Design of an Intelligent Educational Evaluation System Using Deep Learning

Nowadays, online education has been a more general demand in context of COVID-19 epidemic. The intelligent educational evaluation systems assisted by intelligent techniques are in urgent demand. To deal with this issue, this paper introduces the strong information processing ability of deep learning, and proposes the design of an intelligent educational evaluation system using deep learning. Inside the algorithm part, the low-complexity offset minimal sum (OMS) is selected as the front-end processor of deep neural network, so as to reduce following computational complexity in deep neural network. And the deep neural network is adopted as the major calculation backbone. In this paper, our OMS deep neural network parameters are 23 and 57 compared with other parameters, which can save about 59.64% of the network parameters, and the training time is 11270 s and 25000 s respectively, which saves the training time 54.92%. It can be also reflected from experiments that the proposal further improves the performance of unbalanced data classification in this problem scenario.


I. INTRODUCTION
With the data explosion of the past decade, deep learning has experienced unprecedented leaps and bounds, playing an important role in solving a range of challenges in artificial intelligence [1]. Deep Neural Network (DNN) based software, which has been successfully deployed in many real-world applications, has become a key driver for many new industrial applications [2]. Notable applications of deep learning include image classification, speech recognition, and autonomous driving [3]. Along with excellent classification results, DNNs may also exhibit incorrect behavior due to hidden defects, which can lead to serious accidents and losses [4], [5]. To ensure safety, like conventional software, testing techniques are often used to detect incorrect DNN behavior and improve DNN quality [6]. However, in automated testing of DNN-based systems, it is often not possible to directly define test predictions for the correct output given a given input [7]. To obtain the test prediction information, it is The associate editor coordinating the review of this manuscript and approving it for publication was Shuai Liu . often necessary to spend expensive workforce to label the test data, which significantly slows down the quality assurance process [8].
In classroom teaching, there is a need for classroom assessment [9]. In particular, the current development of terminals such as smart phones have significantly reduced the quality of many classroom teaching and interactions [10]. However, in the prior art, although there are enough solutions to analyze classroom assessment through classroom videos, there is no relevant novelty on how to use foreground and background pixel pairs and grayscale information to extract video foreground objects and further analyze classroom assessment [11]. In practical applications, deep neural networks often incorporate a variety of known structures [12]. It can be understood that a deep neural network can be understood as a network containing multiple hidden layer structures [13].
In the process of students' political education, there exists an ''emphasis on explicit education but not implicit education'', which seriously restricts the effectiveness of students' political education [14]. The purpose of political education indicates that the educated will be trained to become a cer-tain society or a certain class of people, so the cultivation of a rational and peaceful social mentality in the new era requires that political education should focus on the goal of ''building moral character'' and on the function of educating people [15]. Political education is essentially the work of nurturing people, so all work is student-oriented, around students, attention to students, service to students [16]. We will further explore the law, update the concept, innovate the model and build the ecology, and promote the tasks of university network political education of university network political education [17]. It is a major political task to promote and deepen the research of college network political education [18].
Deep neural networks can access the underlying essential features or rules [19]. Thus, it seems that deep neural network technology has a technical support function for political education theory courses [20]. This paper selects formative evaluation and outcome evaluation of ideological and political theory courses as the two evaluation methods of model construction. In the specific work of political education, we always take ''building moral character'' as the fundamental goal, take ''students as the basis'', pay attention to the process of students' growth and success, and at the same time pay close attention to the formation of the positive and healthy social mentality of students [21]. The political education work penetrates the whole process, to truly educate people in all aspects and the whole process. The implementation of political education work will ultimately be based on the growth of youth, which is conducive to the transformation of political education of college students from indoctrination education to generation education. In this article, wisdom education is explained as a large content, while political education is a part of wisdom education, or it can be understood that we just choose political education as wisdom education for research.

