KnowSoda
TechnologyAI & Data
Machine Learning: fun facts and quiz
The essentials in 7 ideas, with fun facts, an illustrated story and a quiz to see how much you know.
The essentials
Training Data
Machine learning models learn from examples, called training data. The more varied and accurate the examples, the better the model usually gets.
Key fact
Garbage in, garbage out: bad data makes bad models.
Learning with Labels
In supervised learning, each example comes with the right answer, like photos labelled “cat” or “dog”. The model learns to predict the labels of new photos.
Key fact
Labelling data is often done by people, by hand.
Neural Networks
Neural networks are loosely inspired by the brain: layers of simple units pass numbers along and adjust their connections as they learn.
Key fact
“Deep learning” just means networks with many layers.
Predictions & Errors
A model makes predictions with some confidence, and it can be wrong. Engineers test it on examples it has never seen to check how well it really works.
Key fact
Testing on unseen data shows whether a model truly learned.
Overfitting
A model that memorises its training examples instead of learning general patterns is overfitting. It aces practice but fails on new data, like a student who memorised answers.
Key fact
More varied data helps prevent overfitting.
Learning by Reward
In reinforcement learning, a system learns by trial and error, earning rewards for good moves. That’s how AI came to master games like Go and chess.
Key fact
AlphaZero learned chess by playing millions of games against itself.
Machine Learning You Use Daily
Machine learning powers face unlock, spam filters, voice typing, translation and photo search. It usually works quietly in the background.
Key fact
Your phone can recognise faces in your photos without sending them online.
Fun facts
Some image models learned to spot cows partly by looking for green grass, so they failed on cows standing on a beach.
When a CAPTCHA asks you to pick out traffic lights, you may be helping to label data.
The first neural network machine, the Perceptron, was built in 1958 and learned to tell simple shapes apart.
Changing a few pixels, invisible to people, has fooled a model into seeing a panda as a gibbon.
One pneumonia-spotting model partly learned to recognise which hospital’s scanner took each X-ray, not the illness.
In 2015, an AI learned dozens of Atari games from nothing but the screen pixels and the score.
In the 1990s, early neural networks were already reading handwritten numbers on cheques and post.
Story: The Horse Who Could Count
The Horse Who Could Count
Berlin, 1904. A horse called Clever Hans amazes crowds by tapping out the answers to sums with his hoof. Experts can find no trick.
Then psychologist Oskar Pfungst notices that Hans only gets it right when the questioner knows the answer. The horse is reading tiny cues in their faces.
AI researchers still talk about the “Clever Hans effect”: a model that seems smart, but has learned a shortcut instead of the real task.
True or false?
Tap your answer to see if you got it, and why.
-
Question 1 of 10
The examples a model learns from are called training data.
Show the answer
True. Its learning material.
-
Question 2 of 10
Supervised learning uses labelled examples.
Show the answer
True. Each example comes with its answer.
-
Question 3 of 10
Deep learning uses neural networks with many layers.
Show the answer
True. That’s what “deep” means.
-
Question 4 of 10
Models are tested on data they haven’t seen before.
Show the answer
True. To check they can handle new cases.
-
Question 5 of 10
An overfitted model does badly on new data.
Show the answer
True. It memorised rather than learned.
-
Question 6 of 10
Reinforcement learning uses rewards.
Show the answer
True. Good outcomes are rewarded.
-
Question 7 of 10
Face unlock uses machine learning.
Show the answer
True. It learns what your face looks like.
-
Question 8 of 10
The quality of training data doesn’t matter.
Show the answer
False. Poor data leads to poor results.
-
Question 9 of 10
Supervised learning never needs any answers.
Show the answer
False. The labels are the answers it learns from.
-
Question 10 of 10
Neural networks are made of real brain cells.
Show the answer
False. They’re maths running on computers.
Keep going in the app
KnowSoda learns what you know and brings back what you’re about to forget. Machine Learning has 6 more ideas at the Advanced and Expert levels.
Sources
- Google — Machine Learning Crash Course
- Encyclopaedia Britannica — machine learning, neural network
- Silver et al. (2018), AlphaZero, Science 362
- LeCun et al. (1998), gradient-based learning applied to document recognition
- Vaswani et al. (2017), attention is all you need
- Epoch AI — Trends in machine learning compute
- ImageNet
