Please note: This seminar will take place online.
Ben Armstrong, PhD candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Kate Larson
I design an iterative process of learning classifier performance and removing redundant classifiers from an ensemble by having them "delegate" to another classifier which reduces the computational cost of training compared with training a full ensemble. Conceptually, liquid democracy enables the removal and reweighting of ensemble members in a way that empowers more accurate classifiers while retaining diversity and not centralizing weight to a harmful extent.
I show numerically the amount of computation saved through this delegative pruning method compared with full ensemble training. Via simulation on existing datasets I explore the differences in performance between a variety of delegation methods, demonstrate that our approach can be more accurate than a full ensemble, and compare with existing boosting methods.
To attend this PhD seminar on Zoom, please go to https://uwaterloo.zoom.us/j/96023562868.