Human-Centered Computational Intelligence Information SystemsView this Special Issue
Construction and Application of Music Teaching Resources Based on Recurrent Neural Network
Along with the rapid development of informational techniques, educational techniques are becoming increasingly important in the field of education, encouraging the reform and innovation of traditional educational concepts and teaching methods. Various music teaching aids have also become widely available. However, there is a scarcity of assistive software that is truly appropriate for music classroom instruction. A neural network is an artificial model that models and connects neurons, which are the basic units of the animal or human brain, to stimulate the nervous system’s learning, association, memory, and pattern recognition functions. We propose a recurrent neural network (RNN)-based music teaching resource construction, method, and application in this paper, focusing on the needs of music distance learning, analysing the user role needs of teachers, students, and system administrators involved in teaching and learning, and designing a mobile teaching platform in line with music teaching based on the characteristics of music teaching. The experimental results show that the system’s accuracy increases steadily with the number of iterations, and the accuracy fluctuation is stable, with an average accuracy of 72 percent and good system stability. As a result, research into this topic has significant implications for improving the quality and efficiency of music education in China, reforming the current school music education model, developing social music education, and realising music education for all and lifelong music education.
The advancement and development of network and multimedia technologies has provided new support for educational reform and development, as well as a new way of learning, namely e-learning . Music courses, which are popular among students, are generally valued by schools and society . Art education is an important way to provide high-quality education in schools . With the introduction of computers, many problems that existed in traditional teaching, such as single teaching mode, lack of teaching resources, and boring teaching contents, are now being addressed. Traditional classroom education is changing from its original single mode. In particular, in recent years, the rapid advancement of Internet technology has enriched and diversified the original online education model based on the B/S model, and people have gradually accepted the mode of completing various learning tasks via the Internet. It has transformed the content, pedagogy, teaching approach, and teaching mode of instruction.
According to the current stage of China’s school music curriculum construction, whether explicit or implicit online music resources, real music resources can be developed and used to promote the harmonious development of students’ body and mind, as well as serve as music teaching resources. The RNN exhibits extremely complex dynamic behaviour because of its fully interconnected structure. Neural networks [4–6] have good fault tolerance, learning, adaptability, and associative memory and can process information in parallel. RNNs outperform traditional educational models, when compared to traditional models. Speech recognition, video action analysis, image recognition [7–9], and signal classification [10, 11] have all been successfully applied with RNNs in practise.
With the advancement of educational techniques in China, the rapid growth of network technologies , and the diverse educational reform needs of various disciplines, combining traditional subjects with modern network technologies, makes a lot of sense. Following the computerization of music teaching sources, teaching resources are rapidly disseminated on the network using computer techniques and the Internet, making it easier for students to obtain music teaching sources and knowledge . Learning will no longer be influenced by geography or time, and an increasing number of students will benefit from education informatization . More music classrooms are appearing in multimedia teaching systems, and more music teachers are using the Internet to access textbooks and course production materials . The ultimate goal of taking advantage of the benefits of online music resources and judiciously integrating and allocating music teaching resources is to enrich school music classrooms. The thesis’ RNN-based music teaching system can be applied to music education, allowing students to create a comprehensive music learning platform through regular study, homework practise, and information notification functions. The music learning environment improves the learning effect by changing the previous mode of only learning and not practising on the Internet. The following are the paper's innovative points:(1)Using constructivist theory as a guiding basis for curriculum development and design, the system is used to combine RNN with music subject teaching and learning.(2)This is an attempt to teach resource recommendations through this topic. A solution to meet the demand for teaching resource recommendation is devised and realised based on the characteristics of teaching resource recommendation in the teaching aid system and the mature technology of the existing recommendation system.(3)In the field of pedagogy, RNN’s findings enable computers to receive education and improve intelligence, and RNN's research findings are applied to the education process and teaching to improve education efficiency, stimulate educators’ innovative thinking, and generate new teaching models.
