Active Learning : A way to boost ML Models with Limited data

Just as students do not learn much just sitting in classes and listening to teachers , memorizing pre-packaged assignments, and spitting out answers. In order to make a significant impact they must share their learning experience , talk about it , write about it ( just as i am doing today) , relate to the experiences. Active learning activities can be done by groups , pairs or by an entire class .

Check this notebook for QSAR based Active Learning example .

In Machine Learning, the challenge of learning from a little labeled data
instances is that the traditional supervised machine learning algorithms may produce the undesired results because instances may not be a good representative of the dataset. This process significantly reduces the label-
ing efforts, and the research have led to two learning techniques i) semi supervised learning ii) Active Learning .

AL performs nicely with the small amount of labelled data by fully utilizing the massive amount of unlabeled data through the query mechanism. It reduces the learning and annotation cost of data. Active learning is most appropriate when there are numerous unlabeled data instances and can be easily synthesized, and one can anticipate having to label many of them to train an accurate system. The key difference of active and semi-supervised learning is that in active learning learner chooses the instances (queries) from the unlabeled dataset, about which…