Adversarial variational domain adaptation for semi-supervised image classification.

Abstract

For success fully training deep neural networks, we usually need a large amount of annotated data in order to avoid the overfitting and being able to generalize to new data. In most of real cases, getting labels is difficult and time consuming. In this work we address the problem of transferring knowledge obtained from a vast annotated source domain to a low labeled or unlabeled target domain, reducing the efforts to get labels on the target. We propose Adversarial Variational Domain Adaptation (AVDA), a semi-supervised domain adaptation method based on deep variational embedded representations. The idea of AVDA is to use a mixture of Gaussian distribution as a prior for the latent space, mapping samples that belong to the same class into the same Gaussian component, independently of the domain membership, using approximate inference. We use adversarial methods to align source and target distributions in latent space for each class independently. We tested our model using the digits dataset, which contains images of handwritten digits and images of number of houses. We empirically show that on a semi-supervised scenario, our approach improved the state of the art for digits dataset from 0.3 to 1.5% of accuracy using only 1 and 5 labels per class. Also, we tested out model using images of galaxies from the Cosmic Assembly Near-infrared Deep Extragalactic Legacy Survey (CANDELS, [23]) as source and the Cluster Lensing and Supernova Survey with Hubble (CLASH, [62]) as target. We empirically show that using few labels our model presents a significant speed-up in terms of the increase in accuracy, and the model keeps improving as more labels we add.