Learning discriminative representations for deep clustering
摘要
Clustering is of central importance to many computer vision applications such as image understanding, indexing, searching, and product quantization. Recent advances in deep neural networks have made possible many propitious solutions for clustering analysis. This paper presents a deep autoencoder model, namely CDC (Correlation losses for Deep Clustering), to learn discriminative representations for clustering. For such a purpose, we have utilized the distance-based correlation in the probabilistic space to design novel loss functions. The designed correlation loss is incorporated at the outputs of encoder and decoder for local structure preservation. Noticeably, a novel objective function that captures the probabilistic affinity space is presented to enhance the latent representations targeted for an improved clustering quality. Moreover, we utilize a clustering-oriented loss based on Kullback-Leibler (KL) divergence to enhance the latent features. The proposed model has been validated by extensive experiments in comparison with the state-of-the-art methods and demonstrated strong scores in clustering accuracy on popular benchmark datasets: Reuters-10K (