Variants of self-organizing maps

Abstract

Self-organizing maps have a bearing on traditional vector quantization. A characteristic that makes them more closely resemble certain biological brain maps, however, is the spatial order of their responses, which is formed in the learning process. A discussion is presented of the basic algorithms and two innovations: dynamic weighting of the input signals at each input of each cell, which improves the ordering when very different input signals are used, and definition of neighborhoods in the learning algorithm by the minimal spanning tree, which provides a far better and faster approximation of prominently structured density functions. It is cautioned that if the maps are used for pattern recognition and decision process, it is necessary to fine tune the reference vectors so that they directly define the decision borders.

Self‐Organising Map Principles Applied Towards Automating Road Extraction from Remotely Sensed Imagery

“…However, it is also possible to define the topological relationships among the nodes in the input space instead of in the network (Kangas et al, 1990;Kohonen, 2001). The adaptability allowed by defining network topology in the input space proved to be better suited for constructing road network topology.…”

A note on self-organizing semantic maps

“…2(a)-(c). Since owl and hawk and horse and zebra have the same attribute vector representations (columns (5,6) and (14,15) in Table I), they are mapped to the same vectors by both extraction methods. In [6], the normalization of the symbol part of each datum for these two pairs of animals causes their images to occupy different cells as seen in Fig, 2(a), which seems artificial at best-after all, these animals have identical attribute (context) vectors.…”

A batch self-organizing maps algorithm based on adaptive distances

“…In Kangas et al [11] two novel SOM-based approaches were presented: one based on dynamic weighting of the input signals at each input of each neuron, which improved the ordering when very different input signals are used; the other based on a definition of neighborhoods in the learning algorithm by the minimal spanning tree, which provided a better and faster approximation of prominently structured density functions. Grozavu et al [12] presented two new approaches based on stochastic self-organizing maps which had both clustering and weighting capabilities: local weighting observation (lwo-SOM) which adapted weights to filter observations during learning process; local weighting distance (lwd-SOM), which weighted distance between observations and prototypes.…”