Unsupervised Embedding Learning by Noisy Similarity Label Optimization

Koki Tsubota, Kiyoharu Aizawa

Abstract: We propose a similarity label optimization framework for unsupervised embedding learning. Existing works use similarity labels obtained from instance labels, which are image identifiers, for unsupervised embedding learning assuming that images of different instance labels have different semantic classes. However, because of the significant semantic gap between classes and instances, instance labels are not enough for learning embeddings. To alleviate this problem, we consider the similarity labels as noisy labels and optimize similarity labels and neural network parameters in an alternating fashion. Experimental results demonstrate that the proposed method outperforms a baseline method by 1.2% in terms of accuracy on the CIFAR-10 dataset and in 1.3% in terms of recall at k (k = 1) on the Stanford Online Product dataset.