II. RELATED WORKS
Political education is an indispensable part of cultivating high-quality talents in society and the backbone of cultivating innovative talents, but political education for college students is a huge and complicated educational project. To make political education effectively implemented into the study life of college students [22]. In the long historical development of the whole society, political education in different historical periods has shown diversity and difference. Samek W et al. proposes to use a variety of delightful means, teaching in construction achievements, teaching in music, teaching in literature, teaching in the tour, etc., throughout political education, so that people can be educated subtly [23]. In the middle and late 1980s, some education scholars began to apply the theory and methods of fuzzy mathematics to the system of political education evaluation and conducted tests, evaluations, and statistical analyses of its actual effects, based on which the theoretical concept of ''political education evaluation'' was proposed. On this basis, they proposed the theoretical concept of ''political education evaluation'' and explored the principles, methods, and assessment criteria of this concept to determine its correctness and practicality. Cui X et al. proposed the weighted average method for the comprehensive weighted evaluation of political education in the process of research and evaluation of political education [24]. Ding in the study of the political education evaluation system pointed out that the process of implementing political education evaluation should follow the principles of policy orientation, effect fit for purpose, and assessment comprehensiveness for deployment and implementation [25].
Since the end of the last century, the field of artificial intelligence has flourished, producing several brand-new results one after another, many of which can be attributed to the wide application of deep learning algorithms. In the fields of computer science and automatic control, deep neural networks have been widely used to deal with various real-life related test problems, which have greatly improved the productivity of the related industrial chain. In the fields of computer science and cybernetics, deep neural networks are widely used to deal with various real-life test problems, such as rocket facilities, nuclear facilities, steel production, chemical industry, power supply systems, natural gas extraction, mineral extraction, advanced manufacturing, water hub control, environmental monitoring, medical imaging, railroad, and urban rail transportation management, civil aviation, urban water and gas supply, and heating, and other fields closely related to [26]. It has greatly improved the productivity of these industrial chains and ensured the safety and reliability of the corresponding work. To effectively overcome the effects of different viewpoints, different scales, partial occlusion, different lighting conditions and complex backgrounds on the traditional artificially designed features, the regression model based on deep neural network features mainly uses the advantage of a deep neural network to extract features to automatically obtain data features and design regression algorithms for prediction based on the extracted features [27]. Commonly used methods, such as image analysis processes operating in parallel with the network code, ensure that the target to be tracked or detected for analysis is searched within a multidimensional array in a specific detection space. If a target is found, the tester can then further tune the network code based on the feedback. For deep neural networks in the classical sense, it is easy to improve the transformed output of the system structure by using reasonable and effective parameters to moderately expand its ''width'' and ''depth''. However, the ''depth'' of a deep neural network should not be overextended, because the corresponding difficulty of training the neural network will naturally increase gradually with the increase of the depth of the neural network, which is not conducive to testing and verifying the overall execution efficiency of the deep neural network.
Through sorting out and referring to the existing literature, scholars have done a lot of theoretical and practical research on deep learning and real-time classroom evaluation. rising trend. It mainly provides the following inspirations for this study: First, the research on deep learning is flourishing, and the research results on its conceptual characteristics, practical strategies and evaluation implementation are abundant. They provide a theoretical framework for this study to analyze and construct deep learning. Detailed theoretical reference; second, classroom real-time evaluation has also developed to a certain extent at present, including research on connotation and function, research on the status quo of implementation, and research on implementation strategies. Discussion in this aspect has obvious enlightening effect on collecting data, clarifying research ideas and choosing research methods in this study. Although the results of deep learning are gradually increasing, the academic rationale is strong, and the operability and practicality are relatively weak. And the focus of practical research is mainly on deep learning technology innovation and machine learning, and there are few literatures that use deep learning theory as an entry point to systematically study the interaction between teachers and students in classrooms. Secondly, classroom real-time evaluation is one of the most frequently used and real-time evaluation methods in classroom teaching. Researchers generally draw corresponding conclusions by studying current phenomena, and are still in the stage of experience summary in terms of operational method research. The lack of theoretical guidance makes instant evaluation remain superficial and marginalized in classroom teaching, and does not fully play its due role. Finally, deep learning is one of the core research topics of learning science, which provides important theoretical support for optimizing real-time classroom evaluation. The performance and operability and the pertinence of deep learning need to be further investigated and explored. Therefore, this provides a certain space for the development of this research. In the existing research, the use of DNN/ML method for education requires a relatively large amount of data and computing resources, and the quality of the data is not very good, resulting in the accuracy of the results is not very good. Based on their research, we will further update and improve the dataset algorithm to make the results more accurate.

