Robust continuous learning in a WTA neural network for clustering symbol strings

K-means and the SOM are two well known algorithms that can be applied to the continuous learning of data. However both implicitly make the assumption that the inputs to the learning are independent and identically distributed (iid) which facilitates the choice of learning parameters. The probability distribution of iid inputs with a cluster structure is modelled by a static mixture model while in the non-iid case a dynamic mixture model is used. The K-SCM (symbol string clustering map) algorithm is described as a robust means of clustering symbol string data requiring no time varying learning rate and hence does not assume that the inputs are iid.