Hierarchical Boosting and Filter Generation
par/by Marc-Olivier LaBarre
Département d'Informatique et de Recherche Opérationnelle
Université de Montréal
Boosting is a well-known learning meta-algorithm that constructs a strong hypothesis by iteratively compiling a score of weighted weak hypotheses. Boosting and the well-known AdaBoost have been applied succesfully on many benchmark datasets, and numerous modifications have been proposed to expand the reach of such algorithms. However, little have been said about the structure with which Boosting and AdaBoost builds its strong hypothesis. The basic version of AdaBoost can be shown as constructing a single layer of weak hypotheses, in a way similar to a neural network.
This talk will present a way to have AdaBoost include multiple layers in the construction of a strong hypothesis, with the goal to improve the reach of the algorithm itself. We also present a way to group features through filters inspired from image processing methods like the Harr-like features used by Viola and Jones for face recognition.