Shaping latent representations using Self-Organizing Maps with Relevance Learning

Recent work indicates that Deep Clustering (DC) methods are a viable option for unsupervised representations learning of visual features. By combining representation learning and clustering, traditional approaches have been shown to build latent representations that capture essential features of the data while preserving topological characteristics. In this sense, models based on Self-Organizing Maps models with relevance learning (SOMRL) were considered as they perform well in clustering besides being able to create a map that learns the relevance of each input dimension for each cluster, preserving the original relations and topology of the data. We hypothesize that this type of model can produce a more intuitive and disentangled representation in the latent space by promoting smoother transitions between cluster points over time. This work proposes a representation learning framework that combines a new gradient-based SOMRL model and autoencoders. The SOMRL learns the relevance weights for each input dimension of each cluster. It creates a tendency to separate the information into subspaces. To achieve this, we designed a new loss function term that weighs these learned relevances and provides an estimated unsupervised error to be used in combination with a reconstruction loss. The model is evaluated in terms of clustering performance and quality of the learned representations and then compared with start-of-the-art models, showing competitive results.