Definition
How did you learn what a dog looks like? Did anyone give you a rulebook?
Humans learn from experience: what they see, hear, try and are told, connected to everything they already know. Machines learn from data: many examples, from which a computer finds patterns it can use on new cases.
In simple words
A small child sees a few dogs, hears someone say "dog", and soon spots dogs everywhere, even cartoon ones. A machine usually needs thousands of labelled dog photos to do the same, and it may still be confused by a cartoon dog. People learn from a few examples and lots of background knowledge. Machines learn from many examples and nothing else.
Why it matters
This difference explains why data matters so much in AI, why AI can make strange mistakes, and why collecting good examples is one of the most important jobs in building any AI system.
How it works
| A person learning | A machine learning | |
|---|---|---|
| Learns from | Senses, experience, teachers, books | Examples stored as data |
| Examples needed | A few | Thousands or millions |
| Background knowledge | Huge: links new ideas to old ones | Almost none beyond its data |
| Handles something odd | Usually copes | May fail in surprising ways |
Both improve with feedback. A child is told "that's a cat, not a dog". A machine is shown its mistakes and adjusts.
Try it yourself
Learning the word “mango”
See what a person needs, and then what a machine needs.
Teach a partner a new word
Do this with a friend or family member, then compare it with how a machine would learn.
- Invent a word, such as "blorp", meaning "a round fruit with a seed in the middle".
- Show your partner three examples (a mango, a peach, a plum) and two non-examples (a banana, a grape bunch).
- Ask them to point to a blorp among new fruits. Count how many examples they needed.
- Write a list of everything a machine would need to learn the same word.
Think about it
- Why did your partner need so few examples?
- Which fruit might confuse a machine, and why?
Human, machine or both?
Who does each of these when learning?
Can check a million photos in minutes
Needs thousands of labelled examples
Uses common sense from daily life
Gets better after being corrected
Learns a new game from one explanation
Common mistakes
- Thinking machines learn the way children do. They learn from data only, with no life experience.
- Thinking more data always fixes everything. Wrong or one-sided data teaches wrong lessons.
- Thinking a machine that learned something understands it. It found a pattern; understanding is not proven.
Check yourself
Summary
- People learn from experience, senses and teachers; machines learn from data.
- People need few examples because of background knowledge; machines need many.
- Both improve with feedback, which is why correcting AI matters.
Finished reading?
Sources: Elements of AI (University of Helsinki) · AI4K12: Five Big Ideas in AI