TwiddleLab / Artificial Intelligence / How Machines See and Hear

Project: Teach It to See Critters

Lesson 10 of 10 · 8 min · Free, no account needed

What you will learn

• A vision project is a loop: choose the classes, label photos, train, test on new photos, study the misses, add better data.
• More of the same photos stops helping. Photos of the hard cases (dark, turned, look-alikes) fix what the model misses.
• Your model: pixels, 4 learned filters, pooling, 3 answers. Real camera-trap models are the same idea, much bigger.

Start the lesson

How the lesson goes

1. Four critter species, one camera trap. Build the model that tells three of them apart.

   Pick the 3 species your camera should know, then Check.

2. Label 12 portraits for your model.

   Drag each portrait to its species, then Check.

   A model copies its labels, mistakes included. Clean labels on clear photos are the base of every vision project.

3. How many portraits per species does it need?

   Move the slider. The model retrains each time and takes a test on 30 new portraits.

4. Train your model and watch it learn.

   Press Train. It runs right here in your browser.

5. Everything from this course, in one model.

   Tap to reveal the layers.

   Pixels: each portrait is 16 × 16 = 256 numbers.

   Learned filters: 4 filters of 3 × 3 that training shaped, each making a feature map.

   Pooling: the maps shrink to 4 × 4, so small shifts don't matter.

   Output: weights turn the pooled maps into one probability per species. Tested on portraits it never trained on.

6. Find its mistakes. Tap every test portrait it got wrong.

   Under each portrait is the model's answer. Then Check.

   Testing on new portraits shows real-world mistakes. A model that's 100% on its training photos can still miss whole kinds of new photos.

7. Why are these portraits hard? Sort them by cause.

   Drag each into a pile, then Check.

   The model only learned from bright, front-facing portraits. Low light, a turned head and an in-between ear shape are all things it never saw.

8. Classic goal: get your model to 90% on the test portraits.

   Add the photos it needs. Budget: 30 extra photos.

9. Ecologists use models like yours to sort millions of camera-trap photos. Why bother?

   Estimate the human work, then Check.

   3,000,000 × 5 s = 15,000,000 s; divided by 86,400 s a day that is about 174 days of nonstop work.

10. A model gets 27 of 30 test portraits right. What is its test accuracy, in percent?

    27 right out of 30 = 27 / 30 = 0.9 = 90%.

11. Look inside your model, then take your project card.

    Tap each layer: Pixels, Filter maps, Pooled, Answer.

Previous lesson: Faces, Cameras and Fairness · Part of the course: How Machines See and Hear

Lessons are interactive and need JavaScript. Turn it on to play them.