A two layered neural network is considered as Kohonen's dot-product type SOM model. It defines pattern classifiers through step by step self-organization. This note examines the role of learning rate /spl alpha/ and forgetting rate /spl delta/ in such SOM algorithms. The properties we consider are the relation between stability of winner functions and topographic mapping formation. We propose three classes of networks defining corresponding winner functions. They depend on the two parameters or their ratio K, and the former class includes the latter and the most restrictive class is here called K-topographic. The main result is 1) we can define such topographic networks depending on the ratio K and 2) once a network belongs to this K-topographic class we can maintain the property in the evolution process if we choose /spl alpha/ and /spl delta/ appropriately. Thus the stability and topographic property are related in this generalized SOM algorithm.
[1]
Youichi Kobuchi,et al.
SOM algorithms and their stability consideration
,
2004,
ESANN.
[2]
Gilles Pagès,et al.
Almost sure convergence of the one-dimensional Kohonen algorithm
,
1997,
ESANN.
[3]
Shouji Sakamoto,et al.
Convergence property of topographic mapping formation from cell layer to cell layer through correlation learning rule
,
2000,
Neural Networks.
[4]
Shouji Sakamoto,et al.
Cellular topographic self-organization under correlational learning
,
2003,
ESANN.
[5]
Shouji Sakamoto,et al.
Stability of generalized topographic mappings between cell layers through correlational learning
,
2004,
Neural Networks.