Can Active Learning Experience Be Transferred

摘要: Active learning is an important machine problem in reducing the human labeling effort. Current active strategies are designed from knowledge, and applied on each dataset immutable manner. In other words, experience about usefulness of cannot be updated transferred to improve datasets. This paper initiates a pioneering study whether can transferred. We first propose novel model that linearly aggregates existing strategies. The linear weights then used represent experience. equip with popular upper- confidence-bound (LinUCB) algorithm for contextual bandit update weights. Finally, we extend our transfer across datasets technique biased regularization. Empirical studies demonstrate learned not only competitive most single datasets, but also performance future tasks.