A Convolutional Deep Self-Organizing Map Feature extraction for machine learning

In this work we propose a new Unsupervised Deep Self-Organizing Map (UDSOM) algorithm for feature extraction, quite similar to the existing multi-layer SOM architectures. The principal underlying idea of using SOMs is that if a neuron is wins n times, these n inputs that activated this neuron are similar. The basic principle consists of an alternation of phases of splitting and abstraction of regions, based on a non-linear projection of high-dimensional data over a small space using Kohonen maps following a deep architecture. The proposed architecture consists of a splitting process, layers of alternating self-organizing, a rectification function RELU and an abstraction layer (convolution-pooling). The self-organizing layer is composed of a few SOMs with each map focusing on modelling a local sub-region. The most winning neurons of each SOM are then organized in a second sampling layer to generate a new 2D map. In parallel to this transmission of the winning neurons, an abstraction of the data space is obtained after the convolution-pooling module. The ReLU is then applied. This treatment is applied more than once, changing the size of the splitting window and the displacement step on the reconstructed input image each time. In this way, local information is gathered to form more global information in the upper layers by applying each time a convolution filter of the level. The architecture of the Unsupervised Deep Self-Organizing Map is unique and retains the same principle of deep learning algorithms. This architecture can be very interesting in a Big Data environment for machine learning tasks. Experiments have been conducted to discuss how the proposed architecture shows this performance.