Unsupervised Learning and Clustering Using a Random Field Approach

In this work we propose a random field approach to unsupervised machine learning, classifier training and pattern classification. The proposed method treats each sample as a random field and attempts to assign an optimal cluster label to it so as to partition the samples into clusters without a priori knowledge about the number of clusters and the initial centroids. To start with, the algorithm assigns each sample a unique cluster label, making it a singleton cluster. Subsequently, to update the cluster label, the similarity between the sample in question and the samples in a voting pool and their labels are involved. The clusters progressively form without the user specifying their initial centroids, as interaction among the samples continues. Due to its flexibility and adaptability, the proposed algorithm can be easily adjusted for on-line learning and is able to cope with the stability-plasticity dilemma.