R. Kamimura and O. Uchida (Japan)
Information maximization, Greedy algorithm, Competitive learning, Gaussian activation function
In this paper, we propose a new computational method for a network-growing method called greedy network growing[1]. We have so far introduced a network-growing algorithm called greedy network-growing based upon in formation theoretic competitive learning. For competitive unit outputs, we have used the inverse of the squares of Eu clidean distance between input patterns and connections. The algorithm has extracted very faithful representations of input patterns. However, one problem is that learning is very slow, and sometimes ambiguous final representations are obtained. To remedy these shortcomings, we introduce a new activation function, that is, Gaussian activation func tions for competitive units. By changing a parameter for the Gaussian activation functions, we can build a network that does not focus on faithful representations of input patterns, but try to extract the main characteristics of input patterns. Because this method are not concerned with detailed parts of input patterns, learning is significantly accelerated and salient features should be extracted. We applied the method to an information education problem. Experimental results confirmed that learning was significantly accelerated and salient features could be extracted.
Important Links:
Go Back
