Unsupervised Feature Learning using Self-organizing Maps

In recent years a great amount of research has focused on algorithms that learn features from unlabeled data. In this work we propose a model based on the Self-Organizing Map (SOM) neural network to learn features useful for the problem of automatic natural images classification. In particular we use the SOM model to learn single-layer features from the extremely challenging CIFAR-10 dataset, containing 60.000 tiny labeled natural images, and subsequently use these features with a pyramidal histogram encoding to train a linear SVM classifier. Despite the large number of images, the proposed feature learning method requires only few minutes on an entry-level system, however we show that a supervised classifier trained with learned features provides significantly better results than using raw pixels values or other handcrafted features designed specifically for image classification. Moreover, exploiting the topological property of the SOM neural network, it is possible to reduce the number of features and speed up the supervised training process combining topologically close neurons, without repeating the feature learning process.