This animation tracks four training samples through two rounds of discrete AdaBoost, showing how misclassified points gain weight while correctly classified ones shrink. It visualizes the exponential update rule, the computation of each weak classifier's vote weight (alpha) from its weighted error, and how combining classifiers via a weighted sign vote corrects some mistakes while leaving others unresolved. Useful for first-year AI students learning why boosting can overemphasize noisy or outlier labels.
Narrated · 16:9 · Preview before teaching · automatic layout checks do not establish subject accuracy
# AdaBoost changes the weight of each training example Four samples have true labels [-1,-1,+1,+1] and initial weights .25 each. The first weak classifier predicts [-1,-1,-1,+1], so its weighted error is .25 and its vote weight is one half of ln(3), approximately .549306. Increase the relative emphasis on the third sample and normalize, producing weights [1/6,1/6,1/2,1/6]. The second classifier predicts [-1,+1,+1,+1]. Its weighted error is 1/6, not 1/4, and its vote weight is one half of ln(5), approximately .804719. Update the sample weights to [.1,.5,.3,.1]. Combine the two classifiers through the sign of their weighted votes: sample 3 is corrected, but sample 2 is still wrong. Show the exponential update formula alongside changing sample sizes. State that this is binary discrete AdaBoost and that noisy labels can attract excessive weight. Use black backgrounds and fixed sample identities throughout. Audience: 一年级人工智能专业学生