Scalable Non-linear Beta Process Factor Analysis

We propose a non-linear extension of the factor analysis with beta process priors. This non-linear Beta process factor analysis (nBPFA) model allows the data vector to be represented by processing an invertible non-linear transformation on standard factor decomposition. We develop a scalable variational inference framework for large scale datasets, which benefits from the idea of generalizing variational auto-encoder (VAE) by introducing another latent binary variable with sparse constraint. This framework also allows the inputs to be real valued, binary or count vector data. We empirically test our algorithm on images and document datasets, and demonstrates competitive results, especially showing better performance preventing overfitting.