Boosting: Updating the weights

• Update the weights using the error
• For each example ex:
  • If the example is correctly classified, wex = wex * (error / (1 - error))
  • If the example is incorrectly classified, leave its weight alone
  • What is the relationship between the weights of correctly and incorrectly classified training examples?
• Once all weights are updated, they must be normalized
  • i.e., the sum of all weights must always equal 1.0
  • Otherwise, the weight update function above would fail (why?)
• Normalization
  • Find the total weight
  • Divide each weight by the total