What is Backpropagation?

It's the single most important algorithm for training modern AI. But you don't need complex math to understand it. You just need to think like an archer.

Step 1: The First Attempt (The "Forward Pass")

Imagine an AI is shown a picture of a cat for the first time. Its goal is to hit the "cat" bullseye. It makes a guess based on its initial, random settings. This is the **Forward Pass**. Like an archer's first shot, it's probably going to miss.

Step 2: Measuring The Miss (The "Loss Function")

The AI compares its guess ("dog") to the correct label ("cat"). It then calculates the "error"--how far off its prediction was. The archer does the same thing. They don't just see a miss; they measure it: "I was 10cm high and 5cm to the left."

Step 3: The Correction (The "Backward Pass")

This is the magic. The AI sends the error signal backward through its network. It asks: "Which of my internal 'knobs' contributed most to this specific error?" The archer does this intuitively, thinking: "Because I missed high and left, I must adjust my aim lower and to the right."

Step 4: The Next Attempt (Iteration)

With its internal "knobs" slightly adjusted, the AI makes another guess on the next piece of data. This time, it's a little closer. After repeating this cycle thousands or millions of times, the archer's aim becomes perfect. The AI has learned.

The Heartbeat of Learning

This loop--Forward Pass, Calculate Error, Backward Pass, Adjust Weights--is the fundamental heartbeat of how most AIs learn. It's a simple, elegant process of guided trial and error, repeated at a massive scale.

← Back to AI Glossary