A parallel computing method for K series models based on ODE solution

Abstract

The present invention discloses a K series model parallel computing method based on ODE solution. Firstly, the K series model neuron equation is taken as the research object. On the epileptic electroencephalogram data set, the computing performance of different ODE solutions of the model neuron equation is compared and analyzed. Under the condition of ensuring the accuracy, the ODE solution with high computing efficiency is selected. Finally, the optimal parallel computing mode is selected by comparison to perform parallel computing on the K series model, so as to further improve the computing efficiency of the K series model and solve the problem that the prior art pays more attention to the model recognition rate and ignores the low model computing efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of olfactory nervous system models, and in particular to a K series model parallel computing method based on ODE solution. Background Art

[0002] The research progress of the olfactory nervous system can be divided into three aspects according to the research level:

[0003] (1) Research at the macro level mainly focuses on the research of olfactory nervous system models, such as neuron modeling and modeling of certain parts of the olfactory system. The most representative of these is the K series model, which is an olfactory model that has been continuously developed and improved by Professor Walter J. Freeman, former president of the International Neural Network Society and a neurobiologist at the University of California, Berkeley, in his nearly three decades of research. Among them, the KIII model focuses on modeling the olfactory bulb layer. This model has many characteristics that are different from general neural models, such as model structure and model characteristics.

[0004] (2) Research at the microscopic level mainly focuses on the mechanism, for example, how the sense of smell encodes odor information and the working mechanism of neurons. Monahan et al. used mice as the research object and revealed from the genetic level why the olfactory nervous system can distinguish tens of thousands of odors.

[0005] (3) Research at the functional level mainly focuses on the relationship between olfactory processed odor information and higher-level neural activity. Most studies use electroencephalogram (EEG) or functional magnetic resonance imaging (fMRI) as research methods.

[0006] However, in the application research of existing K series models, more attention is paid to the model recognition rate, while the problem of low model calculation efficiency is ignored. Summary of the invention

[0007] In order to solve the above technical problems, the present invention provides a K series model parallel computing method based on ODE solution, which improves the computing efficiency while ensuring the accuracy by combining the optimal ODE solution with the optimal data parallel computing method.

[0008] In order to solve the above technical problems, the present invention provides a technical solution: 1. A parallel calculation method of a K series model based on an ODE solution, characterized in that:

[0009] Step 1: pre-process the obtained initial sample data to form multiple groups of sample data, and divide each group of sample data into multiple training sets and test sets according to different proportions;

[0010] Step 2: Select training sets and test sets, train and test different ODE solutions for the K series model neuron equations, and select the ODE solution with high computational efficiency as the optimal ODE solution while ensuring accuracy;

[0011] Step 3: Solve the neuron equation of the K series model according to the optimal ODE solution obtained in step 2, select the training set and test set data, use different data parallel methods for parallel computing, and select the optimal data parallel method according to the parallel computing evaluation index while ensuring the accuracy;

[0012] Step 4: Calculate the K series model neuron equations and apply the K series model based on the obtained optimal ODE solution and optimal data parallel method.

[0013] Furthermore, in step 1, the process of preprocessing the initial sample data is: dividing the length of the collected original epileptic EEG signal and the number of channels of the original epileptic EEG signal, expanding the amount of initial sample data, and combining them into multiple groups of sample data of different sizes.

[0014] Furthermore, the division ratios of the training set and the test set are: 5:5, 6:4, 7:3, 8:2, 9:1.

[0015] Furthermore, in step 2, the process of obtaining the optimal ODE solution is:

[0016] 1) Randomly select one set from multiple training sets and test sets as the training set and test set for the K series model neuron equation ODE solution;

[0017] 2) Each ODE solution is solved three times based on the selected training set and test set, and the calculation time and recognition results of the three solutions are obtained, and the average value is taken as the experimental result;

[0018] 3) Analyze the experimental results of different ODE solutions, and select the ODE solution with high computational efficiency as the optimal ODE solution while ensuring accuracy.

[0019] Furthermore, in the step 2, the optimal ODE solution is the forward Euler method (FEuler).

