Zero-shot Learning Via Res-Gan Network

Zero-shot learning is one of the hotspots in the field of transfer learning.Unlike traditional image classification,zero-shot learning aims at identifying data categories that have never been seen before,that is,the samples classified and identified during the testing stage are not involved in the training of classifier models.In the process of solving zero-shot learning problem,in order to transfer knowledge from seen class to unseen class,classification model needs to construct a mapping from bottom features to class labels through visual attributes and other auxiliary knowledge.Attribute is a feature that can be labeled manually and observed in an image.It is a high-level description of image content and can be understood by both machine and human.A large number of studies have shown the role of attribute learning in target recognition,image description and zero sample learning.The process of zero-shot learning can be seen as the mapping between image and visual attributes.In order to facilitate research,most zero-shot learning does not directly use pictures,but uses existing feature extraction models to extract image features.The process of zero-shot learning can be regarded as the mapping between feature space and attribute space.The effective method of zero-shot learning is to map the data in feature space and attribute space to an embedded space,and then process the data in the embedded space using KNN and other classification algorithms.At present,in order to complete the mapping from feature space and attribute space to embedded space,the most widely used research method is based on neural network.And most of the methods based on neural network are shallow full connection,which makes it difficult for network mapping to have high accuracy.In order to solve this problem,we propose a new neural network structure Res-Gan,which is a generative adversarial networks based on residual structure.The purpose of introducing a residual structure is to increase the depth of the network.Since the optimization process of the neural network uses a back-propagation algorithm based on gradient descent,the gradient of the front layer of the network is affected by the back layer in the process of back propagation,and the gradient gradually decreases,which is prone to the problem of gradient disappearance.Therefore,the depth of the neural network is difficult to increase,which makes it difficult to continue to improve the performance of the network.In order to increase the depth of the network,we first proposed to introduce the residual structure idea into the study of zero-shot learning.The residual structure greatly increases the depth of the network,which further enhances the performance of the network.In addition,we also combine the residual structure and the GAN network.The GAN network is a generative adversarial network,including a generative network,which responsible for the simulation data,and a discriminative network,which is responsible for judging whether the data is analog data or real data.We incorporate the residual structure into the generative adversarial network to form Res-Gan,which optimizes network performance.In addition,we have done some work on the activation function and attribute space of the network.Since the previous research used a single activation function,which makes the nonlinear mapping ability or operation speed of the network have a certain impact,we fully compared the advantages and disadvantages of some activation function,and designed a neural network with multiple activation functions.The nonlinear mapping capability of the network is enhanced,and the operation speed is reduced.In the attribute space,we weight the attributes and not treat all the attributes equally.Rather,the important attributes are enhanced,and the secondary attributes are suppressed to improve the effect of zero-shot learning.