measured, as well as the choice of the bias metrics, may need to be guided by social, legal, and other non-technical considerations. Building consensus and achieving collaboration across key stakeholders (such as product, policy, legal, engineering, and AI/ML teams, as well as end users and communities) is a prerequisite for the successful adoption of fairness-aware ML approaches in practice.
• #53: Here are some best practices for evaluating fairness and explainability in the ML lifecycle. Fairness and explainability should be taken into account during each stage of the ML lifecycle, for example, Problem Formation, Dataset Construction, Algorithm Selection, Model Training Process, Testing Process, Deployment, and Monitoring/Feedback. It is important to have the right tools to do this analysis. To encourage engaging with these considerations, here are a few example questions worth asking during each of these stages. Fairness as a Process: We recognize that the notions of bias and fairness are highly application dependent and that the choice of the attribute(s) for which bias is to be measured, as well as the choice of the bias metrics, may need to be guided by social, legal, and other non-technical considerations. Building consensus and achieving collaboration across key stakeholders (such as product, policy, legal, engineering, and AI/ML teams, as well as end users and communities) is a prerequisite for the successful adoption of fairness-aware ML approaches in practice.
• #54: This concludes the overview of Amazon SageMaker Clarify, with a focus on the explainability functionality. Please refer the SageMaker Clarify webpage and the AWS blog post for additional information, including best practices for evaluating fairness and explainability in the ML lifecycle. Thank you for listening to this talk.
• #55: Ben
• #56: Cf. accessibility Cf. urban design
• #59: Microsoft’s AETHER (AI & Ethics in Engineering & Research) Advisory Board Cross-industry initiatives such as Partnership on AI It would be desirable to create an internal advisory board within each tech company on AI and ethics (in case one doesn’t already exist), consisting of senior leaders from different business lines and different functional roles (AI & Engineering, Product, Legal, etc.) across the company (similar to Microsoft’s AI and Ethics in Engineering and Research (AETHER) advisory board). By forming working groups consisting of computer scientists and engineers, social scientists/ethicists, policy experts, lawyers, and product leaders, this board could be tasked with designing AI ethics guidelines, best practices, and tools for the company as a whole, so that all AI efforts across a company are aligned with the company’s responsible AI principles.
• #61: Examples of complex failures: Failure to deflect/terminate contentious topics Refusing to discuss when disapproval would be better Polite agreement with unrecognized bias Differentiate between bot input and bot output in training data Remove offensive text from bot output training But don’t remove from bot inputs  allow learning of good responses to bad inputs End-to-end system to support fairness
• #64: - (1) Differentially private model training, meeting practical requirements [model updates/evolution over time; at par on accuracy with vanilla model]; (2) membership inference attacks on ML models, could be used towards quantifying leakage from a specific model privacy-preserving mechanisms for different teams within a company to develop joint models on highly sensitive datasets.
• #65: Lessons from privacy & fairness challenges  Need “Fairness, Explainabilty, and Privacy and Fairness by Design” approach when building AI products Collaboration/consensus across key stakeholders (product, engineering, AI, PR, legal, social scientists, policy experts, …, end users / customers) NYT/WSJ/ProPublica test :)