📖 Chapter 3 of 9 · AI For New Learner Advance 1 (Multi Indian Language) · Class 7
Chapter 3: How AI Really Learns
In this chapter
- The REAL 6 steps every AI engineer follows.
- Arjun from Lucknow trains an AI to read handwriting.
- Why ‘training data’ decides whether AI is smart or silly.
- What does ‘90% accuracy’ really mean?
- When AI keeps learning — and when it stops.
The Machine Learning Pipeline
Real AI engineers — at Bengaluru companies, ISRO, IIT labs — follow the same 6 steps to BUILD an AI. We will study each step. By the end of this chapter, you will understand the process that powers ChatGPT, UPI fraud check, and the Mumbai Metro AI — all of them.
Did you know?
Training a big AI like ChatGPT can cost over Rs 800 CRORE (10 million dollars) in electricity and computers. India is building smaller, smarter AI — like Sarvam AI and Krutrim — that work well at much lower cost.
Story 1: Arjun’s First AI
From: Lucknow, UP · Arjun, age 12
Arjun lives in Lucknow, the City of Nawabs. His older brother is studying at IIT Kanpur. During the holidays, Bhai gave Arjun a real challenge: “Build an AI that can READ the marks I write on your test papers — like 7, 9, 23.”
First, Arjun COLLECTED — 1000 photos of handwritten numbers from his friends. Then he CLEANED — removed blurry photos. Then he LABELLED each: ‘this is a 7’, ‘this is a 9’. He SPLIT them: 800 for training, 200 for testing.
Bhai helped him TRAIN a small AI on the 800 photos. The AI looked for patterns — what makes a ‘7’ a 7 — the slanted line, the dash, the curve. Then they TESTED on the 200 photos AI had never seen. Result: 92 out of 100 correct!
Arjun was thrilled. “This is exactly what big companies do — just bigger. They use 800 MILLION photos, but the SAME 6 steps.” Arjun realised: AI is not magic. It is a PROCESS. And he could do it too.
Story 2: When the AI Failed
From: Lucknow, UP · Arjun’s lesson
Arjun’s AI worked great in tests. But then he tried it on his Dadi’s handwritten shopping list — and AI failed badly! It thought ‘6’ was ‘0’. It read ‘4’ as ‘9’. Why?
Bhai smiled. “Look at the training data. Whose handwriting did you collect? Friends, age 11-12. Dadi is 75. Her handwriting style is different — bigger, with shaky lines. AI never SAW that.”
Arjun added 200 photos of his Dadi’s, Mummy’s, Papa’s writing. He re-TRAINED the AI. He re-TESTED. Now accuracy went up to 96%. Then he DEPLOYED the AI as a small Android app he gave to his grandfather.
Arjun wrote in his diary: “AI is only as smart as its DATA. Garbage in = garbage out. Variety in = wisdom out. Real AI engineers spend 80% of time on data, only 20% on the AI itself.” That night his Bhai gave him an old AI textbook from IIT. “Welcome to the field, brother.”
Solved illustrations
Deep-dive into machine learning
1. What are the 6 steps of the Machine Learning Pipeline?
Show answer
Answer: (1) COLLECT data, (2) CLEAN it, (3) SPLIT into train and test sets, (4) TRAIN the AI on training data, (5) TEST on unseen data, (6) DEPLOY to real users. Engineers say: ‘Garbage in, garbage out’ — so steps 1-2 are most important.
2. Why do we SPLIT data into training and testing?
Show answer
Answer: If we test on the SAME data the AI studied, of course it will pass! Like writing your own question paper. We hide 20% of data, train on 80%, then test on the hidden 20%. This shows if the AI really LEARNED, or just memorised.
3. What does ‘92% accuracy’ mean?
Show answer
Answer: Out of 100 test photos, AI got 92 correct, 8 wrong. Sounds great — but if 8 of those were medical X-rays, those 8 mistakes are serious. For life-critical AI, we need 99.99% or higher. Accuracy is a number; meaning matters more.
4. Why did Arjun’s AI fail on Dadi’s handwriting?
Show answer
Answer: Because Dadi was not REPRESENTED in the training data. Only young friends’ handwriting was used. This is called BIAS. Real AI systems must include data from ALL types of people — old, young, urban, rural, all states.
5. Why do AI engineers spend 80% time on data, not AI?
Show answer
Answer: Because the AI part is well-known (formulas exist for many years). The DATA is unique to your problem — and bad data ruins even the best AI. Just like good cooking needs good ingredients more than fancy equipment.
The 6-step ML pipeline
Every AI in the world — from ChatGPT to UPI fraud check — is built with these same six steps, in order.
Try this at home
Imagine you have to build an AI to recognise Indian flags from photos. List which 6 steps you would do. Discuss with friends — your data sources will be interesting!
Questions — Chapter 3
Think like an AI engineer
A. Order them — write the correct order (1 to 6)
____ TEST ____ COLLECT ____ TRAIN ____ CLEAN ____ DEPLOY ____ SPLIT
B. Choose the BEST answer
- Arjun’s AI used ____ training photos. (a) 100 (b) 800 (c) 10000
- AI gives wrong answers MAINLY because — (a) computer is slow (b) training data is not enough or biased (c) Wi-Fi is weak
- To TEST an AI properly, we must use — (a) the same data it trained on (b) new data it has never seen (c) no data at all
C. Short Answer
- Why did Arjun’s AI fail on Dadi’s handwriting?
- What does ‘garbage in, garbage out’ mean for AI?
- List ONE real-world problem you would solve with AI, and what data you would need.
Brain Buster
Visit teachablemachine.withgoogle.com with a parent. Try the 6 steps yourself — for FREE!
Check your understanding — Chapter 3
Tap an answer to check it instantly. These are for practice — nothing is saved or sent.
Q1. What are the 6 steps of the Machine Learning pipeline, in order?
- Train, Test, Collect, Clean, Split, Deploy
- Collect, Clean, Split, Train, Test, Deploy
- Deploy, Train, Collect, Test, Clean, Split
Show answer
B. Collect, Clean, Split, Train, Test, Deploy
Q2. Why do we SPLIT data into training and testing sets?
- To save space
- So we can test on data the AI has never seen
- Because computers are slow
Show answer
B. So we can test on data the AI has never seen
Q3. Why did Arjun's AI fail on Dadi's handwriting?
- Dadi wrote too fast
- Dadi's writing style wasn't in the training data — this is bias
- The app had a virus
Show answer
B. Dadi's writing style wasn't in the training data — this is bias
Q4. What does 'garbage in, garbage out' mean for AI?
- AI produces literal garbage
- Bad training data leads to a bad AI
- AI needs to be cleaned daily
Show answer
B. Bad training data leads to a bad AI