Network learning path of university political education based on simulation data and sparse neural network

In the face of the impact of the New Coronary Pneumonia epidemic, Schools need to actively engage in online teaching in response to the Ministry of Education’s call for “uninterrupted teaching”. Ideological and political education is an important way to train socialist successors and is the basis for establishing students’ correct outlook on life, values and the world outlook. Therefore, in this paper, sparse neural network algorithm is introduced to complete the construction of ideological and political online education platform for colleges and universities. Through the design of simulation experiments, we can know that the sparse model can still maintain the stability and accuracy of the network under the condition of black box attacks, and even after a certain amount of tailoring, it can still exceed the accuracy of the original network. The experimental results show the superiority of this platform. In this paper, the platform system is roughly divided into three layers: user layer, data storage layer and functional logic layer. The evaluation is carried out from four dimensions: teaching resources, teaching activities, teacher–student interaction and teacher–student evaluation. According to the results, students have higher systematic evaluation than teachers. Among them, the recognition of teaching resources and activities is high, which proves that the platform in this paper is effective. The online teaching platform designed in this paper can make full use of the characteristics of network technology, realizes the reform and innovation development path of the ideological course in schools, and enhances the attraction of the political course and the enthusiasm of students to learn. This paper designs a kind of network education system by introducing sparse neural network into ideological online education.


Introduction
In modern society, with the rapid development of science and technology, government departments of various countries have fully understood the importance of information technology and applied it in the process of social development and educational transformation (Mardis et al. 2018). On June 23, 2019, The State Council issued the Opinions on Comprehensively Deepening Education and Comprehensively Improving the Quality of Compulsory Education through Education Reform. Based on the background of the times and the integrated development of ''education ? Internet'', it put forward a deployment, and according to the requirements of school management, teacher teaching and student learning, it established an online teaching resource system covering all levels of compulsory education, and then explored an effective Internet teaching path. The Ministry of Education has formed relevant opinions on the construction and use of online learning resources in accordance with the important instructions of the CPC Central Committee, the State Council and the central leadership. The opinions mainly respond to the following requirements: first, the requirements for enriching high-quality resources; the second is to ensure the requirements of network platform operation; third, the requirements for the integration and utilization of network resources and education and teaching. The above requirements are also the successful scientific planning made by the state for China's education tasks based on the needs of modern education development and the active use of modern information technology. Since the outbreak of the epidemic in 2020, online teaching has gradually become the ''new normal'' of teaching. Network teaching not only promotes the development of normal teaching and obtains results, but it also brings teaching challenges, such as how to effectively combine network resources to promote teaching transformation (Pokhrel and Chhetri 2021). The Ministry of Education of China aims to give full play to the role of platform resources, elaborating on the ways in which teachers use resources to improve the teaching structure, students use resources to achieve autonomous learning, teachers and students use resources to respond to major public events, and serve the society (Mahmood 2021). It also puts forward requirements for improving teachers' own technology application ability, and needs to increase the intensity of teacher education and training. It can be seen from the above suggestions that the integration of online resources and online learning will be the ''new normal'' of future teaching forms, and the integration will be a new challenge for teachers. Based on this background, this paper introduces sparse neural network algorithm to realize the design of ideological and political network teaching platform for higher education, so as to expand a kind of ideological and political education transformation path, helps ideological and political teachers improve their ability to use resources, thus promoting the quality of ideological education, and promoting its integration with information technology. The online teaching platform of ideological and political education can effectively promote the construction of courses, and cultivate and provide highquality and comprehensive development talents for the country.

