Parallalizable Deep Self-Organizing Maps for Image Classification

Self-organizing Maps (SOMs) are neural network architectures which are used to learn from unlabeled data. Despite being proven to be useful in many areas, SOMs have limited capability of performing high level feature abstraction due to its shallow structure. As a solution, deep self-organizing maps (DSOM), have been proposed. DSOMs enable multiple levels of abstraction in unsupervised learning with a hierarchical deep structure. However, training of DSOMs is computationally expensive, limiting its usability. This paper presents a DSOM architecture that is easily parallelizable and hence more computationally efficient (PD-SOM). The presented architecture has three main advantages: 1) Unsupervised learning based image classification, 2) High-level feature abstraction and 3) Less computationally expensive training while maintaining high accuracy levels. The PD-SOM architecture was implemented on the benchmark MNIST hand written character dataset. To test the robustness and the generalization capability of the presented PDSOM architecture, testing was done with: 1) small training sets and 2) varying degrees of noise. It was shown that the presented architecture consistently outperformed the previously proposed DSOM while showing similar to 18% decrease in training time for the MNIST dataset.