| Summary: | 博士 === 國立中央大學 === 資訊工程學系 === 106 === Probabilistic framework has emerged as a powerful technique for representation learning. This dissertation proposes probabilistic latent variable model-based representation learning methods that involve both discrete and continuous latent spaces. For a discrete latent space, a hierarchical representation that is based on the Gaussian hierarchical latent Dirichlet allocation (G-hLDA) is proposed for capturing the latent characteristics of low-level features. Representation is learned by constructing an infinitely deep and branching tree-structured mixture model, which effectively models the subtle differences among classes. For a continuous latent space, a novel complex-valued latent variable model, named the complex-valued Gaussian process latent variable model (CGPLVM), is developed for discovering a compressed complex-valued representation of complex-valued data. The key concept of CGPLVM is that complex-valued data is approximated by a low-dimensional complex-valued latent representation through a function that is drawn from a complex Gaussian process. Additionally, we attempt to preserve both global and local data structures while promoting discrimination. A new objective function that incorporates a locality-preserving and a discriminative term for complex-valued data is presented. Then, a deep collaborative learning framework that is based on a variational autoencoder (VAE) and a Gaussian process (GP) is proposed to represent multimedia data with greater discriminative power than previously achieved. A Gaussian process classifier is incorporated into the VAE to guide a VAE-based representation, which distinguishes variations of data among classes and achieves the dual goals of reconstruction and classification. The developed methods are evaluated using multimedia data. The experimental results demonstrate the superior performances of the proposed methods, especially for situations with only a small number of training data.
|
|---|
