Supervised Competition Using Joined Growing Neural Gas
Dušan Fedorčák, Michal Podhorányi · International Conference on Computing Technology and Information Management · 2014
Competitive learning is well-known method to process data. Various goals may be achieved using competi- tive learning such as classication or vector quantiza- tion. In this paper, we present a dierent insight into the principle of supervised competitive learning. An in- novative approach to the supervised self-organization is suggested. The method is based on dierent handling of input data labels which encode the classication. When the label has appropriate format then it is possible to use it within the competitive process in the same way as any input data element. Such approach is as eective as standard supervised methods and has some positive attributes such as the soft classication ability. or industry areas. This article is focused on competitive networks taught by supervision. An alternative to common supervised competitive learning is suggested and it is explained how labeled data may be processed within the standard metrics-based competitive learning. There are some learning methods (such as Learning Vector Quantization) which can utilize the competitive learn- ing principle but are applied to labeled data. They are called supervised competitive learning methods and might be mentioned also as semi-supervised learning. The main dierence between unsupervised competitive learning methods and supervised competitive learning methods is the structure of the input dataset. As usual, the supervision above the learning process must be based on some information added to the system and such information is a class label assigned to each input signal. The learning process is supervised through these labels and the common result of super- vised competitive learning is a classier trained on given patterns i.e. pairs of input signals and their labels. Later on this article, a new method of semi-supervised learning is presented. Our research has been inspired be some very interesting ideas about learning from context. The literature suggests that in some cases it is vital to focus on understanding the input dataset and leave the classication problem until later (7). This leads us to an idea that the pair of input data and its label should be taken as an indivisible object. The joined information may be passed to the network rather than using the label for the output correction, which is the common approach. If this idea is applied to competitive learning, useful results can be obtained. First, a brief description of the standard supervised competitive method (LVQ) will be given, therefore the dierence between the standard approach and our