Research And Application Of Restricted Boltzmann Machine In Background Modelling And Text Modelling

Background modelling and text modelling both are fundamental problems of com-puter vision and natural language processing respectively. These two research areas are both at the cutting edge of artificial intelligence.Background modelling aims to generate background without foreground from frames in a given video. Traditional background modelling approaches focus on local informa-tion around a pixel of a frame, while global information is ignored, which is ubiquitous in real and complex video. For example, the illumination change caused by closing lamp and the shadow caused by moving objects will influence the background. In this thesis, the proposed model based on restricted Boltzmann machine utilizes global in-formation and take two adjacent frames into consideration. An adaptive regularization term is designed to force the backgrounds generated by two adjacent frames to be con-sistent and stable, so that it guarantees that the background information will be modeled precisely. To the best of our knowledge, it is the first time to apply restricted Boltzmann machine to the task of background modelling.Text modelling is a research task of extracting information from text and it is a basic task in the natural language processing field. This thesis considers the famous word em-bedding toolbox word2vec. To date there is no rigorous theory proposed for word2vec. Inspired from restricted Boltzmann machine, this thesis proposes a matrix factorization based model——explicit matrix factorization (EMF) and proves that EMF is equivalent to the skip-gram negative sampling (SGNS) model in word2vec without any approxima-tion or relaxation. The proposed EMF model provides better interpretation than other explanations. In addition, an extended supervised EMF model is designed to combine word analogy queries, and it significantly boosts the accuracy in word analogy task.