defrance@uni2a.unige.ch (03/26/91)
Hello world. I need to find any reference to studies about the optimal size of the competitive layer in a Kohonen network; particularly I'd like to know which relationship exists between the size of the competitive layer, the size of the input layer and the number of patterns. The problem is the classification of 2D electrophoresis gels. I *think* that this problem is unpractical for Kohonen nets (it works quite well with backprop), because the size of the patterns is typical ~3000 (real numbers), there is a lot of random or noisy spots (probably +90%), and between 20 and 100 gels would be used for training. Even if there weren't theoretical problems with this kind of data, more than ~10 units in the competitive layer would be unmanageable (in time). That's for the empirical reasons why I believe that Kohonen is not suited to this application. The problem is that this work (the application of NN to the classification of bi-dimensional electrophoresis) is the biggest part of one of my students master thesis, and I like to have something more formal for him. Any pointer? remarks? Thank you in advance. - massimo ----------------------------------------------------------------------------- Massimo de Francesco e-mail: massimo@cui.unige.ch Research and teaching Assistant voice : +4122 787-6580 CUI - University of Geneva fax : +4122 735-3905 CH-1207 GENEVA (Switzerland)