[0020] Furthermore, in step 3, the process of obtaining the optimal data parallel method is:

[0021] 1) Randomly select one set from multiple training sets and test sets as the training set and test set for parallel computing;

[0022] 2) Each data parallel method is solved based on the selected training set and test set to obtain the parallel computing performance evaluation index of the solution;

[0023] 3) Repeat steps 1) and 2) three times, and take the average of the three performance evaluation indicators of the same data parallel method as the experimental result;

[0024] 4) Analyze the experimental results of different data parallel methods, and select the optimal data parallel method based on the evaluation indicators of parallel computing while ensuring accuracy.

[0025] Furthermore, when performing parallel computing, the "shuffle partitioning" method is used to divide the data according to the number of parallel cores.

[0026] Furthermore, the evaluation indicators are acceleration ratio and parallel efficiency.

[0027] Furthermore, in step three, the optimal data parallel mode is Spark's parallel computing mode.

[0028] The beneficial effects of the present invention are:

[0029] 1. This application starts with the K series model and its neuron equations, and makes full use of the existing ODE solution research results on the epileptic EEG data set, analyzes and compares the computational efficiency and recognition results of different ODE solutions in the model neuron equations, and screens out the optimal ODE solution. It optimizes the computational efficiency while ensuring the accuracy. At the same time, it also conducts parallel computing research based on different parallel computing methods to obtain the optimal parallel computing method, which accelerates the K series model and further improves the computational efficiency of the model. At the same time, it ensures a good recognition effect of the model, and solves the problem that the prior art pays more attention to the model recognition rate, resulting in low model computational efficiency.

[0030] 2. In this application, the initial sample data is expanded and processed to enrich the training and test data, and different division methods are used to divide the training set and the test set to form multiple groups of training sets and test sets to ensure the difference of the data. At the same time, the training set and the test set are randomly selected to reduce the interference of human factors and improve the test accuracy. In particular, through multiple tests, the average value is used as the experimental value to reduce the errors caused by the system and calculation, thereby further improving the test accuracy.

[0031] 3. This application adopts a data-parallel parallel computing method. In this way, data and calculations under the same global data are divided into different processes, thereby reducing the computing and storage pressure on a single core (device). At the same time, the data is "scrambled and divided", the data set is randomly shuffled, and then the shuffled data is divided into a corresponding number of small parts according to the number of CPU threads or processes and distributed to the number of threads or processes. Its computational complexity is much smaller than that of global random sampling, which reduces the processing power occupied and makes more full use of samples, indirectly improving the running speed.

[0032] In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only fifteen of the drawings in the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0034] Figure 1 It is a structural schematic diagram of the present invention;

[0035] Figure 2 Clustered bar charts of calculation time for models based on different ODE solutions;

[0036] Figure 3 Clustered bar charts of model identification results based on different ODE solutions;

[0037] Figure 4 It is a line chart of the calculation time of the K series model based on multi-threading with different numbers of threads;

[0038] Figure 5 is the speedup ratio S of the K series model based on multithreading with different numbers of threads p Line chart;

[0039] Figure 6 is the parallel efficiency E of the K series model based on multithreading with different numbers of threads p Line chart;

[0040] Figure 7 This is a line chart of the recognition results of the K series model based on multi-threading with different numbers of threads;

[0041] Figure 8 It is a line chart of the calculation time of the K series model based on multiple processes with different numbers of processes;

[0042] Fig. 9 is the speedup ratio S of the K series model based on multi-process with different numbers of processes p Line chart;

[0043] Fig.10 is the parallel efficiency E of the K series model based on multi-process with different numbers of processes p Line chart;

[0044] Fig.11 It is a line chart of the recognition results of the K series model based on multiple processes with different numbers of processes;

[0045] Fig.12 This is a line chart of the calculation time of the K series model based on Spark with different numbers of partitions;

[0046] Fig.13 is the speedup ratio S of the K series model based on Spark with different numbers of partitions p Line chart;

[0047] Fig.14 is the parallel efficiency E of the K series model based on Spark with different numbers of partitions p Line chart;

[0048] Fig.15 This is a line chart of the recognition results of the Spark-based K series model with different numbers of partitions. DETAILED DESCRIPTION

[0049] Embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Although certain embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as being limited to the embodiments described herein, which are instead provided for a more thorough and complete understanding of the present invention. It should be understood that the drawings and embodiments of the present invention are only for exemplary purposes and are not intended to limit the scope of protection of the present invention.

