Cost functions for pairwise data clustering

Cost functions for nun-hierarchical pairwise clustering are introduced, in the probabilistic autoencoder framework, by the request of maximal average similarity between input and the output of the autoencoder. Clustering is t hus formulated as the problem of finding the ground state of Potts spins Ha miltonians. The partition, provided by this procedure, identifies clusters with dl:nse connected regions in the data space. (C) 2001 Elsevier Science B.V. All rights reserved.