Discriminative Batch Mode Active Learning (M23)
Yuhong Guo, Dale Schuurmans University of Alberta Problem: batch mode active learning select a set of unlabeled instances to query Approach: discriminative, optimistic optimization Select queries by simultaneously guessing their labels and optimizing classifier Objective: likelihood on labeled and guessed data and entropy regularizer on unlabeled data Optimize objective over (relaxed) query-label indicators Compare true labels to guessed values, down-weight unlabeled contribution in the objective if misled