Abstract
To achieve satisfactory generalization abilities, machine learning models usually require large amounts of labelled data. However, data labelling is very costly, even in the era of big data. Transductive transfer learning looks at taking advantage of the availability of labelled data in a source domain to tackle the lack of labelled data in a task domain, using learning mechanisms to reduce the gap between both domains, so that the trained models can perform robustly in the target domain.
In this thesis, we investigate transductive transfer learning using a deep generative approach in order to solve challenging computer vision tasks such as object recognition and semantic image segmentation. Three novel models are developed for these purposes. First, a Deep Adversarial Transition Learning framework is proposed to project data from the source and target domains into intermediate, transitional spaces through the employment of adjustable, cross-grafted generative network stacks, and an effective adversarial learning strategy between the transitions, involving variational autoencoders (VAE) and generative adversarial networks (GAN). Secondly, a Cross-Domain Latent Modulation method is proposed, using deep representations of domain data to procure cross-domain generations using VAEs with cross-domain modulation added into the reparameterization process. The mechanism can be used for both unsupervised domain adaptation as well as image style translation. Thirdly, we propose a LeakingGAN model, making a GAN's discriminator polluted by information leaking from the generator. Combined with the mean-teacher mechanism, this leads to a powerful semi-supervised semantic segmentation method with enhanced transferability. Empirical studies carried out using a number of benchmark datasets have validated the effectiveness of these proposed generative transfer learning models.