The lesson
12 orders, 1 pot, no panic
Friday night, 7:30, and the room is full. Twelve orders are up at once. Nobody panics. My hands move on their own. I check the pot, I check the tickets, I check the bread, and the bowls keep coming out right. People think that is one skill, the skill of a cook. It is not. It is twenty small skills stacked on top of each other, each one tiny and dull on its own, and together they run my whole kitchen.
A stack of tiny questions
The machine that runs the modern world is built the same way. A neural network is a stack of small decision units. Each unit is dull. It takes in a few numbers, weighs them, and passes a judgement along. That is all. But stack enough of them in layers, and the whole stack learns things no single unit could name. A unit does not know what a carrot is. A layer does not know what a vegetable is. The stack knows a carrot when it sees one. The first layer asks tiny questions: is it round? is it long? is it orange? The next layer builds from those answers: is it a root? The next: is it a carrot? Deep learning just means the stack is tall. The weights are what change during training, and nobody programs the word carrot. The stack learns the weights from thousands of labelled photos. That is the deep in deep learning: a tall stack of dull questions that ends in a sharp answer.
Deep learning at work today
Deep learning works around you today. Your phone unlocks by face: the stack asks about edges, then eyes, then your face. The scan reader at a hospital finds the small shadow a doctor spends years learning to see. The camera in a car reads the road sign through the rain. In every one, a tall stack of tiny questions looks at a grid of numbers and ends with one sharp answer.
Run the checks in your head
Now you run the checks. A carrot sits on my counter. Walk the stack in your head. The first layer asks: round or long? Long. It passes the answer up. The second layer asks: root or leaf? Root. The third layer asks: fruit or vegetable? Vegetable. The fourth layer asks: what is long, a root, and a vegetable? The answer sharpens layer by layer. Now ask yourself which single layer knew. None of them. The stack did. Read the scene again, slower this time, and feel where the answer came from.