Machines do not see the world the way you do. You look at a mango. You notice its colour, its smell, the weight of it in your hand. A machine cannot do any of that directly. It can only work with data. Data is a recorded piece of information that describes something in the world.
Data comes in four main forms. Text is words and sentences — a message, a label, a book page. Numbers are counts and measurements — a temperature reading, a price, a distance. Images are pictures captured as grids of colour values. Sound is a recording of audio — a spoken word, a bell, a bird call.
Nila once drew four boxes in her notebook. In each box she wrote one word: text, numbers, images, sound. "These are the four doors," she said. "Everything a machine reads has to come through one of them."
Dhruv asked about video. Nila thought about it. "A video is images played one after another, plus a sound track," she said. "So it comes through two doors at once."
A machine does not experience the world. It receives data through one or more of those four forms. Only then can it begin to work.
Here is the idea that runs through this whole chapter. When a machine looks at a mango, it does not see a mango. It sees a row of recorded values: colour = yellow-orange, weight-group = medium, surface = smooth. Those recorded values are the data — the machine works from them, not from the fruit itself. This move is called abstraction: you choose which values to record and set everything else aside. You decide what matters for the task and leave out what does not. A doctor checking a patient records temperature and heart rate, not the colour of the patient's socks. A machine learning from mango images records colour and shape values, not the smell. Abstraction is how data stays useful without becoming overwhelming.
Nila has a notebook page. Down the left side, she lists eight household objects. First come a steel cup, a wooden spoon, a rubber eraser, and a glass bottle. Then come a plastic comb, a cotton cloth, a ceramic plate, and a paper bag. Across the top, she writes three column headings: material, weight-group, and use.
"Each column," she says, "tells me one thing about every object. The material column tells me what the object is made of. The weight-group column tells me whether it is light, medium, or heavy. The use column tells me what the object is for."
Each of those columns is a feature. A feature is one recorded property that describes an example. When you build a data table, each column heading names a feature. Each row holds the feature values for one example.
Dhruv checks: "So the material column is one feature? The weight-group column is another? And use is a third?" "Exactly," Nila says, tapping the headings with her pencil. "Three columns. Three features."
This is the structure of almost every data table a machine works with. Rows are examples. Columns are features. The table holds many examples, all described by the same shared set of features. That shared structure is what lets a machine compare one row to another.
Nila adds a fourth column to her table and writes type at the top. For the steel cup she writes metal object. For the wooden spoon, natural-material object. For the rubber eraser, synthetic object. For the glass bottle, natural-material object too. Then she stops — she leaves the type cell blank for the plastic comb, the cotton cloth, the ceramic plate, and the paper bag. "The rows with a type filled in are labelled," she explains. "The label is the answer. It tells me — or a machine — what this example is. The rows with the type cell blank are unlabelled. The answer is missing." A table can hold both kinds of row at once: some examples already show their answer, and some do not yet.
Dhruv picks up the comb row. "Can I predict the type from the other features?" he asks. "Try," Nila says. Dhruv reads across the row: material = plastic, weight-group = light, use = grooming. "I'd say synthetic object," he says. "Plastic is a synthetic material."
Nila checks his guess against the pattern in the other rows. It matches. "You just did what a simple classifier does," she says. "You used the features to predict the label."
This connects to machine learning. In Chapter 1 you met supervised learning. It is a machine. It learns from labelled examples. Nila's labelled rows are exactly what supervised learning needs. Each row shows feature values plus a correct answer. From many such rows, a model gets built. That model predicts the label for new, unlabelled rows.
Unsupervised learning works the opposite way. It receives only the feature columns. There is no answer column. It looks for groupings or patterns on its own. The unlabelled rows in Nila's table are exactly that input. Unsupervised learning works from rows like these.
So far, Nila's table has used text and numbers as feature values. But two of the four data forms — images and sound — do not arrive as neat table cells. They arrive as raw recordings. A photograph is not a word. A bell chime is not a number. So how do they become data a machine can work with? The same move as before: abstraction. You choose which properties to record, and you record them as values.
Nila takes two cards. One is a picture of a pencil case. The other is a short audio clip of a single bell chime. She treats each card as a row in a table.
For the pencil-case image, she lists what she can name just from looking at it: colour (blue), shape (rectangular), size-group (medium), surface (fabric). "Each of those is a visual feature," she says. "I'm not recording the whole photograph. I'm recording the properties that matter for the task."
For the bell chime, she lists pitch (high), duration (short), and pattern (single strike) — audio features. "So the image and the sound are both just rows in a table now?" Dhruv asks. "Once you name the features, yes," Nila says.
"The machine doesn't see a photograph. It sees colour = blue, shape = rectangular, size-group = medium, surface = fabric. It doesn't hear a bell. It reads pitch = high, duration = short, pattern = single strike." This naming step is data analysis: looking at raw input and asking which properties matter for the task.
Dhruv covers the sound card's label. He asks whether he could still say something useful about it without knowing the answer. Nila looks at the feature values. "You could group it with other short, high-pitched, single-strike sounds," she says. "You wouldn't know its name. But you'd know it belongs with sounds that share those features."
Features are useful even without a label. The label tells you the answer. But the features let you compare and group — with or without it.
Data arrives in one of four forms — text, numbers, images, sound. Abstraction picks the properties that matter for a task and names them as features. A table's rows split into labelled, where the answer column is filled in, and unlabelled, where it is left blank. And a classifier — simple or machine — uses the features of an unlabelled row to predict its missing label, the way Dhruv predicted the comb's type from its material, weight, and use. The next chapter picks up from here: pattern recognition looks at features across many rows and finds what comes again, using the features named in this chapter as its raw material.