How it learned

The Perceptron's real power wasn't recognition. It was improvement.

Nobody told it what to look for in an image. Nobody wrote rules for recognising shapes. The machine started with its weights set roughly, looked at examples, and corrected itself when it got one wrong.

Not a big shift. A small one. Then the machine looked at the next shape.

After enough examples, the weights settled into a position that worked. Not because anyone set them there. Because the machine had corrected itself, over and over, until the errors stopped.

This is what Rosenblatt meant by learning. Not a programmer deciding what matters. The machine deciding, through repetition and correction.