Which of the following can be used for clustering of data?

Concept

Clustering is an unsupervised task: the data arrive as plain input vectors with no target label attached, and the learner has to group them purely by similarity. A neural model can therefore cluster only if its learning rule needs no target values at all — that is, only if it learns competitively, letting its units organise themselves around the input distribution instead of correcting an output against a known answer.

Applying it here

A self-organizing map (Kohonen map) is trained in exactly that way. For each input vector x presented to the map:

1. Compare x with the weight vector of every unit on the map grid and pick the closest one — the best-matching unit.

2. Move that unit's weight vector a small step toward x, and move its grid neighbours toward x by a smaller amount, so that nearby units come to represent nearby inputs.

3. Shrink the learning rate and the neighbourhood radius, then repeat over the whole data set until the weights stabilise.

Nothing in this loop consults a target label — only the input vectors themselves. On convergence each unit’s weight vector is a prototype (in effect a cluster centroid), and every input is assigned to the unit that matches it best, so the map partitions the data into clusters and additionally preserves their topology on the grid.

Contrasting the four models

The radial basis function network is worth one extra remark: its hidden-unit centres are very often initialised by running a separate clustering algorithm such as k-means. The clustering there is done by that algorithm, not by the RBF network, so it does not make the network itself a clustering model.

Hence, of the four models offered, the one that can be used for clustering of data is the self-organizing map.