Efficient Unsupervised Clustering with Variational Information Bottleneck: A Fusion of Contrastive Learning and Alternating Optimization

We propose an unsupervised clustering method based on VIB framework, which is composed of encoder, decoder and cluster. The encoder maps the input to a representation of the potential space; A clusterer groups representations; The decoder reconstructs the representation into the original input, using the distance between the input and the reconstructed output to reverse train the network. We introduce an efficient representation learning framework through combinatorial contrast learning and information bottleneck techniques. We also combine a self-expression layer with a multimodal decoder, which is consistent with many methods based on spatial fusion. In addition, we use advanced cluster optimization techniques to alternately optimize the network training process and the clustering process. A large number of experiments on four datasets show that this proposed characterization further outperforms traditional autoencoder-based clustering methods and some methods using similar principles [9].