Responsible AI metrics

1. Responsible AI metrics

Let's review some responsible AI metrics.

2. AI project lifecycle

A responsible AI project demands careful data management across its entire lifecycle, from data acquisition, modeling, deployment, and subsequent monitoring.

3. Responsible AI project

We address responsibility at every stage, ensuring legal compliance, fairness, diversity, transparency, accountability, and security. We’ll focus our metrics discussion on model fairness. To start, we should define model fairness for our project to ensure it is fair, unbiased, and has equitable outcomes for everyone. We will focus on the conceptual application of these metrics.

4. Protected characteristics

To ensure fairness, lack of bias, and equitable outcomes, we must identify groups vulnerable to unfair treatment and discrimination. This can be based on protected characteristics like race, ethnicity, gender, or socioeconomic status or specific to the project.

5. Data acquisition

We use these groups to assess bias and fairness across all project stages. In planning and data acquisition, we’ll want to check for things like equal outcomes and conditional demographic disparity. There may likely be laws and regulations to consider; these will be addressed in the following chapter.

6. AI in facial recognition

A relevant example can be seen in AI facial recognition projects that have had high overall accuracy but have failed at recognizing specific ethnicities or genders. This was due to a lack of data availability, diversity, and representation.

7. Equal outcomes and demographic disparity

Equal outcomes metrics ensure that the benefits and outcomes are equal across protected and non-protected groups. Conditional demographic disparity observes the differences and inequalities between protected and non-protected groups. We can also use descriptive statistics and distributions to assess diversity and representation in the data and apply appropriate corrective measures such as weighting and balancing. This may be hard to know at the beginning, and we may need to revisit after the first testing iteration. Ensuring we have enough diverse data is a good starting point, and always keep track of any tests and transformations for proper transparency and accountability.

8. Modeling

In the modeling stage of our AI project, we can check for equal performance, which is not always the same as equal outcomes. One example is AI for medical diagnoses, as some may be more common in certain protected groups. It’s key to evaluate metrics such as false negatives, false positives, and accuracy to achieve the appropriate balance. During this stage, we also need to keep in mind the explainability of our model to remain transparent and accountable while keeping privacy and security in mind. We carefully document data and algorithms and calculate explainability scores to check if the model's predictions are understandable to humans. Two explainability scores are the Local Interpretable Model-agnostic Explanation (LIME) and the Shapley Additive Explanation (SHAP). The LIME score breaks down the model decision into smaller, more interpretable parts, and the SHAP score tells us how different features contribute to the results.

9. Deployment and monitoring

After an ethical committee has audited our model, we can move to the deployment and monitoring phase. Here we assess model drift, or changes to the model’s performance over time. These changes can come from changes in the real world, such as changes to a company’s customer demographics meaning the results are no longer accurate. To address the model drift, we monitor distributions and technical performance metrics and adjust our model accordingly, keeping track of versioning.

10. Applying metrics

Understanding the protected characteristics helps us use the appropriate metrics to ensure our model is fair and responsible across the different dimensions. We have not gone through an exhaustive list of metrics here; many more exist and are often specific to use cases. Always consult the appropriate legal and domain experts for confirmation and advice. Before we wrap up, we must also remember to conduct privacy and security checks and protect the data and project from breaches and unauthorized access.

11. Let's practice!

Now, let's practice!

Create Your Free Account

By continuing, you accept our Terms of Use, our Privacy Policy and that your data is stored in the USA.