Deep Evidential Clustering of Images

This paper presents a new image clustering method (DEEM) based on convolutional neural networks and the theory of belief functions used to encode uncertainty between clusters. The algorithm learns to generate mass functions for a given image through a training process that minimises a loss between the conflict computed from pairs of images and their dissimilarities. DEEM extends NN-EVCLUS and provides a gateway to the entire realm of deep learning, capitalising on all its advancements. It enables the full exploitation of the benefits offered by customisable layers, sophisticated optimisation algorithms, and other state-of-the-art techniques. DEEM can learn from the data itself, without requiring external labels but we can incorporate prior on labels if available as proposed in NN-EVCLUS. The first results are shown on the MNIST dataset (digit recognition).