Ενεργή μάθηση με μηχανές διανυσμάτων στήριξης
Active learning with support vector machines
In the field of Machine Learning all learning methods require a substantial amount of labeled data in order for the model to be properly fitted. In today’s world of IoT (Internet of Things) and Big Data, where everything is controlled and monitored by software applications, unlabeled data are very easily acquired as they are continuously generated. However, the process of finding and annotating the true class to those dataset’s instances, often requires more effort and time than the actual training of the model. Active learning aims to tackle this problem by enabling machine learning algorithms to perform equally well without reliance on the existence of huge training datasets. To accomplish this, an active learning algorithm is allowed to query an oracle (usually a human expert) for the true label of an unlabeled training example. There are a number of different strategies as well as learning scenarios that can be followed for this interaction which will be presented in later sections of this report. Active learning algorithms are basically wrapping around traditional supervised learning methods such as Support Vector Machines (SVMs), Logistic Regression etc. Apart from the topic of Active Learning, this report offers a walkthrough of the theory behind Support Vector Machines and tries to present the various researched methods that combine these two topics.