An animated linear classifier tracks its weight vector, bias, and decision boundary as it processes labeled 2D points with classes -1 and +1. Starting from w=(0,0), b=0, each misclassified sample triggers the update rule w ← w + y·x, b ← b + y, shown step by step with recalculated boundaries. The animation highlights that one correction does not guarantee an immediately correct sign, then contrasts a separable dataset with XOR to show why separability is required for convergence. Useful for first-year AI students studying online learning algorithms.
Narrated · 16:9 · Preview before teaching · automatic layout checks do not establish subject accuracy
# A perceptron learns by correcting signed mistakes A linear classifier displays its weight vector, bias, and decision boundary beside two-dimensional training samples. Use labels -1 and +1, and treat zero signed score as requiring an update. Start with w = (0,0), b = 0, and learning rate 1. A positive sample (2,1) changes the parameters to w = (2,1), b = 1. A negative sample (1,0) then has score 3 and changes them to w = (1,1), b = 0. Show that this sample still has the wrong sign after that single update: correction steps do not guarantee immediate success. Link the update w <- w + learning_rate*y*x to the perceptron criterion. Compare a separable point set with XOR, explaining why standard convergence requires separability. Use a black background, colored samples, and boundaries calculated from the displayed parameters. Useful for understanding online updates and their limitations. Audience: 一年级人工智能专业学生