Integrate active and meta-cognitive learning with extreme learning machine

Abstract

"The extreme learning machine has become very popular due to its fast training and good generalization performance. The two considerations for extreme learning machine are fixed network structure and redundancy in training data. In this thesis, we propose efficient semi-supervised learning algorithm to integrate active and metacognitive learning with extreme learning machine to arrive at non-redundant data and optimal network structure for better classification. Semi-supervised learning makes use of unlabeled data along with small labeled data for training. Active learning a special case of semi-supervised learning is applied to select non-redundant data and to reduce the labeling costs. In active learning, the algorithm can iteratively query the user for label of new data points. Metacognitive learning proposes the addition of neuron while training which is used to achieve optimal network structure. Also, the use of regularization has shown improvement in accuracy. The proposed approach is evaluated for classification on various benchmark datasets from UCI machine learning repository. This datasets has features extracted for various real world problems.The proposed method is compared with three state-of-the-art methods based on accuracy, training time and amount of data required for training. Performance evaluation shows that proposed approach gives similar or better accuracy than other approaches with reduced training time."