AN ew -Groups Neural Network
Jui-Cheng Yen · 2002
In this paper, a new neural-network model called GROUPSTRON is proposed to identify the groups' elements from a data set. Based on both the divide-and-conquer principle and the coarse-and-fine competition, GROUPSTRON divides the identification process into rounds and then sequentially identifies each group's elements from the data set. All the elements in the first group are larger than those in the second group and this relationship holds for the successive groups. The proof that GROUPSTRON converges to the correct state in every situation is also given in this paper. Moreover, the convergence rates of GROUPSTRON for three special data distributions are deduced. Finally, simulation results are given to demonstrate the effectiveness and design philosophy of GROUPSTRON. NTERESTS in artificial neural networks (ANNs) in the past two decades are driven by its attractive features such as learning capability, massive parallel computing capability, robustness, etc. Competitive learning networks (1)-(5) play a very important role in ANNs and have become a fundamental building block of many complex systems that have been extensively adopted to solve problems in pattern classification, telecommunications, etc. (4), (5). The -groups networks, es- pecially the -winners-take-all ( -WTA or 1-group) networks (6)-(10) and 1-WTA networks (11)-(14), are the key elements in competitive learning. The -groups problem is to identify the groups' elements from a data set. All the elements in the first group are larger than those in the second group and this relationship holds for the successive groups. A network for identifying the groups' elements from a data set is called a -groups network. In this paper, a new -groups network called GROUP- STRON, which is a generalization of WINSTRON (6), is proposed. The network design is based on both the divide-and-conquer principle and the coarse-and-fine competi- tion. The process to identify the groups' elements is divided into rounds and all the rounds are sequentially conquered via the coarse-and-fine competition. The ideas of competition design are based on the assumption that neurons of the human brain are self-organized to perform pattern recognition. This