2. Related Work
2.1. Construction of Music Teaching Resources
Learning resources are the foundation of online teaching. The development trend of online education is to integrate and classify various excellent learning resources to create a teaching resource library. Thanks to the music teaching assistant system, students can select the time, place, and learning content to suit their situation. Through the teaching platform, it provides professional support for teaching and learning in the music business and provides great convenience for remote students through the cell phone network. This diversified learning mode not only satisfies the normal classroom learning but also allows students to complete practice or exam operations through the platform, making learning more effective. Therefore, research on the construction of music teaching resources is imminent.
Guan and Zhang have successfully launched the Moodle platform. As it is developed based on advanced teaching concepts, it provides a free modular online course platform that is simple to operate and easy to use in practical applications. Its openness, flexibility, interactivity, and many other advantages have gained worldwide recognition. It has received a lot of attention and interest from leading researchers and professors within . Yuan has also developed many similar online learning service systems. The basic design model is the same, i.e., the full use of computer multimedia technology to impart knowledge to learners through video and audio . Zuo and Jia divided the content of music education into four areas: feeling and appreciation, performance, composition, and music and related culture. On the one hand, the knowledge space of the original curriculum area is extended, and new learning content is injected, and on the other hand, a new learning area is added . Jian et al. established a fee-based online learning system. By purchasing learning resources through payment, one can complete the paid resources for an unlimited number of times. This type of learning has become a useful aid to campus education to some extent . Zhang and Huang established dynamic web-based technology support with the two-way transmission of network information, which provides a strong technical guarantee for web-based discussion, communication, interaction, and collaboration between music students and coaches .
2.2. Research on RNN
Network technology has developed in the direction of broadband, high-speed, and multimedia, and network teaching has penetrated every corner of the world with its novel and convenient features. The networking of teaching activities determines the birth of an online teaching platform, which in turn promotes the new learning style of online learning. Being an essential branch of machine learning, RNN can be applied to a vast array of AI application scenarios. Thus, nonindependence between elements of the input and output data streams can be modeled. In addition, more sophisticated RNNs can model the serial relationships and temporal dependencies of data in multiple dimensions simultaneously.
Saito et al. ushered in the era of artificial neural network research by modeling a formal mathematical description of neurons and the structure of networks and showing that individual neurons can perform logical functions . Liu and Huang proposed the LSTM, which successfully solved the leakage gradient problem . Experiments have shown that LSTM not only can effectively solve the gradient disappearance problem but also is easy to use and gradually becomes a standard method to deal with the gradient disappearance problem. Lv et al. assumed that the strength of synaptic connections between neurons is variable and proposed a learning criterion for neurons on this basis, which laid the foundation for neural network learning algorithms . Benrabah et al. proposed an online learning resource recommendation system that records explicit feedback from users on recommended content and implicit feedback on recommended content conveyed by users' browsing behaviour . Xia and Yan created a different type of learning neural network processing unit, the adaptive linear element, and also discovered a powerful learning rule, commonly referred to as the Widrow-Hoff learning rule .
Neural networks have embarked on a steady path of development, and various neural network theoretical models and learning algorithms have been proposed one after another, making promising progress in biomedical and other fields.
3. The Concept of Building Music Teaching Sources on the Basis of RNN
3.1. Teaching Repository System Structure Design
Any software or system must be stable, and good stability can greatly improve the user experience . The system is built on the Java EE framework, and the client runs on Android. The system is divided into three tiers to provide maintainability and scalability: functional representation tier, business logic tier, and data tier. Students can easily request learning services on the web that are tailored to their specific needs, and the entire service delivery process is automated. The formulas for calculating the weights and correction thresholds are as follows:
The scope of application of the system is mainly music classroom teaching, a professionally demanding application scenario where the subject of music itself is much focused on sensory experience and quality . The unified client simplifies system development, maintenance, and use by centralizing the specific implementation of core system functions on the server. The architecture of the music teaching system is shown in Figure 1.
The structural design of the teaching resource database system is shown in Figure 2.
