An Investigation Into the Effectiveness of Latent Variable Models for Domain Adaptation

Abstract

The proliferation of machine learning with neural networks (NNs) has revolutionized fields such as computer vision and natural language processing. However, their successes often overshadow two important weaknesses of neural networks: (i) their reliance on large amounts of training data and (ii) the assumption of independent and identically distributed (i.i.d.) data. Because of these weaknesses, the vast majority of NNs today are applicationspecific machineries tuned to one task and one data domain.

This thesis investigates the effectiveness of a latent variable model for unsupervised domain adaptation, aiming to bridge the gap between two different data distributions while leveraging only labeled data samples from one, and unlabeled data samples from the other. A novel generative modeling framework is proposed to address this problem, incorporating recent advances in probabilistic modeling and variational inference techniques from the neural network literature.

Empirical results of the proposed approach seem promising, and indicates adequate transfer of the labeling knowledge of the model across disparate data domains without requiring manual re-labeling or domain-specific adjustments. Moreover, the proposed approach has also shown potentials in solving the related domain translation problem. Despite these fortunes, the existing approach has shown limitation in solving more complex scenarios of unsupervised domain adaptation, speficially those involving more vibrant differences between domains.