III. DEEP NEURAL NETWORK-BASED EVALUATION SYSTEM MODEL FOR POLITICAL EDUCATION OF COLLEGE STUDENTS
A. DEEP NEURAL NETWORK ALGORITHM Neurons in different layers are connected by weighting to form a neural network. Neurons apply activation functions on the weights of their input connections to obtain computational results. While neural networks are essentially based on the computer's recognition of the external environment, a neural network with deep learning capabilities can be understood as a neural network containing several hidden layers [28]. A multi-layer neural network can represent complex functions with fewer parameters, which is equivalent to making the neural network have sufficient self-awareness learning ability. And verifying such a neural network enables the researcher to see whether its actual self-learning ability meets the requirements.
An optimization problem is a problem of solving the minimum (or maximum) value of an objective function under certain constraints or without constraints. The data perspective describes the objectives of the classification problem and the regression problem. In this section, the task of deep neural networks is described from the optimization problem perspective. In the field of deep neural network learning, the objective of the optimization problem can be specified as finding a suitable mapping function in a space of mapping functions such that the empirical risk is minimized or the neighborhood risk is minimized. Assume that the mapping function f has a fixed form and is realized by a vector of real numbers w ∈ R d with its expressiveness and w is called the weights. Then we have.
To summarize and analyze various algorithmic techniques in deep learning, and discuss the challenges and opportunities for machine learning and possible future directions, it is necessary to validate the performance of deep neural networks such as robustness to continuously improve the training and design of neural networks themselves. The goal of optimization is to find the appropriate mapping function in the mapping function space F such that the loss generated by the model's incorrect prediction is minimized [29]. To reduce the generalization error, in addition to reducing the training error using optimization algorithms, problems such as overfitting need to be addressed.
As shown in Figure 1, a schematic diagram of the deep neural network algorithm flow is shown. For the training sample dataset, each sample starts with the same initial weight, and the first classifiers use a resampling technique to draw sub-training datasets from the samples. For samples that cannot be correctly classified by the first neural network, their weights are increased in the next round so that the next neural network can better classify samples that cannot be correctly classified by the previous neural network. Each training subset is generated based on the classification performance of the previous individual neural network. When all member neural networks are trained, the output of all neural networks are considered together when deciding the class of samples. In Figure 1, we visualize the computational logic of the algorithm, where multiple DNN networks are employed and an optimal structure is found. This process is achieved by manually tuning hyperparameters. DenseNet-40 adopts the third strategy to build topology, which proposes an effective dense connection topology. The model parameters used in this model have been greatly reduced compared with other models. Each sample is assigned a weight, representing the probability of the sample being selected into the subset, and each time a new member neural network is added, the weight of the correctly classified samples will be decreased and the weight of the incorrectly classified samples will be increased, thus achieving correct classification of difficult samples.
Neural networks simulate human learning behavior and consciousness by mimicking the organization of the human brain and are used to process the complex information in the network to achieve the desired purpose [30]. The neuron is the basic unit of the neural network and the basic information processing unit of neural network operation. Neurons are equivalent to nerve cells in the brain, and neurons are connected by prominences, each of which has its weight or strength as a feature. The dimensionality of DNN parameter optimization increases significantly, therefore, DNNs are usually trained using unsupervised pre-training, where the data abstracted from the previous layer is used as input to the next layer instead of optimizing all network parameters from scratch, to achieve better parameter initialization before supervised training. Parameter optimization of deep neural networks is a high-dimensional global optimization problem, and the WWO operator can be applied directly. We tried to use several popular evolutionary algorithms to optimize the deep neural network, among which the WWO algorithm shows superior performance. The sample training results of the deep neural network algorithm are shown in Figure 2.
