Semi-Supervised Pairing via Basis-Sharing Wasserstein Matching Auto-Encoder

Semi-supervised pairing, learning explicit pairing relation between unlabeled input and target distribution in an otherwise semisupervised setting, has recently attracted lots of attention [16, 12], since data labeling is often time-consuming and requires lots of human labor. In this paper, we propose Basis-Sharing Wasserstein Matching Auto-Encoder to tackle the problem of semi-supervised pairing. Our model inspired by the success of robust representation learning for matching cross-modal latent space distribution [15, 10] and good statistical properties of optimal transport [1, 4]. In particular, Wasserstein distance in the optimal transport family has been shown to work on GANs and autoencoders [2, 14]. We propose to match the latent code distribution of the unlabeled dataset, with the labeled data points as “anchor points”. This should help the network making association between the distributions from two different domains. A similar idea that uses a different method has been shown to work on some language tasks [3]; our work explores a new approach based on distribution matching. Through preliminary experiments, we show that the proposed algorithm can successfully incorporate the unlabeled data for improving the classification accuracy on MNIST and CIFAR10 datasets.