DeepJDOT: Deep Joint Distribution Optimal Transport for Unsupervised Domain Adaptation

Abstract

In computer vision, one is often confronted with problems of domain shifts, which occur when one applies a classifier trained on a source dataset to target data sharing similar characteristics (e.g. same classes), but also different latent data structures (e.g. different acquisition conditions). In such a situation, the model will perform poorly on the new data, since the classifier is specialized to recognize visual cues specific to the source domain. In this work we explore a solution, named DeepJDOT, to tackle this problem: through a measure of discrepancy on joint deep representations/labels based on optimal transport, we not only learn new data representations aligned between the source and target domain, but also simultaneously preserve the discriminative information used by the classifier. We applied DeepJDOT to a series of visual recognition tasks, where it compares favorably against state-of-the-art deep domain adaptation methods.

Unsupervised Domain Adaptation With Optimal Transport in Multi-Site Segmentation of Multiple Sclerosis Lesions From MRI Data

“…We aim at leveraging the available information {X s , Y s , X t } to learn a classifier f , that is a labeling functionf which approximates f s and is closer to f t than any other functionf s . In order to solve this unsupervised domain adaptation problem, the Optimal Transport theory can be employed Damodaran et al, 2018).…”

“…Subsequently, Damodaran et al proposed a deep learning strategy to solve these two drawbacks (Damodaran et al, 2018). Their Deep-JDOT framework (i) minimizes the cost c in a deep layer of a Convolutional Neural Network, which is more informative than the original image space, and (ii) solves the problem with a stochastic approximation via mini-batches from the source and target domains.…”

“…In order to define the probability distributions, we employ image patch samples rather than image samples as in Damodaran et al (2018). The use of image patches enables an higher number of samples and, therefore, a more precise estimation of the true distribution µ.…”

“…Generative adversarial approaches may generate image samples that are highly different from the actual MRI MS images and therefore make the network learn useless representations. One of the most recent works in Unsupervised Domain Adaptation proposes a solution for a classification task based on Optimal Transport, which learns a shared embedding for the source and target domains while preserving the discriminative information used by the classifier (Damodaran et al, 2018). Our framework is based on the latter approach.…”

Learning a Domain-Invariant Embedding for Unsupervised Domain Adaptation Using Class-Conditioned Distribution Alignment

“…MMD measures the distances between the distance between distributions simply as the distance between the mean of embedding features. In contrast, more recent techniques that use a shared deep encoder, employ the Wasserstein metric [42] to address UDA [4], [6]. Wasserstein metric is shown to be a more accurate probability metric and can be minimized effectively by deep learning first-order optimization techniques.…”