Generative Model With Deep Learning

In recent years,artificial intelligence has been a popular research topic.It is envisioned that artificial intelligence will be a promising technique that encourages the innovation in the future.For all the technologies within this field,deep learning plays an important role undoubtedly,which has made a huge achievement in many application scenarios,such as computer vision,reinforcement learning,natural language processing,etc.There are two dominant categories of models in deep learning,discriminative model and generative model.Compared with discriminative model,generative model has the advantages in modeling the data,which could be extensively applied to many tasks.Based on such motivation,a new generative model with deep learning named generative model with kernel density estimation,is proposed in this dissertation to generate images.Different from other popular generative models like GANs and VAEs,the proposed model captures the distribution of given data directly.To be specific,the model estimates the distribution of given data and generate samples using kernel density estimation,then learns to minimize the distances between two distributions to finally get the samples of same distribution as given data.Our experiments showed that the proposed method can efficiently generate samples from a given distribution,which can be not only one-dimensional but also high-dimensional.The dissertation also applied generative model to the field of computer vision,where images can be cross-domain transformed from one style(determined by a specific attribute)to another while other semantic information stay unchanged.Compared with traditional cross-domain transforming model,the model has the advantage of separating the process in training and generating,which could shorten the time of image generating.