Transductive Adversarial Networks .

Transductive Adversarial Networks (TAN) is a novel domain-adaptation machinelearning framework that is designed for learning a conditional probabilitydistribution on unlabelled input data in a target domain, while also onlyhaving access to: (1) easily obtained labelled data from a related sourcedomain, which may have a different conditional probability distribution thanthe target domain, and (2) a marginalised prior distribution on the labels forthe target domain. TAN leverages a fully adversarial training procedure and aunique generator/encoder architecture which approximates the transductivecombination of the available source- and target-domain data. A benefit of TANis that it allows the distance between the source- and target-domainlabel-vector marginal probability distributions to be greater than 0 (i.e.different tasks across the source and target domains) whereas otherdomain-adaptation algorithms require this distance to equal 0 (i.e. a singletask across the source and target domains). TAN can, however, still handle thelatter case and is a more generalised approach to this case. Another benefit ofTAN is that due to being a fully adversarial algorithm, it has the potential toaccurately approximate highly complex distributions. Theoretical analysisdemonstrates the viability of the TAN framework.