Intelligent Engineering Systems through Artificial Neural Networks
68 Hybrid Method for Better Active Learning
Download citation file:
Most natural learning systems are not simply passive learners. To apply the Nature's paradigm on inference algorithms, active learning has been proposed. This approach intervenes in the selection of training data dynamically. The previously developed active learning methods are classified into two main categories: the model-dependent and the space-filling methods. Both methods have deficiencies; the former ones may sample data from the same areas, the latter ones may collect unusual or unimportant data. This paper combines these two methods into one hybrid algorithm that not only avoids some deficiencies of its originators, but instead takes advantage of their merits. In contrast to the greedy algorithms that are commonly used for active learning, this approach removes unimportant data. Our experimental evaluation supports the theoretical analysis. The generated training datasets not only represent the data space of interest comprehensively, but also they maximize the information content.