ImitAL: Learned Active Learning Strategy on Synthetic Data

Active Learning (AL) is a well-known standard method for efficiently obtaining annotated data by first labeling the samples that contain most information based on query strategy. In past, large variety of such strategies has been proposed, with each generation new increasing runtime and adding more complexity. However, to best our knowledge, none these excels consistently over number datasets from different application domains. Basically, existing AL are combination two simple heuristics informativeness representativeness, big differences lie in often conflicting heuristics. Within this paper, we propose ImitAL, domain-independent novel strategy, which encodes as learning-to-rank problem learns an optimal between both We train ImitAL large-scale simulated runs purely synthetic datasets. To show was successfully trained, perform extensive evaluation comparing strategy 13 datasets, wide range domains, 7 other strategies.