Related work
A small network prediction model based on self-encoder and multiple classifiers (Sparse Feature Extraction (SFE ?) is proposed in the literature (Liu et al. 2020). First, build a small part of the mining network for one-dimensional data, introduce some samples randomly into the input layer of the network, so that the network has less dependence on some features, alleviate the over fitting phenomenon, improve the system generalization ability, and enhance the model performance (Zhu et al. 2021;Smaragdis et al. 2008). Secondly, life cycle sparsity is introduced into the hidden layer to solve the problem of nonlinear data merging to a certain extent, and a small part that can be used for prediction is extracted from many random parts (Cao et al. 2014). A one-dimensional convolution and short-term memory network prediction model based on dimension reduction principle analysis is proposed in the literature. Compared with common one-dimensional features, images contain higher dimensional features, which require more complex networks to obtain effective features for prediction (Reza et al. 2022). The PCA-1DCNN-LSTM image prediction model proposed in the literature combines the PCA feature selection module based on 1DCNN-LSTM, which can effectively filter feature subsets, thus avoiding redundant information in the prediction model (Zhang et al. 2021). The paper intuitively expands the weight distribution of the robust network and proves the robustness of the model. There is a tradeoff between model accuracy and sparsity, and robust model parameters are very sensitive to sparsity (Chekuri et al. 2007;Moreira et al. 2009). According to this sparse feature, this paper proposes a hierarchical pruning algorithm based on reducing sensitivity to compress the robust network. In the literature, a one-step random confrontation training based on structural sparsity is proposed (Zhou et al. 2020). By combining the pattern structural sparsity with one-step random confrontation attack, one-step confrontation training is used to improve the stability of the model (Lin et al. 2020). The effectiveness of the proposed method is verified by the experimental results of CIFAR-10 and Mnist public datasets, and compared with other fast confrontation training methods, this method also has certain accuracy advantages (Zhang 2018).
3 Research on high performance computing based on sparse neural network

Sparse feature extraction network
Sparse self-coding network hidden layer: driven by big data, the features obtained by neural network have certain redundancy, and there is correlation between environmental factors in landslide data and grid cells. The sparse feature extraction network has two hidden layers and the life cycle of sparse feature extraction is shortened. At the same time, ReLU (Rectified Linear Unit) nonlinear activation function is used in the hidden layer, which increases the nonlinear expression ability of the network model. ReLU function can be expressed as formula (1). Next, find the k% maximum activation of Ri and set the remaining activation to zero.
Among g represents an approximate set of support vectors, (g) C represents its complement. In order to estimate the activation statistics of specific hidden units in all samples, this paper uses mini batch for statistics. At the same time, a threshold is set to achieve k% sparsity. k is defined as: where, u represents the degree of sparsity, batch_ size represents the mini batch size. There are only a few nonzero elements in each row of the characteristic matrix. Output layer: the loss function of sparse feature extraction network can be expressed as formula (4).
where W is the weight, b is the deviation, m is the input number, k is the weight falloff parameter. The global loss function of SFE ? model is mean square error (MSE), which can be expressed as: where yˆ, y and n represent the output data, the original input data and the number of samples, respectively. SVM model: Decision function for classification: In order to simplify the form of the formula, this paper first defines the target function l(xt) of layer t: It can then be expressed as: At the same time, for the last layer, qh L / qx L = 0, because hL is a predetermined minimum. Therefore, the partial derivative of the last layer is: For t = 1,…, L -1, we can get the partial derivative of the objective function with respect to xt as follows: Since hL is a predetermined minimum value, it usually satisfies hL \ \ h2 (t = 1…. L-1), so Eq. 12 is close to: When t = L:

