Learning to Cluster

Benjamin Bruno Meier, Thilo Stadelmann, Oliver Dürr

Abstract: We propose a novel neural network architecture to learn the task of clustering end-to-end: salient features for any similarity criterion specified through weakly labeled training data are extracted with an embedding network; during evaluation, the network groups similar data of any modality together, by assigning a probabalistic cluster index, and further gives a probabilistic estimate of the number of clusters. The method is evaluated on 2D point data, speaker data from the TIMIT corpus, and images from the COIL-100 dataset, reaching promising results.

TL;DR: We propose a novel architecture for probabalistic clustering which can be trained in an end-to-end fashion requiring only weakly labeled data.

Keywords: Deep Learning, Clustering, Learning Representations