[0050] Example

[0051] like Figure 1-15 As shown, a parallel computing method for K series models based on ODE solution is provided, which is characterized by:

[0052] Step S1, preprocessing the acquired initial sample data to form multiple groups of sample data, and dividing each group of sample data into multiple training sets and test sets according to different proportions;

[0053] Step S2, selecting a training set and a test set, training and testing different ODE solutions for the K series model neuron equation, and selecting an ODE solution with high computational efficiency as the optimal ODE solution while ensuring accuracy;

[0054] Step S3, solving the neuron equation of the K series model according to the optimal ODE solution obtained in step 2, and selecting training set and test set data, using different data parallel methods for parallel computing, and selecting the optimal data parallel method according to the parallel computing evaluation index while ensuring the accuracy;

[0055] Step S4: Calculate the K series model neuron equations and apply the K series model according to the obtained optimal ODE solution and optimal data parallel method.

[0056] In step S1, the epilepsy EEG data forming the initial sample data is the EEG data collected by the Epilepsy Research Laboratory of the University of Bonn, Germany. The epilepsy EEG data set consists of 5 subsets of data sets, which are marked as Z, O, N, F and S. Each subset contains 100 single-channel EEG signals with a length of 4097, a duration of 23.6 seconds and a sampling frequency of 173.61 Hz. Subsets Z and O were collected from a control group consisting of 5 healthy people.

[0057] The process of preprocessing the initial sample data is: dividing the length of the collected original epileptic EEG signal and the number of channels of the original epileptic EEG signal, expanding the amount of initial sample data, and combining them into multiple groups of sample data of different sizes.

[0058] Specifically, the length of the original epileptic EEG signal is divided into 30, 40, 50 and 60 segments, and the corresponding sub-sample lengths are 136, 102, 81 and 68. The number of channels of the original epileptic EEG signal is divided into 2 segments, and the number of channels of the corresponding sub-sample is 50. Therefore, there are 4 sample data sizes through combination, namely: 50*136, 50*102, 50*81 and 50*68.

[0059] The division ratios of the training set and the test set are: 5:5, 6:4, 7:3, 8:2, and 9:1.

[0060] In step S2, the present application selects the K series model and olfactory neuron equation proposed by Professor Freeman as research objects. All neurons in the K series model (each neuron represents a neural cluster or cell tissue) can be described by a unified second-order differential equation, as shown in formulas (1) and (2).

[0061]

[0062]

[0063] The process of obtaining the optimal ODE solution is as follows:

[0064] 1) Randomly select one set from multiple training sets and test sets as the training set and test set for the K series model neuron equation ODE solution;

[0065] 2) Each ODE solution is solved three times based on the selected training set and test set, and the calculation time and recognition results of the three solutions are obtained, and the average value is taken as the experimental result;

[0066] 3) Analyze the experimental results of different ODE solutions, and select the ODE solution with high computational efficiency as the optimal ODE solution while ensuring accuracy.

[0067] Experimental environment,The experiment was conducted on the MATLAB R2018b platform, and the model and RK4 solution algorithm were,programmed and implemented. The operating system is Windows 10 professional, the processor is AMD Ryzen 5 2600X Six-Core Processor 3.60GHz, the memory is 16GB DDR4, and the hard disk is 3TB.

[0068] Among them, the solutions of the neuron equations in the experiment include forward Euler (FEuler), backward Euler (BEuler), trapezoid Euler (TraEuler), two-step Euler (TwoEuler), improved Euler (IEuler), 2nd-order Runge-Kutta explicit method (RK2), 3rd-order Runge-Kutta explicit method (RK3), Admas 2nd-order implicit method (Admas2y), Admas 3rd-order implicit method (Admas3y), Admas 4th-order implicit method (Admas4y), etc. Each ODE solution is not described here. For different ODE solutions, the calculation time (seconds / s) and recognition results (accuracy / %) of the model are analyzed. In the experiment, 3 experiments are conducted for each case, and then the average value is taken; the selected training set: test set = 8:2, and the experimental results are shown in Tables 1 and 2.

