Improving Performance of Self-Organising Maps with Distance Metric Learning Method

摘要

Self-Organising Maps (SOM) are Artificial Neural Networks used in PatternRecognition tasks. Their major advantage over other architectures is humanreadability of a model. However, they often gain poorer accuracy. Mostly usedmetric in SOM is the Euclidean distance, which is not the best approach to someproblems. In this paper, we study an impact of the metric change on the SOM'sperformance in classification problems. In order to change the metric of theSOM we applied a distance metric learning method, so-called 'Large MarginNearest Neighbour'. It computes the Mahalanobis matrix, which assures smalldistance between nearest neighbour points from the same class and separation ofpoints belonging to different classes by large margin. Results are presented onseveral real data sets, containing for example recognition of written digits,spoken letters or faces.