Summary and Next Steps

Course summary

Congratulations! You just completed a long journey in the fascinating field of AI fairness. We started with almost zero knowledge, and now we are equipped with all the tools and expertise to implement fair practices in data science products.

In the beginning, we discussed different aspects of fairness, why it is essential, and how tricky it can be to find issues in this matter. We introduced a hypothetical Space University to discuss various aspects safely but saw a real-life problem very clearly.

Then, we moved to various metrics for fairness. We saw no single best one, and selecting a proper one might be more complicated than in regular machine learning practice. Moreover, we found that trying to increase the wrong metric can make fairness issues even bigger. During the course, we found many tricky situations, such as the Simpsons paradox, so we are careful when developing new models.

Then, we moved to various mitigation techniques. We started with recognizing possible sources of bias and knowing what kind of approach can be utilized in which scenario. Then, we learned how to use Fairlearn and AI Fairness 360 in practice.

Last but not least, we explored how biases can manifest in natural language processing. Starting with stereotypical associations, we learned how to measure and eliminate them. Finally, we learned how to perform systematic testing that helps us ensure that we don’t discriminate against anyone (or at least minimize it).