rAs Kids📱 🌿✅ / 🍃🐛❌ → 94%
🎒 What you need
- one shared phone with a browser and internet for about 20 minutes
- two visibly different classes of real things: healthy vs blemished cassava or maize leaves, ripe vs unripe mangoes
- daylight and a plain background such as a cloth or bare ground
✋ Do this
- Open a free in-browser training tool (Google's Teachable Machine runs in a phone browser and needs no account).
- Create Class A and capture around 40 images, rotating and re-angling the object as you go. Variety inside a class is what makes it robust.
- Create Class B the same way. Deliberately include some awkward examples — half-shade, slightly blurred, held differently.
- Train. Then test on items the model has never seen and record the confidence numbers for ten of them in a table.
- Note something important: the two confidence numbers always add to 100%. The model cannot say 'I don't know' — it can only divide certainty between the boxes you gave it.
- Add a third class called 'nothing' — bare ground, hands, random objects — and retrain. Test again. Did the false alarms drop?
💡 Why it works
That third class is a genuine professional technique. A two-class model forced to choose will confidently mislabel anything unfamiliar, which is how a leaf-disease app ends up diagnosing a shoe. Giving it a background or 'none of these' class gives it somewhere honest to put the unknown. Real deployed systems also set a confidence threshold below which they refuse to answer at all.
🔥 Challenge
Build the same classifier twice: once with 20 images per class, once with 60. Test both on the identical ten items and report the difference in a table. Does more data always win?
📖 New words
classone of the categories a classifier can output
confidencethe model's score for each class, which is not the same as being right
thresholdthe confidence level below which you refuse to trust the answer
🧪
Model Trainer
Tap when you have finished this module.