Standard simplex induced clustering with hierarchical deep dictionary learning

Clustering remains a fundamental but crucial task on unsupervised pattern recognition. A recent novel deep dictionary learning system where hierarchical discriminative dictionary learning layers are embedded within a neural network showed either competitive or state-of-the-art results for image classification. We now propose a new method for image clustering, extending hierarchical deep dictionary learning to the unsupervised learning setting. Clustering is induced in the neural network in a simple way by requesting vector outputs associated to categories to approximately lie in the corners of a standard simplex and by approximating a given expected value on such vectors. This is carried out in a scalable way due to the batch-wise updates as clustering is learned in the deep system. We evaluate our proposal on four benchmark datasets and we either achieve competitive results or outperform state-of-the-art clustering methods. Under the unbalanced scenario clustering is known to be more challenging. For this our method additionally considers Bayesian updates to automatically learn the correct cluster proportions for a small number of clusters.