This animation shows three classifiers voting on samples, illustrating how majority voting combines predictions. It builds a grid of samples and errors to demonstrate that under independent errors with 0.3 individual error rate, majority vote error drops to 0.216, while perfectly correlated errors yield no improvement. It distinguishes hard voting, weighted voting, and probability averaging, and outlines bagging, boosting, and stacking as ensemble strategies. Useful for first-year AI students learning why diverse, uncorrelated errors—not just more models—drive ensemble performance.
Narrated · 16:9 · Preview before teaching · automatic layout checks do not establish subject accuracy
# Combining classifiers helps when their errors differ Three classifiers vote on the same sample, producing [+1,+1,-1], so majority voting selects +1. Expand the scene into a grid of samples and model errors to reveal the importance of complementary mistakes. Under the explicit teaching assumption of independent errors, each model has error probability .3, and the majority vote has error 3*(.3 squared)*.7 + .3 cubed = .216. Compare this with perfectly correlated mistakes, where copying the classifier does not reduce its .3 error. Distinguish hard voting, weighted voting, and averaging comparable class probabilities. Briefly organize ensemble strategies into parallel resampling, sequential reweighting, and learning a combination from out-of-fold predictions. Do not present independence as a fact about real models. Use black backgrounds, cyan successful predictions, orange errors, and simple voting diagrams. Useful for understanding why diversity matters more than merely increasing the number of models. Audience: 一年级人工智能专业学生