TwiddleLab / Artificial Intelligence / How Machines See and Hear
Faces, Cameras and Fairness
What you will learn
- A face network turns each photo into a face embedding: a list of numbers, a point. Photos of one person land close together.
- A match threshold decides how close counts as the same person: wider lets strangers in, tighter locks owners out.
- If a group was rare in the training photos, its people end up crowded together and get more false matches.
- Fair systems test errors per group, train on varied data and keep a human check where mistakes hurt.
How the lesson goes
Your phone unlocks with your face in a second. Would it work as well for everyone?
Make a guess, then Check.
A face network turns each photo into a point. Measure some distances.
Same color, same person. Tap two photos at a time.
Set the unlock threshold for group A: no strangers in, owners not locked out.
Each stored photo gets a circle. Slide its size.
What the phone actually does.
Tap to reveal.
The point for a face is a face embedding: a list of numbers learned so that one person's photos land close together.
The distance that still counts as a match is the match threshold.
Every threshold trades false unlocks against lockouts, like the alarm line in AI-3.
In NIST's 2019 test of 189 algorithms, many matched the wrong person 10 to 100 times more often for some groups than for others.
Now use your threshold on group B, whose faces were rare in the training photos.
Switch between the two groups.
Retrain with more group B photos. What happens to the crowd?
Slide the number of training photos.
Game: fair unlock. Make it work for both groups with the least extra data.
Each data step costs 1. Then set one threshold for everyone.
Math: a system makes 1 false match per 100,000 comparisons. A city runs 1,000,000 comparisons a day. How many false matches?
Expected false matches = rate × comparisons = 1/100,000 × 1,000,000 = 10.
Where is face recognition a small risk, and where can a mistake hurt someone?
Drag each use into a pile, then Check.
The same false-match rate means a retyped PIN in one use and a wrongly stopped person in another. High-stakes uses need per-group testing and a human check.
Write your rule card for fair face recognition.
Put 3 rules on the card, then Check.
One overall accuracy can look great while one group gets many more false matches. Testing per group finds it, better data fixes it, and a human check catches what's left.
Previous lesson: Sound Is a Picture Too · Next lesson: Project: Teach It to See Critters · Part of the course: How Machines See and Hear
Lessons are interactive and need JavaScript. Turn it on to play them.