Weight matrix generation and application
The sparse weight matrix obtained from the sparse variational self-encoder can be regarded as a special sparse connected network, but because it is more adaptive than the usual meta learning methods, it is easier to fall into local optimization and difficult to preserve common features. This will weaken the overall performance of the classification model. Therefore, this paper considers combining the features of the two meta learning methods, using the general meta learning method to save the common features from the initial values, and using the hidden space meta learning method to extract the local structure of the system. Based on the above LASSo weighted regression, the sparse vector zn is obtained through the variational auto encoder coding, and then the sparse weight matrix P is obtained, as shown in Formula (15): Among them, c represents the level of judgment of sparse weight, is an adjustable super re-recession, which is used to control the impact of sparse weight on the sparseness of the connected network, and indirectly control the sparseness of connecting networks.
In order to make the model output sparse, a special gradient descent method must be used to ensure that the sparse dimension converges to 0. Otherwise, the traditional gradient descent method will oscillate around 0, and the final result cannot be discrete. This paper uses the near end gradient descent to solve this problem, and ensures the maximum output by setting the zero crossing parameter in the next update to 0. When using near end gradient descent to update in SAMCN, you must first change the regularization sparsity used in SAMCN to a regular term in general form. Without losing generality, let matrix B be the combination of two regular terms in sAMCN, as shown in Eq. (17): Then the loss function in Formula (16) can be simplified as (18): Then the common terms of this general form can be optimized by using the near end gradient descent. The matrix Un is the numerical matrix of the parameter, as shown in Eq. (19): The definition of prox function is shown in Eq. (20):

Single step training results based on sparse structure
The basic idea based on gradient attack is to find the optimal disturbance direction along the gradient increasing direction of the model loss function, as shown in Eq. (21).
x adv ¼ arg max On the contrary, the starting point of optimization based attack is to find the optimal countermeasure disturbance on the premise of meeting the misclassification of countermeasure samples generated in the model: Find the optimal anti disturbance along the gradient rising direction of the model loss function: PGD (Projected Gradient Descent): L-BFGS attack: L-BFGS is the first white box attack method, which is an optimization attack algorithm based on boundary constraints.
Black box accuracy results of models under different pruning rates are shown in Table 1. The accuracy of PGD attacks is higher than that of FGSM attacks, regardless of Mnist or Cifar-10 datasets. This is because PGD is a multistage iterative attack, which will affect the replacement model when it generates disturbance and over fitting. It can be found in the Mnist dataset that with the improvement of pruning rate, the accuracy of the black box algorithm in this paper is almost the same as that of the original network model. In Cifar-10 dataset, the accuracy of sparse model under black box attack is higher than that of the original network. Especially when the pruning rate is 30%, the accuracy of FGSM and PGD of pruning model is improved by 1.04 and 0.89%. This is because it is difficult to estimate model parameters of sparse network model compared with ordinary model. It can be seen from the experimental results that the sparse model after hierarchical pruning can still maintain the stability and accuracy of the original network under the black box attack, and even exceed the accuracy of the original network under a certain pruning rate. Table 2 shows the stability of the training model of this method and the number of attacks in the iteration. As the number of disturbances increases, the accuracy of this model decreases, but it will eventually become stable. And the low deviation of the design method in this paper is consistent with the multi-phase PGD confrontation training. It shows that the stability of this method is not bad with the increase of attack times.
The specific experimental data, such as test accuracy and loss curve, can be obtained from Fig. 1. The data analysis shows that the training loss decreases with the experimental process, and the accuracy increases continuously. At the 5000th iteration, these two values tend to be stable, indicating that the learning of the model has reached stability.  The basic design structure of the network teaching platform for ideological and political education in colleges and universities is shown in Fig. 2. The main operating characteristic of the system is to take knowledge sharing as the goal, so as to allocate teaching resources reasonably. User layer: the users of online teaching platform for ideological and political education in colleges and universities are usually teachers engaged in information technology teaching, and can provide learning services for them.
Functional logic layer: the functional logic layer provides functional services for students. The platform mainly includes registration and certification, knowledge sharing, resource upload and download, self-study, background management and other functional modules. Among them, registration and authentication are generally used to confirm the user's identity; Knowledge sharing is usually a module of the platform, providing students with learning resources; Uploading and downloading resources are modules that search for ideological and political learning resources that can meet the needs of students according to keywords, and provide complete or partial downloading functions; Self-study module mainly allows students to study according to their own time and part; Background management mainly aims at platform registration and review of uploaded resources, new user registration and platform management.
Physical storage layer: the physical storage layer is usually divided into user information base and knowledge base data. The ideological and political knowledge sharing platform for AI courses uses MySQL database management system to store user information and knowledge resources. The user information database data includes all user information on the registration platform; Knowledge base data includes knowledge sharing data uploaded by users, learner behavior data, downloaded learning resource data, etc.

