A Novel Incremental Class Learning Technique for Multi-class Classification
In this paper, a novel technique for multi-class classification, which is independent of the number of class constraints and can learn the new classes it encounters, is developed. The developed technique enables remodelling of the network to adapt to the
- PDF / 454,899 Bytes
- 8 Pages / 439.37 x 666.142 pts Page_size
- 82 Downloads / 261 Views
Marine Engineering College, Dalian Maritime University, Dalian, China [email protected], [email protected] 2 School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore, Singapore {vijayakr001,raja0046}@e.ntu.edu.sg
Abstract. In this paper, a novel technique for multi-class classification, which is independent of the number of class constraints and can learn the new classes it encounters, is developed. The developed technique enables remodelling of the network to adapt to the dynamic nature of non-stationary input samples. It not only can learn the new classes, but also the new patterns created in the input. The proposed algorithm is evaluated using various benchmark datasets and a comparative study of classification performance shows that the proposed algorithm is superior. Keywords: ELM OS-ELM Incremental class learning learning Multi-class classification
Sequential
1 Introduction Classification problems are one of the age-old problems in the computational intelligence community. In classification, the problem of grouping input sequences into one of more than two disjoint classes is called multi-class classification. Based on learning paradigms, multi-class classification problems are divided into two categories. The first one is batch learning where the entire input data are available for training which takes place throughout the availability of data once. Another one is sequential learning where the initial training takes place with available data and later, training happens continuously whenever new datum arrives. In the recent years, much research on computation intelligence pertaining to sequential learning has been reported [1–3]. In multi-class classification, sequential learning algorithms can learn the new data of a fixed number of classes which it has been trained initially. However, whenever a new class is encountered, the network is to be retrained with the entire set of data, which makes it very time consuming. In applications like cognitive robotics, the number of classes varies and the trained data may not be available to retrain the network with the new class encountered. This makes the existing sequential learning algorithms unsuitable for these applications. In this paper, a novel algorithm, which can adapt to the new classes it encounters and changes to the input pattern by automatically changing the architecture of the © Springer International Publishing Switzerland 2016 L. Cheng et al. (Eds.): ISNN 2016, LNCS 9719, pp. 474–481, 2016. DOI: 10.1007/978-3-319-40663-3_54
A Novel Incremental Class Learning Technique
475
network, is proposed. The proposed algorithm updates the network without losing the knowledge of the previous training. The proposed algorithm is different from the CIELM of [4] where it does not consider changes in the input patterns created during the addition of a new class. The Class-Incremental Learning was proposed using the support vector machines [5] where the problem is considered as binary classification. A new class is considered
Data Loading...
