AI Basics for Kids· Lesson 3 of 8
இயந்திரங்கள் எப்படி கற்றுக்கொள்கின்றன?
How Do Machines Learn?
~11 min
You learned to recognise mangoes by seeing hundreds of them — different sizes, colours, angles. AI learns the same way, from millions of examples. But what happens when those examples are one-sided or incomplete? This lesson is where judgment begins.
By the end of this lesson you will be able to— இந்த பாடத்தின் இறுதியில்
- Explain machine learning as learning from examples (not from rules given by a programmer)
- Describe what training data is and why its quality matters as much as its quantity
- Explain what happens when an AI is corrected after a wrong prediction
- Understand that incomplete or one-sided training data produces limited or unfair AI
Let's Learn
What you will learn today
Understand how machine learning works — data, labels, training, and error correction — through concrete analogies.
How Did You Learn to Recognise a Dog?
You were never given a rule like 'dogs have four legs, fur, a tail, and bark'. You saw hundreds of dogs — your parents said 'dog', you noticed patterns — and eventually your brain built its own mental model. You can now recognise a dog you have never seen before, in any lighting, from any angle.
This is exactly how machine learning works — except with far more examples and mathematics.
Three Ways Machines Learn
There are three main types of machine learning:
1. Supervised Learning — the machine learns from labelled examples (data with correct answers provided). Most common. Like a teacher showing flashcards and marking right/wrong.
2. Unsupervised Learning — the machine finds patterns in unlabelled data on its own. Like grouping customers by shopping behaviour without being told what the groups are.
3. Reinforcement Learning — the machine tries actions and receives rewards or penalties, learning by trial and error. Like training a dog with treats. Used for games and robots.
Today we focus on supervised learning — the most widely used.
- Supervised: learn from labelled examples
- Unsupervised: find patterns in unlabelled data
- Reinforcement: learn from rewards and penalties
📐 Supervised Learning Step by Step
Imagine teaching an AI to tell cats from dogs in photos:
Step 1 — Collect data: Gather 100,000 photos of cats and dogs.
Step 2 — Label the data: A human marks each photo 'cat' or 'dog'.
Step 3 — Train: The AI looks at each photo, makes a guess, finds out if it was wrong, and adjusts its internal settings slightly.
Step 4 — Repeat millions of times: After millions of adjustments, the AI gets very good at the patterns that distinguish cats from dogs.
Step 5 — Test: Show the AI photos it has never seen. If it gets most right, the training worked.
The 'internal settings' the AI adjusts are called weights — numbers that represent how much importance to give each feature it notices.
- 11. Collect data
- 22. Label the data
- 33. Make a guess, check, adjust
- 44. Repeat millions of times
- 55. Test on new, unseen data
What Are Neural Networks?
The most successful machine learning systems use neural networks — a structure loosely inspired by the human brain.
A neural network has layers of nodes (like neurons). Each node takes in numbers, multiplies them by its weights, and passes the result on. Through the training process, the weights are adjusted until the network makes accurate predictions.
Deep learning means using neural networks with many layers — sometimes hundreds. Each layer learns to recognise increasingly complex patterns: early layers notice edges and colours; later layers notice shapes and objects; final layers make the actual decision.
The Quality of Data Matters More Than Anything
A machine learning system is only as good as the data it trains on. Garbage data → garbage AI.
If you train a spam filter on only English emails, it will struggle with Tamil spam. If you train a face recogniser mostly on light-skinned faces, it will be worse at recognising darker-skinned faces. The patterns the AI learns are a mirror of whatever data it was given.
Misconception: 'More Data Always Means Better AI'
More data helps — but only if the data is high quality and relevant. Problems:
• Biased data: if all your 'doctor' training examples are men, the AI learns doctors are male
• Irrelevant data: adding random images won't help a cat/dog classifier
• Incorrectly labelled data: a mislabelled cat as 'dog' teaches the AI the wrong pattern
A smaller, carefully curated dataset often outperforms a massive messy one.
Human Learning vs Machine Learning
Human child learning to recognise cats:
• Needs perhaps 10–20 examples
• Learns from a single experience (touch a cat, remember it is soft)
• Understands context (a cartoon cat is still a cat)
• Generalises immediately to new animals
Machine learning to recognise cats:
• Needs tens or hundreds of thousands of examples
• Cannot learn from a single example
• Struggles with context and cartoon versions initially
• Does not know what a cat IS — only what cat images look like statistically
Humans are vastly more sample-efficient. Machines compensate with scale and speed.
Challenge Round
Design a Training Dataset
You want to train an AI to recognise whether a student is confused during a lesson (from their facial expression). Think through:
1. What data would you collect?
2. How would you label it?
3. What could go wrong with the labels?
4. How would you test whether the AI actually works?
How Machines Learn
Machine learning is about finding patterns in data through repeated examples. Supervised learning uses labelled data — the most common approach. Neural networks adjust weights through millions of corrections until they make accurate predictions. The quality and diversity of training data determines everything.
You now understand the core of machine learning: labelled data, pattern finding, and weight adjustment. This is the engine behind almost every AI product you use.
↪ Next lesson: how do computers 'see' images? This is one of the most powerful and surprising abilities AI has developed.
Key Points
- ✓Machine learning means a computer finds patterns in data instead of being given rules
- ✓Supervised learning uses labelled examples (like photos tagged 'cat' or 'dog')
- ✓A neural network adjusts its weights during training to reduce errors
- ✓The more diverse and accurate the training data, the better the AI performs
- ✓Reinforcement learning trains AI through rewards and penalties
Glossary
சொல் அகராதி
Machine learning
இயந்திர கற்றல்
Training data
பயிற்சி தரவு
Pattern
வடிவம்
Prediction
கணிப்பு
Feedback
பின்னூட்டம்
Label
தலைப்பு (data label)
Practice Activities
Quizவினாடி வினா
Answer each question to check your understanding.
In supervised learning, what does 'labelled data' mean?
Fill in the Blanksஇடைவெளி நிரப்புக
Type the missing word and press Check or Enter.
Type the missing word and click Check or press Enter.