Research On Multi-task Technology Based On Bayesian Deep Learning

Multi-task learning optimizes multiple learning tasks by exploiting the shared information of related tasks,thereby improving the model prediction ability of each task.As an important research direction of artificial intelligence,multi-task learning has been widely used in various fields of machine learning.With the rapid development of deep learning in recent years,multi-task learning techniques have achieved satisfactory performance in processing tasks such as vision,text,and speech.However,due to the complexity of the relationship between multiple tasks,there are problems such as lack of shared information and negative transfer in deep multi-task learning.Therefore,how to make a specific task effectively learn the knowledge information from other tasks is still a significant challenge faced by multi-task deep learning for a long time.Bayesian deep learning,which combines Bayesian learning principles with deep neural networks,has become a robust area of research.It has a more vital perception and reasoning ability than traditional deep learning by capturing the uncertainty factors in the model.Considering that there are many uncertainties in the interaction of multimodal information in deep multi-task learning,this thesis introduces the Bayesian deep learning method to study the multi-task learning method from the shared network model structure and task optimization balance strategy.The main work of this thesis is as follows:Aiming at the shared information exchange problem caused by the inherent model structure in multi-task learning networks,this thesis proposes a multi-scale multi-task learning network using the Bayesian gate mechanism.In this thesis,a multi-scale and multi-task knowledge interaction module is designed in the shared network decoding stage to realize the joint learning of different scale features and multiple tasks in the network decoding stage.Each specific task learns scale contextual information in a multi-task learning unit pool and simultaneously communicates information between tasks.In this thesis,a Bayesian knowledge gating mechanism is designed using a Bayesian deep learning algorithm.The gating unit gates the uncertain shared feature flow in different communication stages of the shared network.Experiments show that the proposed model can perform better in multiple vision tasks.Aiming to avoid the negative transfer between tasks and balance the task weights,this thesis proposes a Bayesian multi-task dynamic optimization strategy based on weight update.It solves the multi-task optimization strategy problem by introducing Bayesian deep learning.This method learns the task weights related to the validation loss using forward-looking information by considering the uncertainty in the future information and uses the Bayesian deep learning algorithm to distribute the validation loss task weights to explore each task in the future gradient.The Bayesian automatic weight method updates the weights of each task in the direction of a better gradient.Experiments on multiple computer vision task datasets and multi-task networks show that this method can more effectively improve the performance of multi-task learning.