Learning Generalizable Batch Active Learning Strategies via Deep Q-networks (Student Abstract)

To handle a large amount of unlabeled data, batch active learning (BAL) queries humans for the labels most valuable data points at every round. Most current BAL strategies are based on human-designed heuristics, such as uncertainty sampling or mutual information maximization. However, there exists disagreement between these heuristics and ultimate goal BAL, i.e., optimizing model's final performance within query budgets. This leads to limited generality heuristics. this end, we formulate an MDP propose data-driven approach deep reinforcement learning. Our method learns strategy by maximizing performance. Experiments UCI benchmark show that our can achieve competitive compared existing heuristics-based approaches.