First, the data bootstrap layer provides the data access interface to the system and the corresponding data operations with the database platform, and the system database uses an SQL Server database with the database access interface defined in the system data layer. When developing and designing the system, care must be taken to extend the depth and breadth of its capabilities so that the overall system meets the needs of the curriculum design while providing a valid baseline and value for the curriculum . In addition to studying uncertain and imprecise knowledge representations, imprecise inference methods should be explored to compute the output of each neuron in the hidden and output layers of the network.
The traditional approach is to use a measure of proximity, expressed in most cases as a sum of squared errors (SSE). The SSE is defined as follows：
--Number of clusters.
--Class. --Number of samples.
When a user visits a detailed page of an instructional resource, the recommendation system recommends instructional resources that are similar to the current resource and that may be of interest to him based on the currently visited resource and the user’s habitual records . All units in a set of network units compete with each other for the ability to respond to external stimulus patterns, and the unit that wins the competition inhibits the response of the unit that loses the competition to the stimulus pattern. A potential representation of the learner’s historical learning record is extracted, taking into account the learner’s characteristics, to generate the final recommendation results for the target learner.
Second, the business logic tier is used to define the operational rules of the system. The business logic tier is on the server side of the system and is used to define the business logic of the system. Since the application logic is on a central server, users can use different hardware, including database systems, network operating systems, and so on. Only a browser needs to be installed on the customer side, and the browser exchanges information with the database through the Web server. The nodes in the hermit phase of the three-layer system are not chosen arbitrarily; in this system, the numerical number of nodes in the entry and exit layers can be determined, and the number of nodes in the hermit phase can be given according to the empirical formula：
The information transfer process of neurons has a time lag, which means that the network model must be related to the state of the neurons over time, which reflects the properties of the neurons themselves . The main idea behind RNN-based music teaching resource integration is to use an image acquisition device to collect hundreds of images from 20 commonly used Arduino devices to create image array data. Following that, the image data are used to train a convolutional neural network model, which is then used to classify and identify Arduino devices to generate the measured device categories. Finally, the functional presentation level is the interface between the system’s operational users and the system, which can accept input from the users, send the input or selected operation items from the client to the server, and display the processing results on the server:
Finally, the functional presentation layer is the interface between the system operation user and the system, which can accept input from the operation user, submit the input or selected operation items from the client side to the server side, and display the processing results on the server side at the same time. Its primary responsibility is to programme, assemble, and debug the system from the ground up. Teachers can use the online teaching resource library to help them prepare lessons, from finding the most up-to-date reference materials for subject teaching to searching various media materials related to teaching contents and using the online lesson preparation system and tools to integrate and create teaching materials that are tailored to the course’s teaching design. The data from training samples are added to the network’s input, and the expected output is compared to the network’s output to obtain the error signal, which controls the adjustment of the weights’ connection strength and converges to a definite weight after several training sessions.
3.2. Design of Functional Module of Teaching Resource Database System
The functions of the network teaching resource database determine the structure of the network teaching resource database, which in turn determines the design and development of the network teaching resources. The network teaching resource database system divides the whole system into several basic functional modules, each module provides different services for students or teachers, such as the login and registration module, teacher space module, question and answer module, news announcement module, backend management module (personnel management module), and so on. Each part has different functions. In the case that the nodes in the hidden layer can be set freely according to the need, the three-layer forward neural network can be used to achieve any continuous function with any approximation. Let the output of neurons from layer neuron to layer neuron under samples.
This repository management system mainly consists of three major functional modules, namely, resource query module, resource upload module, and a resource download module. The workflow of the repository service system is shown in Figure 3.
First, the resource query module must include Boolean query functions, relational queries, exact queries, and approximate queries. The online teaching resource library system is a system with access rights control. Therefore, from the operational practicality of the system homepage and the simplicity of the system program, we put the login interface in a homepage to complete, including three login methods of administrator, teacher, and student sessions. According to different user types, users can choose their login methods. After the user logs in successfully, enter keywords in the search box at the top of any page, click Find Resources, and the page will return all information about the resources, including the keywords. The number of training samples is defined as , the weight of the neural network is . The neural network training function is the mean squared difference objective function .where α and β are regularization coefficients.