[0069] Table 1 Model calculation time based on different ODE solutions (seconds / s)

[0070]

[0071]

[0072] Table 2 Model identification results based on different ODE solutions (%)

[0073]

[0074] In order to more intuitively display the experimental results in Tables 1 and 2, Figure 2 and 3 They represent the model calculation time and recognition results based on different ODE solutions on the epileptic EEG dataset. Figure 2 and 3 In the figure, the horizontal axis represents different ODE solutions of the model neuron equation, and the vertical axis represents the calculation time and recognition results of the model based on different ODE solutions.

[0075] By analyzing and comparing Table 1 and Table 2 and Figure 2 and Figure 3 It can be concluded that by analyzing the computational performance of the model based on different ODE solutions, compared with the RK4 solution, the FEuler solution is the most suitable for solving the model neuron equation. This is because the FEuler solution has the shortest average computational time of the model, which is only 7205s, or about 2 hours, while ensuring the recognition accuracy.

[0076] In step S3, the process of obtaining the optimal data parallel mode is:

[0077] 1) Randomly select one set from multiple training sets and test sets as the training set and test set for parallel computing;

[0078] 2) Each data parallel method is solved based on the selected training set and test set to obtain the parallel computing performance evaluation index of the solution;

[0079] 3) Repeat steps 1) and 2) three times, and take the average of the three performance evaluation indicators of the same data parallel method as the experimental result;

[0080] 4) Analyze the experimental results of different data parallel methods, and select the optimal data parallel method based on the evaluation indicators of parallel computing while ensuring accuracy; the evaluation indicators are acceleration ratio and parallel efficiency, and the parallel computing performance of the model is evaluated by these two indicators.

[0081] The expression of the speed-up ratio is shown in formula (3):

[0082]

[0083] Where T(s) is the time of the serial algorithm, and T(p) is the time of the parallel algorithm with p processors. Both solve the same problem, and T(p) ≥ T(s) / p.

[0084] The expression of parallel efficiency is shown in formula (4):

[0085]

[0086] E p It represents the parallel computing efficiency of calling p processors, which indicates the utilization of the processors.

[0087] When performing parallel computing, the "scrambled partitioning" method is used to partition data according to the number of parallel cores; the "scrambled partitioning" method randomly shuffles the data set, and then divides the shuffled data into a corresponding number of small portions according to the number of CPU threads or processes, and distributes them to the number of threads or processes. Among them, there is a special data partitioning method for balanced grouping, which is to divide the data into corresponding small portions according to the hardware configuration, and then randomly distribute these small portions of data to each number of threads or processes.

[0088] Experimental environment,The hardware environment of the experiment is the operating system of Windows 10 professional, the processor is AMD Ryzen 5 2600X Six-Core Processor 3.60GHz, the memory is 16GB DDR4, the hard disk is 3TB, the number of cores is 6, and the number of logical processors is 12. The experimental software environment based on multi-threading and multi-process is carried out on the PyCharm 2020.1 platform, and Python3.6 is selected as the development language; the Spark parallel framework selects Spark's local running mode.

[0089] Among them, the commonly used parallel computing methods are multi-threaded and multi-process parallel computing. This application adds the Spark parallel framework method on this basis, and conducts parallel computing research on the K series model and olfactory neuron equation proposed by Professor Freeman, uses the FEuler solution to solve the model neuron equation, and discusses and analyzes the experimental results of each method.

[0090] The selected training set: test set = 8:2, the training set and the test set are input into the K series model, the experiment is conducted 3 times on different sample data, and the average result of the 3 tests is taken as the experimental result.

[0091] The experimental results of parallel computing based on multi-threading are shown in Table 3, Table 4, Table 5 and Table 6.

