Dynamics of self-organized feature mapping
The dynamics of the feature maps created by Kohonen's algorithm is studied by analyzing the spectral density of synaptic fluctuations both analytically and by means of computer simulations. We consider unsupervised learning as a stochastic process and investigate the usefulness of the Fokker-Planck approach for the case of a topological mismatch between input and output space. A breakdown of the Fokker-Planck description is observed if the mismatch exceeds a critical value.
Unable to display preview. Download preview PDF.