Main functional modules of the system
The front end of online teaching platform is mainly used for teaching case search, resource display and resource retrieval. The user functional structure includes four modules: registration and authentication, self-learning, knowledge sharing, and background management, as shown in Fig. 3.
Register and log in. The user who logs in for the first time needs to register first, and the login can be successful only after the background manager passes the verification. Users must provide their mobile phone number and password when logging in.Self-taught. After the user logs in to the system successfully, he/she can select and upload the model course to the user's course list according to the five aspects of the course's ideological and political objectives, content, activities, resources and learning evaluation, and upload the model course materials such as videos, PPTs, Word and archived data. At this time, if the user agrees to upload the knowledge base, it will be sent to the background for review. If it passes the review, it will enter the shared lesson knowledge base.  Knowledge sharing. Users who have successfully registered and logged in can only search for the course name by keyword in the shared knowledge base. It supports downloading the entire course file or downloading a single file separately. In addition, users can manage individual courses in My Course module. The system will allow users to re edit the courses that have not been uploaded to the knowledge base, and suggest to modify the contents of the course file during the background review.
This research is based on the observation of school morality and law, and interviews teachers timely after classroom observation. Next, according to the content of classroom observation and interview, the teaching at different levels is analyzed, as shown in Table 3:

Overall analysis of network teaching status
The survey divides the evaluation results of teachers and students into four grades: ''very satisfied'', ''satisfied'', ''average'' and ''dissatisfied'' from the aspects of teaching resources, teaching activities, teacher-student interaction, and teacher-student evaluation. The teaching effect of teachers is shown in Table 4: Among them, the recognition of teaching resources and teaching activities is relatively high. In terms of interaction between teachers and students, about 40% of teachers and students think that the effect of traditional offline teaching is not as good as that of traditional offline teaching. Nearly half of teachers think that students' performance in online learning of ideological and political courses is ordinary, and the learning effect needs to be strengthened. To sum up, during the epidemic period, the influence of online teaching was widespread, especially the interaction between teachers and students was less. Students are unable to communicate effectively with teachers and classmates when learning. When students have doubts and blind spots in learning, it is difficult to effectively solve them, which affects the learning effect of students. In addition, teachers' evaluation of students' online learning is low. It can be seen from the interview that some teachers think that some students are not focused enough and easily distracted. When learning, students are easily affected by the surrounding environment. Further improve the effect of classroom teaching, but some functions of the system lack maintenance Student evaluation The system can comment on the works of the middle school students in the resource library Network learning path of university political education based on simulation data and sparse neural… 9961

Analysis of network model training test results
As shown in Fig. 4, the classical pulsation matrices of different convolution kernels are compared with the sparse convolution calculation kernel proposed in this paper. The input characteristic diagram of 224*224*3 is calculated from the number of layers composed of 3*3*3 convolution kernels to obtain the normal time required for the change of normalization trend. It can be seen that when the sparsity is 0, that is, when the neural network is a dense neural network, the classical systolic matrix can achieve better performance. The execution time of the sparse convolution kernel is 2.1 times that of the shrink matrix. This result is because it takes more time to evaluate and skip MAC calculations with a value of 0 if you want to skip MAC calculations with a value of 0. As the sparsity of the network mode increases, the execution time of the sparse convolution core decreases gradually due to the increase in the number of omitted MAC operations. The accelerator designed in this paper can consider the sparsity of weight and mapping at the same time, which can accelerate the effects of static synaptic sparsity, static neuronal sparsity, and dynamic neuronal sparsity. In order to further analyze the impact of feature map sparsity on the performance acceleration of sparse convolution cores, some feature maps in the input feature map are randomly set to 0, and the convolution cores are set to dense cores. The calculated core time is shown in Fig. 5.
Two conclusions can be drawn from Fig. 5. First, we can see that the sparse convolution calculation core designed in this paper can also be fully used to accelerate the sparsity of feature maps. Therefore, in the effective application process, the fractional accelerator architecture proposed in front of the sparse convolution kernel in the neural network model can achieve better performance. Secondly, the accelerator proposed in this paper has a good acceleration effect on the sparsity of feature map, which is related to the accelerator design method used in this paper. The data stream described above uses a sparse data stream with a fixed weight. For the characteristic graph, the flow ping-pong method is adopted, and the 0 value of the characteristic graph is not in the calculation unit. In the experiment in this section, the convolution kernel is set to dense convolution kernel, which results in the same number of computing tasks for PE matrix, so as to achieve load balancing, resulting in better time effect of the acceleration effect of the characteristic graph of the computing core.
5 Research on strategies for improving the teaching of political courses under the network environment

