Skip to main content

Skills application demo

Instruction and application
In Progress

Skills application demo​

Facing fairness in facial recognition demo​

In this skills application demo, we'll apply an end-to-end fairness auditing pipeline on a facial-emotion CNN—reinforcing skills in demographic proxy derivation, slice-level evaluation, visualisation, and bias mitigation for robust and equitable models.

Challenge instructions and resources​

Select the link below to download and access the Jupyter Notebook so you can follow along with this activity. It includes the challenge instructions and the skills application demo files needed.

EnsembleMethods_demo.zip

Activity instructions​

Watch the demo​

Watch the coach demonstrate how to apply techniques to evaluate and refine the performance of the model.

Key takeaways​

  • Fairness Auditing = Measurable Impact: You saw how targeted fixes—post-hoc temperature scaling and tone-aware re-weighting—can close FNR gaps across skin-tone groups, no heavy retraining required.
  • Clear Equity Insights, Better Governance: You learned to connect slice-level fairness metrics from the facial-emotion dataset to real-world governance decisions, powering ethical deployment and stakeholder buy-in.
  • Workflow Confidence from Day One: You’ll own the end-to-end fairness pipeline—proxy derivation, disaggregated evaluation, mitigation, and trade-off analysis—and apply it immediately to govern your ML models.

Skills application​

Facing fairness in facial recognition practical​

In this skills application, you'll apply the techniques show in the demo to compute tone weights, build per-sample weights, fine-tune the last 3 epochs with sample weight and then recompute weighted-F1 and max-FNR gap

≤ 0.05 and F1 ≥ baseline

Learning this skill brings value to the workplace because mastering slice-level fairness evaluation equips you to spot and fix hidden biases, delivers more reliable predictions across all demographics, and drives data-informed, ethical decision-making in your ML projects.

Challenge instructions and resources​

Select the link below to download and access the Jupyter Notebook for this activity. It includes the challenge instructions and the files needed to complete the challenge.

fairnessFacialRecognition.zip

Activity instructions​

Work on the challenges​

Follow the instructions in the Jupyter Notebook to work on the challenges. Use the provided resources to:

  • Compute tone weights.
  • Build per-sample weights.
  • Fine-tune the last 3 epochs with sample_weight.
  • Recompute weighted-F1 and max-FNR gap—aim for gap ≤ 0.05 and F1 ≥ baseline.

Collaborate in the breakout room​

Discuss and share insights with fellow apprentices as you work on the activity challenges. While the activity is designed for individual completion, feel free to ask questions, collaborate, and compare approaches with your group.

Share your findings​

Submit your completed Jupyter Notebook and summarise your findings.

Regroup​

Return to the main session after 25 minutes to discuss key takeaways and insights from the exercise.