Results

**1 - 1**of**1**### Self-Organisation of Hypercolumns based on Force-Directed Clustering

"... Attractor neural networks are often evaluated and found to perform well on sparse random patterns. However, real world data often have a large amount of correlation, which tends to decrease the performance of these neural networks dramatically. This thesis describes the first steps in the creation o ..."

Abstract
- Add to MetaCart

Attractor neural networks are often evaluated and found to perform well on sparse random patterns. However, real world data often have a large amount of correlation, which tends to decrease the performance of these neural networks dramatically. This thesis describes the first steps in the creation of a hidden layer for preprocessing data to address these problems. We describe a novel algorithm for non-parametric clustering, which we apply in the preprocessing step in order to find correlations in the input data. The algorithm uses mutual information as measure and interprets this measure as a force, which is later used in a force-directed clustering. The algorithm is found to work as expected and is migrated to a parallel cluster computer. The parallel performance is very good and the algorithm scales almost linearly with the number of processors used. Självorganisation av hyperkolumner med hjälp av kraftbaserad klustring