Neural Clustering: Concatenating Layers for Better Projections

Sean Saito, Robby T. Tan

Feb 17, 2017 (modified: Mar 09, 2017) ICLR 2017 workshop submission readers: everyone
  • Abstract: Effective clustering can be achieved by mapping the input to an embedded space rather than clustering on the raw data itself. However, there is limited focus on transformation methods that improve clustering accuracies. In this paper, we introduce Neural Clustering, a simple yet effective unsupervised model to project data onto an embedded space where intermediate layers of a deep autoencoder are concatenated to generate high-dimensional representations. Optimization of the autoencoder via reconstruction error allows the layers in the network to learn semantic representations of different classes of data. Our experimental results yield significant improvements on other models and a robustness across different kinds of datasets.
  • TL;DR: We introduce a simple yet effective technique for projecting data that allows for better clustering.
  • Conflicts: u.yale-nus.edu.sg
  • Keywords: Deep learning, Unsupervised Learning

Loading