Second, the resource upload module must have the function of remote submission of materials, and users can submit materials remotely through the Internet. After the user logs in successfully, click the resource upload button to enter the resource upload page, fill in the resource information, and upload the file to upload the local resource file to the library and publish it after the main user has reviewed and approved it. In this paper, in the process of training the convolutional neural network model, images are converted into TFRecord files as the initial data input of the network. Students’ basic knowledge structure is assessed before advanced courses through student learning history and knowledge tests. For example, learning data structures requires proficiency in discrete mathematics and programming languages, as well as Hopkins Computing:
In the multimedia teaching environment equipped with a network, teachers can call various media information and teaching materials from the resource library through the classroom teaching application system, including the teaching materials constructed by teachers and existing in the library. It creates a multidimensional learning space for students and provides powerful support for multimedia portfolio optimization teaching. Therefore, the overhead of deleting a WME is one more overhead of finding memory than that of adding a WME. And the overhead of finding memory at this point is as follows:
In the electronic circuit of the neural network, there is inevitably a time delay due to the limited switching speed of the operational amplifier and the long signal transmission distance. Therefore, the convolutional neural network outputs the network model by multilayer convolution and pooling operations after completing image data processing. The results such as test accuracy values and error loss functions describing the quality of the model are observed and analysed by comparing the error information of the actual output and the desired output. We used here data range normalization, which is done by first finding the range of each variable and then calculating the mean of each variable as follows:
Finally, the resource download module provides content transfer management and supports functions such as media download and upload. After the user logs in successfully, click the resource link to enter the resource detail page, and click the download button to download the resource to the venue for learning and use. The first problem faced when using neural networks to solve real-world problems is how to determine the structure of the neural network, which consists of two main aspects: the configuration of the neural network and the optimal number of hidden layer nodes. Any state depends and only depends on the current input and network state of the previous step; however, this hidden state of each step can contain a nearly arbitrarily large context window; this is possible because the number of states that can be represented is determined by the size of the hidden layer nodes.
4. Application Analysis of RNN in the Construction of Music Teaching Resources
4.1. Stability Analysis of RNN
There is usually only one equilibrium point at the origin for linear systems, but there may be multiple equilibrium points for nonlinear systems, and the equilibrium point is not always at the origin. A learning algorithm is a procedure for completing the learning process. Its function is to change the synaptic weights of the network in an orderly manner to achieve the desired design goal. Although the results of each module’s operation match the reference results in most cases when randomly generated values are used in the weight matrix and the vector to be computed, individual values are still subject to large errors. The error between the network output vector and the standard reference varies with the number of cycles for various common parameter settings, as shown in Figures 4 and 5.
First, the common method is to perform a coordinate translation transformation at the nonzero equilibrium point or to perform a coordinate translation transformation on the nonzero solution to transform the stability problem of the nonzero equilibrium point zero to the stability problem of the origin. Resource attributes can be accurately labeled, which will affect resource retrieval and accuracy when using the entire resource management system in the future. At the beginning of the simulation, there may be problems with the verification platform and the design to be tested. We should start with a simple scenario. This phase of the simulation is more like a cross-checking process between the verification platform and the design. The function fetches an object from the VST soft source plug-in before rendering, then determines whether the object is empty, and simply exits if it is. This process continues until the output data values converge and converge to a steady state or the number of iterations ends and the model training is finished.
Second, because the equilibrium points of the two systems before and after translation in the two coordinate systems are equivalent in terms of stability, all that is needed to determine the stability characteristics of the nonzero equilibrium point of the original system is to study the stability characteristics of the transformed system’s equilibrium point in the new coordinate system, i.e., the origin. The system information message, like some instructional notifications, has two parts: basic information and relevant attachment information, such as online browsing and query information and notifications, as well as background data maintenance. Inputs to some or all of the nodes control the entire network, and the inputs generate triggers that propagate through the network’s weighted connections. If there are any issues or unexpected situations during the simulation, the platform should be checked for defects, the configuration of the design to be tested and the data excitation should be checked for compliance, and the functional implementation between the two is equivalent, then the VST plug-in enters the audio processing phase. Two parameters, inputBuffers and outputBuffers, must be initialised before proceeding to this phase.