Demand for political education under the network environment
In the traditional teaching process, teachers are in a completely dominated ''centralized'' educational framework. The educated gain the right to speak through the media network and change from the traditional passive receiver to the active attacker. Due to the influence of information technology, higher requirements have been put forward for traditional ideological and political education. First, teachers need to fully understand the importance of information technology and make full use of the teaching network carrier in classroom teaching, which is the necessary ''information literacy'' for ideological and political teachers. However, at present, many teachers have formed an unconscious ''teaching inertia'' and are used to relying on the ''comfort zone'' familiar with traditional classroom teaching. Therefore, many teachers have fear and resistance to the use of online teaching network, and the carrier of such ''psychological field'' has been formed for many years. In addition, many ''digital immigrant'' teachers have poor understanding and skills of network equipment, and there is a gap in the use and operation of network equipment technology compared with ''digital natives'' college students. The second is the new requirements of teaching service informatization. Although the online teaching activities carried out during the epidemic were ''forced'' by modern ''emergency'' teaching technology, they still showed the defects of ''normalization'' education technology support. During the epidemic, the comprehensive online teaching mode put forward a stricter test on the software and hardware capabilities of the technology. How to introduce and support all-round technology in teaching is the basis for ensuring that students have classes, improving teaching quality and achieving success.
The third is the new requirements of informatization on teaching organization. Through face-to-face knowledge explanation of offline courses, teachers can timely adjust the teaching difficulties according to the students' homework completion and learning ability fed back from the scene, so as to ensure that students track the teaching progress and understand the teaching content.

