How a machine learns from examples

Every May, Kabir’s nani in Lucknow walks through the fruit market and picks the perfect dasheri mango in about two seconds. She presses it gently, sniffs near the stem, glances at the colour. Kabir tries the same and comes home with three rock-hard mangoes and one that is mush.

Nani never read a rulebook on mangoes. She has squeezed thousands of them over fifty summers, and every wrong pick taught her a little. That slow, quiet correction is exactly how a machine learns. In this lesson we will teach a computer to pick ripe mangoes, one small mistake at a time.

Picture Priya at cricket nets in Chennai, learning to bowl. Her first ball goes wide down the leg side. The coach says, “A little to the left.” Next ball, closer. “Bit fuller.” Closer still. She is not memorising a rule. She is adjusting a tiny bit after every ball, based on how far off she was.

A computer learns in the same three-beat rhythm. It makes a guess. It checks the guess against the right answer and measures how wrong it was. Then it nudges its internal settings so the next guess would be a little better. Repeat that enough times, and the guesses become good.

Let us build nani’s skill inside a computer. First we need examples where the answer is already known. Kabir and nani go through 500 mangoes. For each, they note three things on a scale of 0 to 10: how soft it feels, how strong it smells and how yellow it looks. Then they cut it open and write “ripe” or “not ripe”.

Our model starts with three knobs, one for each feature. The knobs decide how much each feature counts. At the start they are set randomly, so the model is as clueless as Kabir. Step through the loop below and watch the knobs move.

The numbers here are made up for the story, but the shape is real. Accuracy climbs fast at first and then creeps up slowly. And the test score on new mangoes is usually a bit lower than on the practice ones.

Rohan’s cousin topped every class test by memorising last year’s answers word for word. Then the board exam changed the questions slightly, and he struggled. He had learned the paper, not the subject.

Models can make the same mistake. Train too long on too few examples, and a model can memorise each mango, even the bruise on mango #3. It never learns “soft and fragrant means ripe”. It scores nearly 100% on practice and stumbles on anything new.

Our mango model had three knobs. The large language models behind chat apps have billions of them, and the biggest have far more. But the loop is the same one you stepped through.

For a language model, the example is a piece of text with the next word hidden. The model guesses the hidden word, checks the real one, and nudges its knobs. “Chai is best with ___”: guess “sugar”, truth “biscuits”, small nudge. Repeat across a huge library of text, and the model slowly picks up grammar, facts and style.

Here is a neat detail: nobody has to label that text by hand. The next word is already sitting there in the sentence, so the text labels itself. That is a big reason language models could learn from so much data. You will see this next-word game in detail in module 2.