Train a 1960 neural network
ADALINE was one of the first machines that learned from examples. Place some points, press Train, and watch it find the line that separates them, using the same rule Bernard Widrow and Ted Hoff published in 1960.
1. Add points
2. Train
Try the hard data: when the classes overlap, no straight line can separate them. ADALINE still finds the best compromise. That limit is why later networks added hidden layers.
What you're watching
ADALINE takes each point's two coordinates, multiplies them by two weights, adds a bias, and calls the point Class A if the result is positive and Class B if it's negative. The line on the canvas is where the result is exactly zero.
Training nudges the weights after every point, in proportion to how wrong the raw output was. Big mistakes cause big corrections; near-misses cause small ones. Repeat that over all the points (one epoch), and the error shrinks until the line settles.
That idea, adjusting weights to reduce error, is still how today's largest networks learn. See how it works, step by step.