In recent years, deep learning models has made huge attention to the researchers specially computer scientists and data scientists due to the advancement of the optimization algorithms, improvement of the computational power and large volume of data. Deep neural network allows flexible parameterization of data distributions. This research mainly deals with latent classes of deep generative models (DGM) such as variational auto encoder, normalizing flow and generative adversarial network. DGM’s are the combinations of the deep neural network and probabilistic generative models. The goal of generative modelling is to derive the distribution of the input data and to generate new samples from this distribution. Deep generative modelling is proven very much useful for unsupervised learning where data is not labelled and costly to acquire. In practice, learning unlabelled data using existing machine learning methods still challenging and active research area. Therefore, this research aims to work on model selection and inferences of deep generative latent variable models. The theoretical foundation of the background of DGM are discussed in this proposal. Finally, this study formulates research goals by addressing the potential problems in VAE model and model selection using metrics in unsupervised and semi-supervised learning. As an illustration, this proposed framework will be utilized in some real-life applications such as open source data and credit scoring inferences.

Zoom link: https://vuw.zoom.us/my/ecspostgrad