DATE: Wed, Feb 19, 2020
TIME: 1 pm
PLACE: SITE 5084
TITLE: Mixup and Its Applications in Transfer Learning
PRESENTER: Yuan Wu
Carleton University
ABSTRACT:

Large deep neural networks are powerful, but exhibit undesirable behaviors such as memorization and sensitivity to adversarial examples. Mixup is a simple learning principle to alleviate these issues. In essence, mixup trains a neural network on convex combinations of pairs of examples and their labels. By doing so, mixup regularizes the neural network to favor simple linear behavior in-between training examples. Transfer Learning is a hot topic in machine learning, it has many subfields, such as domain adaptation and multi-domain text classification. For unsupervised domain adaptation (UDA), it aims to learn a good predictive model in a target domain without any labeled data by leveraging the knowledge from a label-rich source domain. Recent advances on UDA rely on adversarial learning to disentangle the explanatory and transferable features for domain adaptation. However, there are two issues with the existing methods. First, the discriminability of the latent space cannot be fully guaranteed without considering the class-aware information in the target domain. Second, only discriminating the samples from the source and target domains is not sufficient for domain-invariant feature extracting in most parts of the latent space. In order to alleviate the above issues, we propose a novel dual mixup regularized learning method (DMRL) for unsupervised domain adaptation, which not only guides the classifier in enhancing consistent predictions in-between samples per domain, but also enriches the intrinsic structures of the latent space. The DMRL jointly conducts category and domain mixup on the pixel level to improve the effectiveness of the models. A series of empirical studies on four domain adaptation benchmarks demonstrate that our approach can achieve state-of-the-art performance. For multi-domain text classification, many existing approaches are highly dependent on the availability of sufficient large training data. However, some datasets have abundant labeled training samples, while others have scarce or no labeled data. Here, we propose a mixup regularized adversarial network (MRAN) to tackle multi-domain text classification (MDTC) task. This novel model adopts the shared-private paradigm and applies mixup to conduct interpolation regularization on both category-level and domain-level simultaneously, to help enforce feature alignment among different distributions. Experimental results on two benchmarks demonstrate that our method is superior to existing approaches in the literature.