In this paper, the data set is collected uniformly by our research group, and the data volume is 9124, of which 20% of the data is randomly divided as the test set, and the remaining 80% is used as the training set. DNN obtains the primary features of each layer through layer-by-layer data pre-training; distributed data learning is more effective (exponential); compared with shallow modeling methods, deep modeling can represent actual complex nonlinear problems in a more detailed and efficient manner. Deep learning is generally composed of multi-layered neural networks, where each layer is a linear transformation or a simple nonlinear operation, and a multilayer neural network is a combination of multiple simple nonlinear functions. Shallow layer networks are generally of limited use and can hardly reflect the complex relationship between input and output. Its training method is also very different from traditional neural networks: traditional neural networks set initial parameters randomly and use the BP algorithm to train the network by gradient descent until convergence. However, it is very difficult to train the neural network with a deep structure, and the forward propagation VOLUME 11, 2023 of residual differences will be severely weakened to the point of loss due to many layers, resulting in poor gradient spread.

B. MODEL IMPLEMENTATION
While deep neural networks are widely used, the existing quality problems have also been paid attention to. In recent years, more and more research has been devoted to solving the testing problem of deep neural networks. As a kind of software system, the deep learning system has the same solutions to some problems as the classical methods that have been widely studied in traditional software engineering. But since deep neural networks are constructed based on data-driven programming paradigms, their statistical nature presents additional and challenging research questions for software testing.
In this paper, python is used as the experimental language and the experimental environment is built under the Tensor-Flow framework. For the experiments, the codewords are BPSK modulated and coupled with Gaussian white noise codewords, and zero codewords are transmitted over AWGN channels to create training data with signal-to-noise ratios ranging from 1 dB to 6 dB. Finally, the bit error rate (BER) in the decoded code words is measured at the output of the network. In the experiment, the weights of the output layer of the network cannot be trained and should be set to 1, otherwise, the network will not learn [31]. This is because: the network can improve the loss by changing the weights of the last layer without improving its decoding ability. As shown in Figure 3, the trends are predicted based on the information results, their trends are predicted, and then the corresponding countermeasures are proposed. The input feature vector of the bionic algorithm is social relations. The collected 3*4-dimensional matrix data are used as network training This characterization ability should also be reflected in the features of the deep neural network, i.e., the features extracted from each layer of the deep neural network with better characterization ability should be aggregated according to the sample labels [32]. To verify the above arguments, in this paper, the DenseNet-40 model trained from the CIFAR10 training set extracts its fully connected layer input features and then visualizes the data by reducing them to three dimensions using principal component analysis and t-SNE, where each point represents a sample feature and different colors represent different categories. As shown in Figure 4, the model using this paper obtained the highest Fisher value compared to the model without data augmentation and using other data augmentation methods. It is of great importance to theoretically explore the connection between the decoding weights and the original training samples. This section mathematically analyzes the key steps of the proposed data enhancement method and derives the relationship between the weighted samples and the original input samples.
In addition to the possibility of the gradient approaching or becoming zero when the locally optimal solution is attached, another possibility is that the current solution is near the saddle point. The objective function reaches a local minimum in some dimensions at the saddle point, but a local maximum in some other dimensions. The performance evaluation of ideological and political education for college students must not only meet the goals and objectives expected by ideological and political educators and the society, but also conform to the internal law of college students' growth and success, and meet the needs of social development as well as individual development needs. The so-called effectiveness means that in terms of the value orientation of ideological and political education, social value and personal value are unified and measured. In the determination of evaluation standards, it is necessary to reflect both the advanced nature of morality and its universality. On the one hand, it guides people to carry forward the noble Social ethics, professional ethics, and family ethics, on the other hand, pay attention to cultivating rich and colorful moral personalities, moral sentiments, and cultivating a multi-level socialist moral culture. Evaluation criteria should not only highlight moral ideals but also cannot be divorced from moral practice.
