Neural networks
Start with Bob’s chilly room. One tiny rule decides how hard to heat it. Put a few rules together and they can draw a shape. Stack them and they can build useful answers from simpler ones. Show them mistakes, and their adjustable numbers move.
Each lesson can be read on its own; the video frames are illustrations, not prerequisites. Move the temperature, bend the shape, and take a training step yourself.
Lessons
Each stop answers one question that the previous stop creates. You can read straight through or revisit a single experiment.
Overview
Bob’s temperature goes into one adjustable rule. Its bend makes a ramp. Several ramps add into a shape. Layers turn early answers into more useful ones. Training compares the network’s answer with examples, sends a correction backward, and moves the adjustable numbers a little. Then it tries again.
Weights and biases are the numbers that can change. The bend is a fixed rule.
Backward messages tell each one which way matters; small steps change them together.
Try explaining it without the special words: What does one number do, why connect the little rules, and how does a mistake change them?
Back to the beginning