A Statistical Learning Model With Deep Learning Characteristics

With the continuous growth of computer computing power and Internet data volume,deep learning has achieved great success in computer vision,natural language processing,speech recognition,and other fields,but the lack of interpretability severely limits its wide use in real-world tasks,especially in security-sensitive tasks.Statistical learning models have better interpretability,less calculation,and higher robustness than deep learning.However,statistical models have lower performance on sparse data such as image recognition tasks than deep learning models,resulting in fewer and fewer people using statistical learning models.When the performance of deep learning models hits the bottleneck,artificial intelligence security issues become more and more important,and there are more and more relevant studies.Many of them use the interpretability of statistical learning models to explain deep learning models.However,these studies have not completely solved the problem of deep learning interpretability,and deep learning interpretability is still an open problem.Instead of improving the interpretability of deep learning models,why not do the opposite and use the features of deep learning to improve the performance of statistical learning? This thesis uses the features of deep learning to improve the performance of statistical learning models,and designing a statistical learning model with deep learning characteristics.The main contributions and innovations of this thesis are as follows:1.This thesis selects four important characteristics in convolutional neural networks for research: 1)translation invariance,2)sparse learning ability,3)feature capture,and 4)background filtering.Then use statistical learning algorithms to simulate these characteristics: use convolution algorithm to achieve translation invariance,use k-NN algorithm to achieve sparse learning ability,use a clustering algorithm to achieve feature capture,and use mathematical statistics to achieve background filtering,and thus design a statistical learning model with deep learning characteristics.2.In the realization of the model,through experiments we verify that convolution calculation can improve the accuracy of the similarity algorithm;Define similar parts in the same sample as the features of the image,and solve these features through the clustering algorithm;calculate the variance of the entire sample and define the location of the variance below the threshold as a static background,calculate the variance of the same sample and define the location of the variance above the threshold as a random background.3.This thesis do comparative experiments on the three models: our model,CNN and k-NN under two benchmark data sets MNIST and CIFAR-10.The experiments show that our model performance is better than ordinary statistical learning models.The calculation amount and robustness are better than the deep learning model.Then we analyze the interpretability of the model and visually display the calculation results of each step of the model in the form of graphs.Finally,we study the influence of hyperparameters on the performance of the model,and propose the improvement space of the model.