The construction of the evaluation system should not only be based on the country, conform to the national conditions, but also face the world, so as to conform to the development trend of economic globalization. The evaluation index system must be less and more refined to avoid complexity. Through the evaluation index system and supplemented by certain measurement means, the evaluation object can be measured to obtain certain information and certain conclusions to achieve the evaluation goal. In the specific measurement process, the specific and quantitative index system is easier to measure than the abstract and qualitative index system, and the final index system is easier to measure than the first-level index system. In terms of the background environment of the evaluation, it pays attention to the combination of static evaluation and dynamic evaluation, strengthens dynamic evaluation and reduces the method of focusing on static evaluation in the past. In terms of evaluation content, it pays attention to those hidden, indirect, long-term effects that are easily overlooked by people. In terms of the scope of evaluation, it pays attention to the diversity of educational means, covering a complete evaluation space from on-campus behavior to offcampus behavior. Accumulate evaluation elements in daily life and realize the diversification of evaluation paths.
Plain gradient descent is a basic optimization method, and its update direction passes in line with gradient theory, with weight updates along the direction where the loss of the optimization problem decreases the fastest. However, deep neural networks are very complex, and the loss function for deep neural optimization problems is not a typical convex optimization, which may have several local optima.

IV. METHODS OF CONSTRUCTING EVALUATION INDEXES FOR POLITICAL EDUCATION OF COLLEGE STUDENTS
The evaluation of political education of college students is a relatively open system for all students. With the change of the times and the introduction of relevant national policies, students, universities, and society will have various needs, so the evaluation system of political education of college students also needs to make corresponding changes to meet the needs of the further development of society [33]. The complexity of social relations, increased scholars have started to pay attention to whether the evaluation method of political education is single, whether the evaluation process is effective, and whether the evaluation index system is reasonable and comprehensive. Political educators are required to always grasp the new national policies and requirements of political education work under the new situation, and integrate them into the evaluation work of political education of college students promptly. The political education work of college students is implemented continuously, based on the realistic layout and implementation of the Party, government, and society's demand for higher education goals. Contrast is a very important feature of the evaluation system, there is contrast for deficiency, adjustment, and growth, and the same is true for the political education evaluation system. Implicit political education of college students has been continuously promoted and strengthened, and certain results have been achieved, mainly in that implicit political thinking has been paid more attention, education practice has been gradually standardized, education carriers have been increasingly diversified, education methods have tended to be diversified and implicit political thinking [34]. The main results are that implicit thought politics is more emphasized, educational practices are gradually standardized, educational carriers are increasingly diversified, educational methods tend to be diversified and implicit thought politics is more recognized. In the on-site individual interviews, when asked ''how much importance your school attaches to implicit political education'', 78 people (26%) answered ''attach great importance to it''; 201 people answered, ''attach more importance to it''. 201 people (67%) answered ''attach more importance''; 8 people (2.7%) answered ''do not attach importance''. ''Don't know'' 13 people, accounting for 4.3%. The total number of people who ''attach great importance'' and ''attach more importance'' is 278, accounting for 93% of the total number of recipients. most college students have personal experience [35]. In the interview, many college students said that in recent years, the political educators in colleges and universities do not only stay in the traditional explicit political education methods but also gradually adopt the social practice, ''curriculum thinking politics'' and other implicit political education methods. In addition, they also noticed that some slogans about political education were posted in the canteen and dormitory of colleges and universities, and more green plants were placed in teaching buildings and libraries. This shows that colleges and universities pay more attention to implicit political education and environmental education work and have gradually implemented it into action.
