Gradual Feature Extraction by Repeated Information Maximization

R. Kamimura and T. Kamimura (Japan)


Information Maximization, Competitive Learning, Multilayered Networks


In this paper, we propose a new multi-layered network model where information is always maximized in each layer. Networks are composed of several competitive layers. In each competitive layer, information is maximized. This successive information maximization enables networks to extract features gradually. We applied the new method to vertical and horizontal line detection and a medical data problem. Experimental results con firmed that information can successively be maximized in multi-layered networks, and the networks can gradu ally extract features.

Important Links:

Go Back