[0092] Table 3 Calculation time of K series models with different numbers of threads (seconds / s)

[0093]

[0094]

[0095] Table 4 Speedup ratio S of K series models at different thread numbers p

[0096]

[0097] Table 5 Parallel efficiency E of K series models at different thread numbers p (%)

[0098]

[0099]

[0100] Table 6 Recognition results of K series models at different thread numbers (%)

[0101]

[0102] In order to more intuitively display the experimental results in Tables 3 to 6, Figures 4 to 7 They represent the computation time and speedup ratio S of the K series model with different numbers of threads on the epilepsy EEG dataset. p , parallel efficiency E p And recognition accuracy. Figures 4 to 7 In the figure, the horizontal axis represents the number of threads called for parallel computing, and the vertical axis represents the computing time of the bionic model with different numbers of threads, S p Value, E of different number of threads p Values ​​and recognition accuracy of different thread numbers.

[0103] By analyzing Tables 3 to 6 and Figures 4 to 7 It can be seen that as the number of threads increases, the calculation time of the K series model does not decrease, but even increases. The parallel efficiency E p The value decreases gradually, with a good recognition accuracy and an average recognition rate of more than 97%. That is, parallel computing based on multi-threading does not improve the computing efficiency of the bionic model.

[0104] The experimental results of parallel computing based on multi-process are shown in Table 7, Table 8, Table 9 and Table 10.

[0105] Table 7 Calculation time of K series models with different number of processes (seconds / s)

[0106]

[0107] Table 8 Speedup ratio S of K series models at different process numbers p

[0108]

[0109]

[0110] Table 9 Parallel efficiency E of K series models at different process numbers p (%)

[0111]

[0112] Table 10 Recognition results of K series models with different numbers of processes (accuracy / %)

[0113]

[0114]

[0115] In order to more intuitively display the experimental results in Tables 7 to 10, Figures 8 to 11 They represent the computation time and speedup ratio S of the K series model with different numbers of processes on the epilepsy EEG dataset. p, parallel efficiency E p And recognition accuracy. Figures 8 to 11 In the figure, the horizontal axis represents the number of different processes called by parallel computing, and the vertical axis represents the computing time of the bionic model with different numbers of processes. p Value, E at different process numbers p Value,recognition accuracy at different numbers of processes.

[0116] By analyzing Tables 7 to 10 and Figures 8 to 11 It can be seen that as the number of processes increases, the calculation time of the K series model gradually decreases, and the speedup ratio S p The value gradually increases, and the parallel efficiency E p The value decreases gradually, with a good recognition accuracy and an average recognition rate of more than 94%. That is, while ensuring the recognition result, the parallel computing based on multi-process improves the computing efficiency of the bionic model.

[0117] The experimental results of parallel computing based on Spark are shown in Tables 11, 12, 13 and 14.

[0118] Table 11 Calculation time of K series model with different number of partitions (seconds / s)

[0119]

[0120]

[0121] Table 12K series model speedup S in different partition numbers p

[0122]

[0123] Table 13 Parallel efficiency E of K series models with different numbers of partitions p (%)

[0124]

[0125]

[0126] Table 14 Recognition results of K series models with different numbers of partitions (accuracy / %)

[0127]

[0128] In order to more intuitively display the experimental results in Tables 11 to 14, Figures 12 to 15 They represent the computation time and acceleration ratio S of the K series model with different numbers of data partitions on the epileptic EEG dataset. p , parallel efficiency E p And recognition accuracy. Figures 12 to 15In the figure, the horizontal axis represents the number of data partitions in parallel computing, and the vertical axis represents the computing time of the bionic model on different numbers of partitions, the S p Value, E at different partition numbers p Value, recognition accuracy on different numbers of partitions.

[0129] By analyzing Tables 11 to 14, we can see that as the number of data partitions increases, the calculation time of the K series model gradually decreases, and the speedup ratio S p The value gradually increases, and the parallel efficiency E p The value gradually decreases, with a better recognition accuracy and an average recognition rate of more than 97%. That is, while ensuring the recognition results, the parallel computing based on Spark improves the computing efficiency of the K series model.

