Generalized Adaptive Boosting and the Geometry of Mistakes
The paper characterizes which joint false-positive and false-negative guarantees allow weak learners to be boosted.
Its framework keeps the two types of error separate, allowing guarantees that reflect their different impacts. The authors give a complete geometric characterization that includes classical boosting and a recent multi-objective extension as special cases. They also present an adaptive boosting algorithm that does not need advance knowledge of the learner’s guarantee set. A further result establishes confidence amplification using the centerpoint theorem from discrete and computational geometry.
ArXiv · AI/CL/LG's note
Its framework keeps the two types of error separate, allowing guarantees that reflect their different impacts. The authors give a complete geometric characterization that includes classical boosting and a recent multi-objective extension as special cases. They also present an adaptive boosting algorithm that does not need advance knowledge of the learner’s guarantee set. A further result establishes confidence amplification using the centerpoint theorem from discrete and computational geometry.
ArXiv · AI/CL/LG's note
score 4