Finally, because any system can perform the above coordinate transformation and the stability of the equilibrium point before and after the transformation is equivalent, we assume that the system’s equilibrium point is at the origin of the coordinates in the following description of the stability problem. The simplest combination of configuration parameters is tested first when simulating and verifying the gate control module using the full verification platform, and the travel data excitation is also set to a fixed value and written to the generated weight matrix. The system examines the information provided by the user and determines whether it is legal or not. If the user is not a legal one, it does not exist; if it is, it’s still determining whether the user is already in the administrator table. They cannot be readded if they have already been added. Advertising is the information publishing class, which is the main class of information and notification class in the design of information notification class. To add, modify, delete information, and perform other operations, the related properties and methods of the info class and notification info class are inherited from Advertising.
4.2. System-Level Verification Analysis of Neural Network
Since learners learn from a variety of different learning resources during the learning process, a large number of learning relationships are generated for learning. Therefore, according to the requirements of module-level validation work, all functions of a module must be tested in module-level simulation validation.
First, the simulation operation revolves around three classes of data: configuration parameters, weight matrices, and feature vectors because the system-level verification platform does not introduce new classes of data after integrating each module-level verification platform. The MLP can be thought of as a function that maps input vectors to output vectors. The behaviour of this function is determined by the network’s connection weights. The model testing accuracy is not high and easy to oscillate, the error loss function does not converge, and the weights do not reach the optimal solution, even though the extracted features can express a degree of abstraction. The introduction of floating-point numbers in the accumulation part, the measure of extending the bit width of the registers for storing the memory state, and the effect of the nonlinear function on limiting the range of values can effectively suppress the errors and relate to the ultimate practical value of the designed hardware gas pedal. Since the length of the recommendation list can have a large impact on the recommendation performance, ontology + neural network and collaborative filtering + ontology are used as the evaluation index of RNN, and the comparison of the quadratic error functions of the three is shown in Figure 6.
Second, before the computation officially starts, a handshake signal is passed through an external port, written to the RAM of each gating module via the controller, and then the control module port reads it at runtime. For the fixed-structure network approach, the network gives the probability of a correct response to other randomly chosen inputs after training. The continuous optimization of the network model parameters through continuous iterative operations is the main task of the training process. A momentum term, which is a certain multiple of the weight change in the previous step, is added to the weight update rule, so that the weight change is not only influenced by the current gradient information but also related to the previous weight change. For the training time, the MGU with 200 hidden nodes runs almost simultaneously with the GRU with 100 hidden nodes. For comparison, Table 1 compares the experimental results of MGU and GRU.
The combination of linear operations is ultimately linear, which means that a multilayer perceptron with multiple linear hidden layers is fully equivalent to a multilayer perceptron with a single hidden layer. Therefore, during module-level verification, the hidden problems in the underlying verification components are verified and corrected, various data interaction operations are verified, and the module-level verification debugging phase system is smooth. The model accuracy changes when different iterations are set. Eight datasets with iterations between 1,000 and 8,000 were selected to observe the change in accuracy. The results are shown in Table 2.
As the number of iterations in the network changes, it is possible to understand how the test accuracy changes from the RNN’s system. The number of iterations increases sequentially, and the accuracy of the system steadily increases, with smooth accuracy fluctuations and an average accuracy of 72%. Therefore, the stability of this system is good.
The optimization measures of range and accuracy of calculating fixed-point numbers in the design were removed, the whole process of using fixed-point formats was tested, and the results were recorded statistically, replacing the lower representation range and lower accuracy with multiple data formats. The network calculation error and inference accuracy for different numerical formats are shown in Figures 7 and 8.
