Extreme learning machine based novelty detection for incremental semi-supervised learning

A variety of problems are related to streaming data e.g. infinite length, concept-drift, non-linearly separable classes, and the possible emergence of “novel classes”. We propose a semi-supervised learning method using an incremental neural network to cope with all these problems. Tracking the concept drift is maintained by using incremental learning. Additionally, the extreme value theory is used as a novelty detector technique to recognize outliers, since the semi-supervised learning is sensitive to them. The extreme learning machine is easily updated and it can be used for multiple classes. Superior properties are shown for the proposed algorithm as compared with an auto-encoder neural network. Particularly, the training time is greatly reduced hence it is adequate for online training.