Unsupervised learning algorithms realizing topographic mappings are justified by neurobiology while they are useful for multivariate data analysis. In contrast to supervised learning algorithms Unsupervised neural networks have their objective function implicitly defined by the learning rule. When considering topographic mapping as an optimization problem, the presence of explicitly defined objective functions becomes essential. In this paper, we show that measures of neighborhood preservation can be used for optimizing and learning topographic mappings by means of evolution strategies. Numerical experiments reveal these measures also being a possible description of the principles governing the learning process of Unsupervised neural networks. We argue that quantifying neighborhood preservation provides a link for connecting evolution strategies and Unsupervised neural learning algorithms for building hybrid learning architectures.
展开▼
