Meta learning for unsupervised clustering

Learning an embedding space is essential in clustering. Deep learning has been used recently for this purpose, yielding impressive clustering results. However, it remains challenging to discover clusters in small data, which are insufficient to train deep networks. To address this challenge, we adopt the meta learning strategy, which learns to learn new tasks efficiently. We propose a novel meta learner, called MLC-Net, which mimics numerous clustering tasks during the training to learn an effective embedding space for new clustering tasks. MLC-Net has three building blocks: encoder, centroid, and prediction blocks. The encoder block transforms input patterns into an embedding space, while the centroid block estimates a representative feature for each cluster, called pseudo-centroid. It makes the embedding space more effective and more reliable, by learning an embedding space and a pseudo-centroid estimator jointly. Extensive experimental results on the Omniglot, MNIST, and Mini-ImageNet datasets demonstrate that MLC-Net achieves the state-of-the-art unsupervised clustering, as well as few-shot classification, performances.