Action: Teaching AI to See

How do you teach AI to find cats in photos?

1. You give the AI lots of photos. Since we are teaching it to find cats, we need lots of pictures of cats. It is just as important to include pictures that do not contain a cat, like other animals.
2. You tell the AI which photos have cats, and where the cats are.
3. The AI identifies patterns. It compares the features of the animals we labelled and notices that most cats have a tail, pointy ears, two eyes, and fur.
4. When you give the AI a new photo, it uses that pattern to find cats. Tail + pointy ears + two eyes + fur = cat!

This is called object detection.

Activity 6: Is there a cat in this photo? (3 minutes)

Our cat-finding AI is shown a new photo. In the photo there is an animal that is furry, has two eyes, pointy ears, and a bushy tail. It is a husky dog.

Activity 7: The right tool for the job (4 minutes)

Which of these photos contains a cat?

1. A cheetah drinking at a watering hole
2. A stuffed toy cat on a bed
3. A pumpkin carved with a cat face
4. A house cat asleep on a sofa

It depends on your definition of "cat"! When you train an AI, how you label the data depends on the purpose of the AI.

• If the AI is being trained to monitor endangered species visiting a watering hole in Africa, you would label the cheetah as a cat, but the toy and the pumpkin as NOT cats.
• If the AI is being trained to use city traffic cameras to help find lost pets, only the house cat counts. None of the others are pets to rescue.
• If the AI is being trained to search a photo library for pictures of cats, you would probably want all four photos returned.

If you use an AI without knowing its purpose or how it was trained, you may get results you did not expect, because your definition is different from the definition of the person who trained it. Does that mean the AI is wrong? No. It means you picked an AI tool that is not suited to your task.

Every AI is trained for a specific task

Baseballs and basketballs are both types of balls, but it would not work very well to play baseball with a basketball. In the same way, each AI is trained to do a specific job, so when you want to do something with AI you need to find the AI best suited for that job.

Here are two real examples of object detection at work in Canada:

• Tracking salmon. You can't know if a fishery is sustainable if you don't know how many fish are coming back. Salmon Vision uses underwater cameras in Indigenous-run fish counting weirs on British Columbia's Central Coast, and an AI model identifies 12 different species of fish with 80 to 90 percent accuracy, so salmon populations can be monitored in real time. Read the story.
• Counting beluga whales. Fisheries and Oceans Canada worked with Esri Canada to train models that detect beluga whales in Arctic waters from aerial, drone, and satellite images with 85 to 88 percent accuracy. The model can analyze thousands of images in hours, a job that used to take months.