Zero-Shot Learning Based On Reverse Prediction

Benefiting from the improvement of computing power and the advent of the era of big data,machine learning has achieved remarkable results in areas such as computer vi-sion and speech recognition.However,with the development of machine learning,peo-ple are no longer content to deal only with supervised issues with large amounts of data,and few samples or even no sample learning problems are on the agenda.Therefore,how to make full use of existing knowledge to help learning without sample has impor-tant research value and significance.This dissertation focus on the zero-shot learning in two settings:transductive and inductive.Zero-shot learning can be considered as a kind of transfer learning,and it usually contains two domains:source domain and target domain.The former contains a large number of labeled training samples,while the latter has no labelling information and it contains a different set of labels than the source domain.At present,zero-shot learning still can not get rid of the dependence on other modal information,such as the semantic space made up of manually labeled attributes.By making the seen source classes and the unseen target classes share the same semantic space,people can build connections between them and establish a bridge for knowledge transfer.In order to solve the zero-shot learning problem,most of the existing methods are based on the projection strategy.In the training stage,they use supervised information in the source domain to learn a projection function which can represent the samples and classes into a common space.In the testing stage,the test samples and classes are first projected into the specified space,and the closest target class to each sample is calculated by the nearest neighbor strategy.Finally,the label of each test sample is obtained.However,these methods usually have projection domain shift problem and hubness issue.In addition,because they use a two-stage strategy for predicting,there will be some information loss.Different from these existing methods,this dissertation proposes to solve the zero-shot learning problem from the perspective of reverse prediction.According to whether the unlabeled samples in target domain are available during the training stage,zero-shot learning can be roughly divided into two categories:transudctive zero-shot learning and inductive zero-shot learning.Under the setting of transductive zero-shot learning,the unlabeled samples in target domain are available.The RevTZSL model proposed in this dissertation helps the knowledge transfer between the seen source classes and the unseen target classes by considering the information in the source domain and the tar-get domain at the same time,and uses the reverse prediction principle to infer the visual characteristics from the class label corresponding to the sample.By this way,the dis-crimination of unseen target classes is accelerated,and projection domain shift problemand hubness issue in zero-shot learning are effectively avoided.With the setting of in-ductive zero-shot learning,unlabeled data in the target domain is not available.To solve this problem,the main idea of this dissertation is to learn a model(RevIZSL)similar to a simple autoencoder model by using the principle of reverse prediction in the source domain,so that it robust enough to have good scalability in the target domain.In addi-tion,it is worth mentioning that the two models proposed in this dissertation can predict the labels corresponding to the test samples in only one step during the testing stage.To verify the effectiveness of the proposed RevTZSL and RevIZSL models,we perform experiments on three standard zero-shot learning datasets and a benchmark fine-gained image classification dataset,and we report the classification accuracy of these two models.The experimental results show that the proposed models have obvi-ous advantages compared with the state-of-the-art methods on the four datasets,espe-cially in the absolute advantage of fine-grained image classification to verify the dis-criminative ability of the proposed models.In addition,by analyzing the parameter sensitivity and convergence speed of RevTZSL and RevIZSL,the effectiveness of the proposed alternative optimization algorithms is further verified.