Sparse-and-Non-Shared Transfer Learning Based On Global Semantic Reasoning

In recent years,deep neural networks have made significant progress in solving various machine learning problems and applications.However,this significant advancement is due to the availability of labeled data on a large scale.It is usually undesirable to manually enough training data with label for specific application tasks.In the absence of labeling data,it is urgent to design a general algorithm to reduce the consumption of manual labeling.Domain adaptive methods can use machine learning methods to train data sampled in one distribution and apply it to data sampled in another branch.Its core is to adapt to changes in data distribution in different domains.However,in actual application scenarios,on the one hand,it is often difficult to find the source domain with the same label space as the target domain.On the other hand,there is an overfitting problem caused by the scarcity of data in the corresponding space of the source domain.In response to the above mentioned,this paper establishes a non-shared and rare transfer learning research problem.It uses conditional adversarial training and external knowledge inference to classify image data.Deep adversarial networks can reduce the distribution differences of feature representations in different fields.Semantic knowledge can enhance the ability of automatic learning.Compared with the traditional domain adaptive method,it can train more efficiently and get a more accurate recognition effect.First of all,this paper takes the restaurant smart receipt scenario as an entry point,and based on the analysis of previous research work such as image classification and domain adaptation,a domain adaptive model(CTAN)based on conditional adversarial network and a priori tree is proposed.Different from the previous method of completely matching and aligning the feature distribution of the source and target domains,this method can jointly borrow the non-shared knowledge at the same level into the shared class of the source domain,and further use the conditional confrontation network to transfer shared knowledge from the source domain to the target domain.And the CTAN model is compared with the existing domain adaptive model.The experimental results of the two datasets verify the effectiveness of the model.Considering that the prior tree only uses shallower hierarchical relationships,in order to make full use of the rich semantic information in the category labels,a conditional adversarial domain adaptation model(GADA)under global semantic inference is further designed.The purpose is to use external knowledge to enhance the local feature expression and knowledge transfer between nodes,which is achieved through a two-stage algorithm.In the first stage,the transfer graph reasoning layer(TGR layer)is embedded behind the convolutional layer.This TGR layer performs semantic reasoning on graph nodes based on a priori knowledge to achieve knowledge sharing between adjacent nodes.In the second stage,it is achieved through the adaptation of adversarial domains to align the multimodal distribution,thereby achieving knowledge sharing between different domains.The proposed algorithm is evaluated on multiple datasets,and the performance is better than the latest algorithm.