Boosting: Biasing the weak learners
- Recall the purpose of weighting the training examples
- Consistently mis-classified examples get high weights
- Hence, new hypotheses will focus on classifying them correctly
- Decision stumps
- Splitting feature/cutoff: Maximize homogeneity of resulting branches
- To emphasize highly weighted examples:
- Perceptrons
- Another possible weak learner
- How could we weight the training examples here?
(next)