Research On Semantic And Non-semantic Feature Mining Method For Insufficient Samples

With the development of deep learning,the structure of neural network model is becoming more and more complex,which often needs a lot of training data to help model learning.In practice,data collection in complex task scenes is difficult and costly,which leads to the lack of high-quality large-scale labeled datasets.Learning from the idea of transfer learning,the pre-trained model provides a higher starting point for task learning,and introduces external similar datasets to make up for the lack of samples' semantic feature space.The appearance of adversarial samples exposes the vulnerability of deep learning model,and then puts forward adversarial training to improve the robustness of the model,at the same time,it also inhibits the learning of non-semantic features of the model and loses generalization performance.In this paper,the semantic features and non-semantic features are studied as follows:(1)Research on semantic feature mining method based on middle-level attributes.For the task of multi-modal human activity recognition,this research designs a middlelevel attribute structure and builds a two-stage modeling framework,from low-level signal to middle-level attribute to high-level activity,which disentangles complex tasks into periodic tasks and reduces the difficulty of model mapping.In the first stage,the pretrained model of external middle-level attribute labeling data is introduced to model different modal signals at the bottom,which reduces the dependence on activity labeling data and alleviating the problem of insufficient samples;In the second stage,the improved LSTM network based on recurrent neural network is used to extract the dimensional features of time series.With experiments on Stanford-ECM datasets,the test results proved the effectiveness of this method,and further analyze the classification contribution of different modal signals according to the different intensity of activities.(2)Research on non-semantic feature mining method based on adversarial samples.For the task of image recognition,this research experiments on the generalization characteristics of non-semantic features,and verified that the non-semantic features and semantic features are different and the non-semantic features have classification effectiveness.It is proposed to mine the non-semantic features in samples by adversarial attacks,and take the adversarial samples as the carrier of non-semantic features,augmente the data of insufficient samples,improve the feature space of training samples,and then improve the generalization performance of the model.The experiments are carried out on CIFAR-10 dataset,and quantitative and qualitative analysis is carried out from the perspectives of test results and feature distribution,which verified the effectiveness of this method in improving the generalization of the model.