Boosting: Weighting the hypothesis

• So far, we have only used weights for examples
• But the weak learners also have weights
  • log((1 - error) / error)
• Why do the weak learners need weights?
  • After training the first learner, we generated new weights for the examples
  • We then train a second learner, which must somehow be biased by the weights
  • We can repeat this indefinitely, generating an ensemble of weak learners
  • The members of the ensemble then vote on the proper label for a newly presented example
• What weight would correspond to an error of zero?