Binary Clustering for Deep Network Trained by Feature Growth

Clustering, a class of unsupervised learning methods, has been extensively studied and applied in machine learning. By designing the training process, we are able to impose the CNN with the ability of clustering. In this paper, we propose a binary clustering framework by implementing the deep network in the clustering process, instead of barely acting as a feature extractor. The proposed training strategy consists of five stages: feature elimination, feature seeding, feature germination, feature growing and feature grafting. Feature elimination works to eliminate the impact of random weights initialization and leave the network unbiased. Feature seeding, feature germination and feature growing altogether realize similarity generation, while feature grafting filters and strengthens the generated similarity. Because of the uncertainty of clustering, we take genres into consideration as evaluation metrics. Only performance analysis is considered while the first and the second samples selected during feature training belong to different genres. Compared with DEC and DEC-DA, our proposed training strategy is shown to achieve the state-of-the-art performance, with clustering accuracy of up to 0.996 for dataset fashion MNIST 4/5 -full and outperforms other methods in terms of stability.