Abstract

We propose a variation of the self organizing map algorithm by consideringthe random placement of neurons on a two-dimensional manifold, following a bluenoise distribution from which various topologies can be derived. Thesetopologies possess random (but controllable) discontinuities that allow for amore flexible self-organization, especially with high-dimensional data. Theproposed algorithm is tested on one-, two- and three-dimensions tasks as wellas on the MNIST handwritten digits dataset and validated using spectralanalysis and topological data analysis tools. We also demonstrate the abilityof the randomized self-organizing map to gracefully reorganize itself in caseof neural lesion and/or neurogenesis.

 

Quick Read (beta)