2 The Learning Task  Given m training examples and their labels  Predict the label of a new example 3 The Idea of Boosting  Combine many "moderately inaccurate" base classifiers into a combined predictor  Generate a new base classifier in each round  Constantly focus on the hardest examples  The final predictor is the weighted vote of the base classifiers  AdaBoost sets voting weights of each new base classifier to reduce an upper bound on the training error.