In the interview, most college students said that colleges and universities are continuously tapping the resources of implicit political education and investing a lot of human and material resources, and the ways of implicit political education practice have moved from single to multiple, breaking through the limitation of time and space and permeating all aspects of the campus. The practical education of implicit political education for college students is gradually becoming standardized. As shown in Figure 5, only 34.76% of college students choose the explicit political education method of ''Civic Education Course'', and more college students prefer the implicit political education method. Although these two-education methods account for a large proportion, not a few college students choose a social practice, cell phone apps, and the internet. In Figure 5, all the scales increase with time and eventually smooth out. The final proportion from high to low is Ideological and political courses, Course ideology and politics, Campus Cultural Activities, Social practice activities, APP Development Educational Resources, Hot event entry. The common methods of weight design are the direct judgment method, empirical weighting method, Delphi method, hierarchical analysis method, and weight factor analysis method. The weight design for index evaluation is mainly a response to objective observation and subjective evaluation of evaluation indexes [36]. The methods used have a strong theoretical basis for the evaluators, only in this way can more college students accept them in the process of evaluation.
To make the evaluation index system more scientific, reasonable, and accurate, we need to conduct reliability tests. In the reliability test of evaluation indexes, we choose one or more groups of evaluation indexes to test the data for a fixed group of evaluated people. In the process of testing, if the results of the test are less inaccurate than the actual situation, we can assume that they are consistent with the characteristics of the group. This also reflects that the design of the evaluation index system is in line with the objective reality. If the data error of the evaluation is large, it is difficult to reflect the objective actual situation of the evaluated group accurately and comprehensively, which will reduce the trust of the whole evaluation index system test. A test prediction is defined as a test reference that determines an expected result compared to the actual result at test time. Similar to the testing of traditional software systems, the testing of deep neural network-based systems also needs to solve the test oracle problem to measure whether the results of the deep neural network after various given inputs meet the expected output criteria. Since deep learning-based software itself is constructed based on a data-driven programming paradigm, testing whether the data has sufficient labels is critical to detecting bad behavior in DNN-based software.

V. CONSTRUCTION OF POLITICAL EDUCATION EVALUATION INDEX SYSTEM FOR COLLEGE STUDENTS
There is a clearer direction and scope for the evaluation of classroom teaching in higher education Civics, and the scope is identified as the university level because Civics is offered at all stages of universities, schools, and colleges in China. A subject, a course to gradually improve and mature, that is bound to go through a long period of development, it must make corresponding changes according to the times, social environment, and historical environment. Civics has a strong contemporary nature, we can easily find through the history of the practice, Civics courses offered, the teaching materials used, the teaching cases selected and the means and methods are closely followed by the times.
The curriculum is the backbone of the school's teaching work, and generally, the curriculum consists of compulsory courses and corresponding elective courses. 14.22% of the students said Yes, and it is a compulsory course, 56.42% said Yes, but it is an elective course, 17.38% of the students said no, and 9.98% of the students said, not sure''. As shown in Figure 6, most colleges and universities only offer elective courses, and even some schools do not offer relevant courses, which is an important bottleneck that affects the function of political education of national defense education in colleges and universities. In addition, teaching materials are an effective carrier of school teaching work, the author surveyed political education-related teaching materials. 35.01% of students said ''yes'', 60.5% said ''no'' and 12.82% said ''not sure'', which shows that most colleges and universities do not issue relevant teaching materials to students. This makes the function of political education in colleges and universities become rootless water and a rootless source.