Finally, when memory cells are integrated into a system-level verification platform, sequencers, and controllers are no longer required as verification components in the input data ports of each memory cell module, and the corresponding sequences in the memory cells are no longer required. The details of the two differ during training in practise, owing to the steep part of the tanh function curve. To construct the embedding of entities and relations, the relations are considered transformations from head entities to tail entities. The decoder network receives the output results, the selected candidate recommendation resources, and the user entity representations via the two-way GRU layer, the two-layer forward neural network, and finally the softmax layer.
Aesthetic education can be implemented through music education, which is an important part of basic education. In recent years, China has made significant progress in developing online music education resources, with notable results. The accumulation of teaching resources, on the other hand, is a long-term process that involves professional and technical personnel, as well as teachers and students. However, the problem of “a small number of regional online music education resource sites and limited coverage” still exists in the development of online music education resources. Although RNNs are widely used in fields such as time series analysis, system identification, and fault diagnosis, their feedback mechanism allows them to process dynamic data and directly reflect the dynamic process of information system characteristics. To that end, this paper proposes an RNN-based approach to the creation and application of music teaching resources, analysing current problems and the teaching process in the music classroom in the context of the digital and information era. A music classroom teaching resource system is designed after a comparative analysis of current popular music teaching aids, and the most suitable RNN is selected for implementation based on the system’s characteristics. This study addresses the system’s need for music curriculum development, realises the integration of RNN and music teaching, serves as a model for the development and application of music teaching resources, motivates music education reform, and promotes the development of music teaching, as well as the development of the entire education and ideology business.
The data used to support the findings of this study are available from the author upon request.
Conflicts of Interest
The author does not have any possible conflicts of interest.
J. Li, T. Dai, Q. Tang, Y. Xing, and S. T. Xia, “Cyclic annealing training convolutional neural networks for image classification with noisy labels,” in Proceedings of the 25th IEEE International Conference on Image Processing (ICIP), pp. 21–25, Athens, Greece, October 2018.View at: Google Scholar
L. Qian and C. Zhu, “Modulation classification based on cyclic spectral features and neural network,” in Proceedings of the 3rd International Congress on Image and Signal Processing, pp. 3601–3605, IEEE, Yantai, China, October 2010.View at: Google Scholar
J. Zuo and P. Jia, “Empirical research on evaluation system of college mathematics teaching quality based on bp neural network[J],” Revista de la Facultad de Ingenieria, vol. 32, no. 14, pp. 188–193, 2017.View at: Google Scholar
W. Jian, W. Wei, and J. M. Zurada, “Computational properties and convergence analysis of BPNN for cyclic and almost cyclic learning with penalty,” Neural Networks the Official Journal of the International Neural Network Society, vol. 33, pp. 127–135, 2012.View at: Google Scholar
M. Liu and J. Huang, “Piano playing teaching system based on artificial intelligence – design and research,” Journal of Intelligent and Fuzzy Systems, vol. 40, no. 1, pp. 1–9, 2020.View at: Google Scholar
M. Benrabah, K. Kara, O. Aitsahed, and M. L. Hadjili, “Constrained nonlinear predictive control using neural networks and teaching–learning-based optimization,” Journal of Control, Automation and Electrical Systems, vol. 32, pp. 1–16, 2021.View at: Google Scholar
Q. Peng, “Optimization of physical education and teaching quality management based on BP neural network[J],” Boletin Tecnico/Technical Bulletin, vol. 55, no. 7, pp. 643–649, 2017.View at: Google Scholar
Q. Yan, S. F. Kong, and H. B. Liu, “On the construction of risk validation model for E-government system based on rough set-BP neural network,” Zhongguo Huanjing Kexue/China Environmental Science, vol. 37, no. 10, pp. 3708–3721, 2017.View at: Google Scholar
Q. Qiu, “The application of neural network algorithm and embedded system in computer distance teach system,” Journal of Intelligent Systems, vol. 31, no. 1, pp. 148–158, 2022.View at: Google Scholar