Teach a no-code AI

The idea

A no-code AI tool hides the programming, but you still collect examples, label them, train the tool and test it, and its quality depends on the examples.

Do you need to code?

Could someone with no programming skill teach a computer to sort waste into paper, plastic and metal?

• No, only programmers can build AI
• Yes, by showing it labelled examples
• Yes, and it would need no examples at all

Be the no-code teacher

This very simple no-code tool has no code to write. You only label each item, one at a time, and the tool trains itself from your labels.

Then it is tested on new items it has never seen.

Same steps, no code

You did what a no-code tool asks. You gave labelled examples, this simple tool counted the clues and trained itself, then it was tested on new items.

It got the foil wrong. It was taught no thin metal, so it had no example of it, and the clues pointed to plastic. The tool hid the code, but not the need for good examples.

A no-code AI tool hides the code, but you still collect, label, train and test, and examples decide its quality.

In a real project you also collect the examples yourself, for instance by taking photos. Real tools that learn from photos or sounds find their own clues, not a list we write.

Notes

A no-code AI tool hides the code, but you still collect, label, train and test, and examples decide its quality.

Four words to remember: collect, label, train, test. And one more: examples matter most.

Check yourself

A class teaches a no-code tool to tell cats from dogs, using photos taken only in bright daylight. What may happen with a dark indoor photo?

Tenzin has labelled 30 leaf photos and pressed the Train button. What should he do next?

A student wants to teach a no-code tool to recognise classmates' faces, using photos from the class chat group. What is the best first step?

• It will cope fine, because no-code tools adjust to any light. No-code tools do not adjust by themselves. They learn only from the examples they are given.
• It will call every dark photo a cat, as a safe choice. There is no reason for one fixed answer. The guess just becomes less reliable.
• It may make more mistakes, since it saw no photos like that — correct. The examples set the limits. Adding photos in dim light and indoors, then testing again, would help.
• Test the tool with new leaf photos it has not seen — correct. Training is not the end. Testing on new photos tells him whether it really works.
• Start using it, because training makes it perfect. Training does not guarantee quality. The tool may have picked up something odd from his examples.
• Label the same 30 photos again, then train once more. The same photos teach the same things. Testing on new photos will show what is missing.
• Use the photos, as they are already shared in the class chat. Sharing a photo in a chat is not agreement to train a tool with it. People must be asked.
• Ask classmates and their parents first, and tell a teacher — correct. Faces are personal data. Consent comes first, with the purpose explained, and a teacher can help decide whether the project is needed at all.
• Take photos of strangers instead, so no classmate minds. Strangers have not agreed either. Photos of people need their permission, wherever they are.