Bidirectional-generative-network-based Research On Generalized Zero-shot Learning And Its Robustness

With the rapid development of deep learning in recent years,deep neural networks have achieved better performance than human in supervised image classification tasks.However,the supervise learned models need a large amount of labeled data in the training stage,and it is almost unrealistic to collect enough and reliable data in the practical application scenarios.In order to solve the problem of lack of data in the real scenario,the concept of transfer learning is put forward.It attempts to transfer the knowledge learned from the known data domain and realize the purpose of learning from the unknown data domain.Among them,generalized zero-shot learning,as a task to simulate knowledge transfer in extreme situations,has received extensive attention and a large number of related research studies have emerged.Although the existing generalized zero-shot studies have achieved good results,common problems such as domain bias have not been well solved.Moreover,although generalized zero-shot learning has a setting closer to the actual scenario,there is a lack of discussion on the robustness of the model in the existing research works.However,in a more practical scenario,not only is data collection one of the difficulties,but the reliability of the collected data itself and the credibility of the intermediary used for knowledge transfer also have a significant impact on the results.Based on the above observations,in view of the current situation that domain bias has not been effectively alleviated in current research of generalized zero-shot learning and the lack of systematic discussion on the robustness of the model,this paper will carry out the following researches with the generative generalized zero-shot learning as the theme:(1)Considering that the existing generative generalized zero-shot learning model still cannot deal with the domain bias problem well,this paper proposes the idea of bidirectional generation and designs a bidirectional generation adversarial network to further alleviate the domain bias problem.Using semantic representation,the bidirectional generation adversarial network maps the constructed visual samples back to the semantic space.Thus the semantic consistency of generated samples of the same class can be guaranteed while constructing samples with class discriminant.(2)In view of the lack of systematic discussion on the robustness of generalized zero-shot learning models,a robust bidirectional generative framework based on adversarial attack is proposed in this paper,and the robustness of generative generalized zero-shot learning models is discussed.In order to verify the effectiveness of the proposed framework,this paper selected representative VAE and GAN-based generative generalized zero-shot learning models to conduct robustness experiments and detailed analysis.(3)Because there is a lack of systematic robust evaluation scheme for generalized zero-shot learning at present,this paper proposes a complete robust evaluation scheme for generalized zero-shot learning.From the construction of possible disturbance situations,the robustness test process,and finally the robustness evaluation criterion,the robustness evaluation scheme proposed in this paper provides a foundation for further robustness research.