High Resolution Self-organizing Maps

Abstract

Kohonen’s self organizing feature map (SOM) provides a convenient way for visualizing high dimensional input features by projecting them onto a low dimensional display space. This map has an appealing characteristic: feature vectors close to one another in the high dimensional input space remain close to one another in the low dimensional display space. Owing to the computational requirements, the display space so far remains of relatively low resolutions. In this paper, we provide an implementation of the SOM by making use of the highly parallel architecture of a graphic processing unit to increase its computational speed to allow a substantial increase in the resolution while keeping the computation to within an acceptable wall clock time. Armed with such an implementation, we find that the high resolution SOM can display intricate details concerning the relationships among the input feature vectors. These details would be lost if a low resolution SOM was deployed. The capability of the high resolution SOM is demonstrated through an application to an artificially generated dataset, the policeman dataset. The dataset allows us to design intricate relationships among the input feature vectors.