Kōkiri Learn

Week 4 of 5

Train an AI, test it and find its blind spots

  1. 1
  2. 2
  3. 3
  4. 4
  5. 5

Session 1 · DESIGN: Improve

Teach a machine

  1. 10 minHow AI learns: it looks at many labelled examples and finds patterns. Show how Spyfish counts from volunteers train DOC's fish AI.
  2. 20 minTrain Teachable Machine to tell two objects apart (harakeke leaf and kawakawa leaf, or can and bottle). Use at least 30 photos of each.
  3. 15 minBreak it on purpose: train a second model with photos from only one background or one angle. Test both. Which is fooled, and why?
  4. 5 minRecord: what did the training data leave out, and who could that be unfair to if this were a real app?

Session 2 · DESIGN: Improve

Is AI right for this job?

  1. 15 minYour teacher asks an AI chatbot three questions about your topic on the projector. Fact-check each answer against a trusted source. Was it right, partly right or made up?
  2. 20 minAI suitability card: for your context, score an AI tool on accuracy, fairness and context. Should it count birds, fish or people? What must a person still do?
  3. 10 minMisinformation check: look at a fake-looking post about your topic. Use the three questions: who made it, what is the evidence, what do other sources say?
  4. 5 minImprove your data story plan using what you learned.

Getting started

Train Teachable Machine with only two classes and a plain background first.

Stretch

Play Code.org AI for Oceans and explain how the AI's training changes when you choose which examples to show it.

Checkpoint

A tested model with a written note on its blind spot, and a draft AI suitability card.

Activities for this week

Useful this week