Strategies for improving the teaching effect
The sudden new epidemic situation has left school teachers with only one or two weeks to prepare for online teaching. There is no doubt that this has increased the difficulty and intensity of teaching. Many teachers do not adapt to this teaching method as soon as possible. In the process of online teaching, the teaching progress will be disrupted due to ''rush'', thus increasing the pressure on teachers and further affecting the effect of online teaching. Therefore, schools should actively organize teachers to carry out information-based teaching training, and systematically train teachers in teaching methods, teaching contents and teaching organization forms. Teachers can also share some high-quality online courses with the working group, organize teacher exchanges and seminars, and constantly improve the teaching quality. The school should start from the reality and orderly carry out online teaching skills training for teachers. First of all, when introducing information technology teacher training courses, schools should carry out training courses according to their own conditions, not only to infuse teachers with a lot of resources, but also to focus on training teachers how to use the network independently. This requires teachers to have a certain understanding of their own situation, including professional level and skills. This type of training needs to solve the actual problems of teachers and carry out individualized and hierarchical training courses, such as the training of college ideological and political teachers, which can be classified from the aspects of history, current events, humanities, tools, etc. The training course allows teachers to learn the teaching software related to the ideological and political course, which is convenient for teachers to organize theoretical knowledge and cultivate students' logical thinking ability. In general, an excellent teaching course can make ideological and political related knowledge practical and give full play to its due effect.
The survey showed that most teachers were not good at online teaching before the epidemic, but during the epidemic, almost all college teachers in China were practicing online teaching. Therefore, recording lessons or live online teaching has become a basic teaching skill for teachers. Using various network platforms and teaching software for teaching has greatly changed teachers' traditional teaching concepts. Teachers use multimedia, books and other classroom teaching tools to impart theoretical knowledge to students in a one-way way, while ignoring the fact that students are also teaching subjects. The principle of ''student-centered'' has not been implemented. The traditional teaching concept of teachers is deeply rooted in teaching practice, and teachers often do not want to change it. First of all, teachers should change their roles, clarify their responsibilities, guide students to learn and disseminate knowledge, and help students establish a correct three outlook. Teachers themselves should also establish the concept of lifelong learning, constantly develop professional knowledge, develop teaching skills, and adhere to the principle that teachers receive education first. Finally, teachers should actively innovate, actively explore problems, explore students, and take students as the main body of teaching. Teachers need to change from teachers to learning designers. The biggest problem of online teaching is that the teacher is far away from the students, and is not in the same space as the students. How the students learn, what behaviors they have, whether they take the initiative to learn, self-discipline, etc. The teacher on the other side of the screen is difficult to control. Teachers need to think about how to teach from the perspective of students' learning, insight into students' needs, design teaching from the perspective of students' learning, and pay attention to, support and promote students' learning. Starting from students' problems and students' learning design ideas, pay attention to students' emotional interaction.
The network environment itself has its own advantages and disadvantages for students' learning and teachers' teaching. In order to effectively use the network environment for assisted learning, students need to improve the selection, thinking and creation of information and resources under the network environment on the basis of abiding by laws and norms.
First, we need to strengthen online legal education and moral education for students. As a virtual space, the network environment is an important extension of the real society. The social order and activities of people in the network space must also be regulated by law. Therefore, students' words and deeds should still respect laws and regulations and conform to the code of conduct. Only when students know the law, distinguish right from wrong, and base themselves on law and morality can they consciously control their words and deeds, access the Internet legally and civilly, make effective use of the Internet, and give play to their subjective initiative in the learning process. Secondly, in the face of numerous resources, students should pay attention to educational resources other than entertainment and social networking, improve their ability to monitor and screen resources, and use critical eyes to ''eliminate the rough and extract the essence'' and ''eliminate the false and seek the truth''. Select learning contents and learning projects that meet their own characteristics and learning needs. In addition, the network environment provides an excellent platform for testing the real effect of ideological and political teaching. In the face of remarks that are contrary to the mainstream ideology of Chinese society under the network environment, students should maintain political vigilance and political discrimination, speak with the ideological and political knowledge they have learned, adhere to justice, and avoid blindly following and blindly criticizing. They need to follow the trend, use Marxist positions, views and methods to analyze and solve problems, and improve their understanding, identification and creation of the political thoughts and values to be expressed in the ideological and political courses.

Conclusion
In the context of the new crown epidemic, schools across the country are actively promoting online teaching. The first is the need for students to learn new courses, the second is the need for epidemic prevention and control, and the third is the need for social stability. Especially for students' families, if online teaching is not implemented, parents and students will be very worried. Whether this method is effective or not, online teaching is, after all, a form of campus learning. The biggest difference between online teaching and political teaching in colleges and universities lies in that, in order to carry out ideological education for students, they should not only be professional knowledge disseminators, but also be thought leaders. In this context, this paper completed the design and improvement of the online platform of ideological and political education for colleges and universities by introducing sparse neural network algorithm. Using this system for teaching can stimulate the enthusiasm of students, enhance the convenience of learning and improve the learning effect.
Funding The authors have not disclosed any funding.
Data availability Data will be made available on request.

Declarations
Conflict of interest The authors declare that they have no conflict of interests.
Ethical approval This article does not contain any studies with human participants performed by any of the authors.