Research On Deep Generative Models Based On Variational Inference Of Flow Structure

In the field of machine learning,there are two main types of tasks: supervised learning and unsupervised learning.In recent years,using unsupervised learning to infer the natural structure of data and exploring data generation with efficient presentation capability is one of the focuses of current machine learning research.Fitting these models by maximizing the marginal likelihood or evidence is typically intractable,thus a common approximation is to maximize the evidence lower bound(ELBO)instead.Variational inference provides a unified framework for generative modeling.An important problem in variational inference is how to select a proper posterior distribution in a scalable way to approximate the real complex posterior distribution.However,when the data dimension increases,the longer serial transformation structure will bring larger training variance and even make the model collapse.On the basis of previous work,the dynamic routing mechanism is embedded in the normalizing flow model,after which the dynamic routing flow variational inference algorithm is proposed.This method is introducing the coupling coefficient to ensure the diversity of function transformation,inference effect compared with the standard of normalized flow structure to achieve the further ascension,to some extent reduced the variational item KL divergence disappear since the encoder and the collapse of the a posteriori probability problem.At the same time,the high-dimensional random variables are divided into sub-modules,and the weights among sub-modules are shared.The expression ability of each sub-module is combined through protocol routing,and a set of weighted mixed function transformation is realized.Moreover,this method is easy to realize parallel processing with the help of experimental platforms such as TensorFlow,so as to improve the inference efficiency.The experimental results show that the dynamic routing flow variational inference model proposed in this paper has significantly improved the inference effect and the posterior distribution estimation.Dynamic routing flow variational inference algorithm infers the probability structure of real data with a given finite number of samples,calculates and approximates the posterior distribution of hidden variables.By introducing a flexible coupling coefficient updating mechanism,the mean absolute deviation of the model is effectively constrained,and the inferential learning ability of the model is enhanced by means of clustering information learning potential space representation.