The purpose of test adequacy is to discover whether a test has good fault revealing capabilities. It provides an objective confidence measure for testing activities. Adequacy criteria can also be used to guide test prioritization. In traditional software testing, code coverage measures how well the test suite executes the program's source code. The higher the coverage of the test suite, the more likely it is that hidden bugs will be discovered. In data-driven deep learning, the model mainly uses the available validation data set for evaluation and testing, and the data set is usually divided into a part of the test data to verify the generalization ability and accuracy of the model. But these datasets often fail to cover the various corner cases that can lead to unexpected behavior. 65.54% of the students are very much looking forward to the successful convening of the 18th National Congress of the Communist Party of China; 92.5% of the students agree with ''building a harmonious society''; 37.5% of the students believe that the Party Central Committee's anti-corruption work has been effective in recent years; To express their attitude, 94.6% of the students clearly hope that the Taiwan Strait issue can be resolved by peaceful and diplomatic means as soon as possible, and they should focus on economic development, which shows the rational patriotic enthusiasm of our students. On the question of whether they are willing to work in the west, 50.2% of the students think that they may consider starting a business. Therefore, in order to test potential quality issues in deep neural networks, inspired by test adequacy in traditional software testing, software engineering researchers propose a number of Multiple test methods are available to evaluate DNNs. On the question of ''whether the evaluation indexes of classroom teaching in Civics are clear and easy to understand'', 45.4% of the students strongly agreed, 35.7% agreed, 9.3% agreed, 8.9% disagreed, and 3.02% disagreed, as shown in Figure 7. The responses to the question ''The evaluation indexes of classroom teaching in the school's Civics class can fully reflect the actual situation of classroom teaching in Civics class'' are, in descending order of proportion, 51.52% strongly agree and generally agree, 24.15% agree, 18.32% disagree, and 6.01% disagree. 6.01% agree with it. According to the survey results, there are certain problems in the setting of the evaluation index system, the evaluation indexes are not clear enough and not easy to understand, and setting of the evaluation indexes cannot cover the whole classroom teaching of Civics.
The evaluation criteria used in the classroom teaching evaluation of Civics and Political Science courses are no different from those used in other general education or professional courses, accounting for 30.17% of the total. In response to the question ''Who sets the evaluation standards for classroom teaching in Civics courses? The percentages of respondents who answered ''who makes the evaluation standard of Civics class'' are 50.87%, 63.09%, 46.63%, 28.42%, 24.43% for the relevant departments of the Ministry of Education and the university. The result of the comparison is the result of the formative assessment of political education. In the traditional assessment of ideological and political courses, teachers are the main assessors. In the assessment process, even if there are uniform assessment indexes, different assessors will have different tendencies. Machine examiners applying deep learning technology, on the other hand, will follow the same assessment criteria based on the original model, maintaining the homogeneity of the criteria in the assessment process while ensuring uniformity and scientific in the output of results. In addition, machine examiners save a large amount of data while assessing, accumulating a large amount of raw data for the development of educational artificial intelligence, and providing a massive database for the scientific development of political education.
Only by continuously strengthening the relevant theoretical research can we ensure the continuous innovation of evaluation concepts, evaluation methods and means, and evaluation content. The complexity and enormity of the evaluation process of classroom teaching evaluation in higher education have put forward higher requirements for theoretical innovation research. ''The construction of basic theory should not be neglected. In addition, the theory of innovation and research to combine with practice. In the process of evaluating the practice of classroom teaching in Civics, we should be good at finding problems and raising them to the height of theory for research, and then use them to guide the development of practice.

VI. CONCLUSION
This paper constructs a political education evaluation system for college students based on deep neural networks. In the deep neural network model optimization, the structural description of the deep neural network is given first, and then specific modifications to the deep neural network algorithm are given in the complexity problem, and the effectiveness of the newly proposed algorithm is demonstrated using experiments. The low-complexity offset minimum sum (OMS) deep neural network algorithm can save about 60% of the number of networks and 55% of the training time.
To address the shortcomings of individual neural networks, this paper proposes a deep neural network integration model based on evolutionary algorithm optimization, which uses an evolutionary algorithm to optimize the weights of each member neural network in the model. Experimental results show that this deep neural network integration method further improves the performance of unbalanced data classification. The assessment and evaluation system of ideological and political theory teaching in colleges and universities should focus on the unity of internalization and externalization of course contents, and the unity of knowledge and action. The deep neural network technology embodies such an evaluation requirement using computation, and it is worth thinking about and exploring to solve the difficult problems in the construction of political education subjects in colleges and universities with the power of science.
Compared with other research results, the research results of this paper are more accurate and more efficient. Therefore, our model can be applied to actual course evaluation, which can save a lot of labor and time costs. This study has achieved certain results, however, due to the limited time and effort and the small number of experimental subjects, the wide applicability of the model cannot be fully illustrated. The model can be used in more university practices in the future to further verify its extensiveness and effectiveness.