Christos Ferles, Yannis Papanikolaou, Stylianos P. Savaidis and Stelios A. Mitilineos, University of West Attica, Greece

The Self-Organizing Convolutional Map (SOCOM) combines convolutional neural networks, clustering via self-organizing maps, and learning through gradient backpropagation into a novel unified unsupervised deep architecture. The proposed clustering and training procedures reflect the model’s degree of integration and synergy between its constituting modules. The SOCOM prototype is in position to carry out unsupervised classification and clustering tasks based upon the distributed higher level representations that are produced by its underlying convolutional deep architecture, without necessitating target or label information at any stage of its training and inference operations. Due to its convolutional component SOCOM has the intrinsic capability to model signals consisting of one or more channels like grayscale and colored images.

Deep Learning, Unsupervised Learning, Convolutional Neural Network (CNN), Self-Organizing Map (SOM), Clustering.