[0130] In summary, Spark-based parallel computing has the greatest improvement in computing efficiency for the K series model and has the highest parallel efficiency. However, in practice it is impossible to call all hardware resources for parallel computing, which affects the speedup ratio S p Value and parallel efficiency E p The values ​​are required, and the appropriate number of partitions can be selected according to the experimental results, which can ensure both the acceleration ratio of the K series model and the parallel efficiency.

[0131] The present application can also apply the obtained optimal ODE solution and optimal data parallel method to other K series models and models modified and improved by the K series models to improve operating efficiency, and is not limited to the K series model and olfactory neuron equation proposed by Professor Freeman described in this application.

[0132] The above description of the various embodiments tends to emphasize the differences between the various embodiments. The same or similar aspects can be referenced to each other, and for the sake of brevity, this application will not go into details.

[0133] Note that the above are only preferred embodiments of the present invention and the technical principles used. Those skilled in the art will understand that the present invention is not limited to the specific embodiments herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in more detail through the above embodiments, the present invention is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. A parallel computing method for K series models based on ODE solution, characterized by: Step 1: pre-process the obtained initial sample data to form multiple groups of sample data, and divide each group of sample data into multiple training sets and test sets according to different proportions; Step 2: Select training sets and test sets, train and test different ODE solutions for the K series model neuron equations, and select the ODE solution with high computational efficiency as the optimal ODE solution while ensuring accuracy; Step 3: Solve the neuron equation of the K series model according to the optimal ODE solution obtained in step 2, select the training set and test set data, use different data parallel methods for parallel computing, and select the optimal data parallel method according to the parallel computing evaluation index while ensuring the accuracy; Step 4: Calculate the K series model neuron equations and apply the K series model based on the obtained optimal ODE solution and optimal data parallel method.

2. The K series model parallel computing method based on ODE solution according to claim 1, characterized in that: In step 1, the process of preprocessing the initial sample data is: dividing the length of the collected original epileptic EEG signal and the number of channels of the original epileptic EEG signal, expanding the amount of initial sample data, and combining them into multiple groups of sample data of different sizes.

3. The K series model parallel computing method based on ODE solution according to claim 2, characterized in that: The division ratios of the training set and the test set are: 5:5, 6:4, 7:3, 8:2, and 9:

1.

4. The K series model parallel computing method based on ODE solution according to claim 1, characterized in that: In step 2, the process of obtaining the optimal ODE solution is: 1) Randomly select one set from multiple training sets and test sets as the training set and test set for the K series model neuron equation ODE solution; 2) Each ODE solution is solved three times based on the selected training set and test set, and the calculation time and recognition results of the three solutions are obtained, and the average value is taken as the experimental result; 3) Analyze the experimental results of different ODE solutions, and select the ODE solution with high computational efficiency as the optimal ODE solution while ensuring accuracy.

5. The K series model parallel computing method based on ODE solution according to claim 1, characterized in that: In the step 2, the optimal ODE solution is the forward Euler method (FEuler).

6. The K series model parallel computing method based on ODE solution according to claim 1, characterized in that: In step 3, the process of obtaining the optimal data parallel method is as follows: 1) Randomly select one set from multiple training sets and test sets as the training set and test set for parallel computing; 2) Each data parallel method is solved based on the selected training set and test set to obtain the parallel computing performance evaluation index of the solution; 3) Repeat steps 1) and 2) three times, and take the average of the three performance evaluation indicators of the same data parallel method as the experimental result; 4) Analyze the experimental results of different data parallel methods, and select the optimal data parallel method based on the evaluation indicators of parallel computing while ensuring accuracy.

7. The K series model parallel computing method based on ODE solution according to claim 6, characterized in that: When performing parallel computing, the "shuffle partitioning" method is used to divide data according to the number of parallel cores.

8. The K series model parallel computing method based on ODE solution according to claim 6, characterized in that: The evaluation indicators are speedup ratio and parallel efficiency.

9. The K series model parallel computing method based on ODE solution according to claim 1, characterized in that: In step three, the optimal data parallel mode is Spark's parallel computing mode.

Images