KR102451231B1 - A simulator for RBF-based networks in neuromorphic chip - Google Patents

Abstract

A GUI-based RBF algorithm simulator for simulating an RBF algorithm operated in a neuromorphic chip according to an embodiment of the present invention, comprising: a simulation module for simulating the RBF algorithm; and a GUI module for inputting options for manipulating the simulation module and outputting a simulation result.

Description

A simulator for RBF-based networks in neuromorphic chip

The present invention relates to a simulator for learning RBF neural networks operated in a neuromorphic chip.

In the era of big data, there is a mixture of atypical texts, images, voices, and videos that are difficult for machines to easily recognize among a large amount of data. However, although the Von Neumann-based computer structure, which is a traditional computing method, is excellent at executing numerical calculations or precisely written programs, it has great limitations in terms of high-level cognitive applications for image and sound processing, processing speed, and power consumption. looks like

Therefore, neuromorphic computing technology has been developed in an attempt to overcome these limitations. Neuromorphic computing was first proposed by Carver Mead and was created by mimicking the main function of the human brain, where large amounts of memory and computation are simultaneously operated, so it has a faster computation speed, is more efficient in terms of power, and requires a smaller footprint. do. In other words, neuromorphic computing is expected to be the most suitable platform for implementing algorithms that go beyond the limits of existing hardware in the field of machine learning.

Neuromorphic computing research has been actively conducted recently. Among them, the CM1K neuromorphic chip is a hardware based on a Radial Basis Function (RBF) neural network and provides fast pattern recognition in embedded systems. The RBF neural network is a classifier used for modeling pattern recognition systems in various fields due to its superiority in learning ability and generalization ability.

Existing studies using the CM1K neuromorphic chip have trained and used RBF neural networks. As a basic method of learning the RBF network of the neuromorphic chip, there is a method of directly learning it in hardware. However, since this intuitive learning method has limitations in that it takes a lot of time and money for learning and testing, a method of verifying and learning the algorithm using a simulator is used as a method to compensate for this.

That is, if the RBF network is simulated and verified using a simulator and then trained, it is possible to provide fast algorithm verification, thereby efficiently evaluating the performance of the algorithm.

However, the existing simulator capable of such a method in the neuromorphic chip does not have a GUI module (Graphic Unser Interface) or, even if a GUI module is implemented, does not have a function to control various performance evaluation options.

The embodiment of the present invention improves the suitability for multiple input data than the existing RBF algorithm through the RBF algorithm of the double layer structure, and improves the user's convenience by facilitating the adjustment of performance evaluation options through the provision of a GUI module (Graphic User Interface) It aims to provide a simulator for learning RBF neural networks operated on a neuromorphic chip that can be improved.

On the other hand, the technical problems to be achieved in the present invention are not limited to the technical problems mentioned above, and other technical problems not mentioned are clearly to those of ordinary skill in the art to which the present invention belongs from the description below. can be understood

A GUI-based RBF algorithm simulator for simulating an RBF algorithm operated in a neuromorphic chip according to an embodiment of the present invention, comprising: a simulation module for simulating the RBF algorithm; and a GUI module for inputting options for manipulating the simulation module and outputting a simulation result.

The GUI module includes: a state option box for adjusting the state of the simulation module; a learning option box for adjusting options required for learning on the simulation module; a classification option box for adjusting options required for classification on the simulation module; an integrated graph box displaying the number of neurons and data used in the simulation; a learning result graph box displaying the learning result as a graph; and a classification result graph box displaying the classification result as a graph.

The state option box may include: a distance calculation method control unit for selecting a distance calculation method between an input vector and a neuron; a chip number selection unit that selects the number of neuromorphic chips used for simulation; and a learning information delete button for erasing the learned neuron information, the name of the data used for simulation, the number of data samples, the total number of neurons according to the number of the neuromorphic chips, and the number of neurons generated after learning can be displayed.

The learning option box may include: a learning number adjusting unit for adjusting the number of learning repetitions; a dual-layer learning algorithm selection unit for learning through a dual-layer structure RBF algorithm; a whole data recording/non-recording unit to record all information of the learning data in the neurons; an influence field value selection unit for setting a maximum value and a minimum value of an influence field generated by neurons; and a learning start button instructing to start learning.

The classification option box may include: a classification algorithm selection unit for selecting the simulation module to classify by using any one of an RBF algorithm and a KNN algorithm; a classification policy selection unit for selecting a classification policy; and a classification start button instructing to start classification.

The integrated graph box may display the number of input data, the number of generated neurons, and the number of classified data.

The learning result graph box may display the number of neurons generated for each category and the number of neurons generated for each number of repeated learning.

The classification result graph box may display the number of status codes classified by category and the total amount of status codes.

The classification policy includes: Best Match for selecting the category of the first firing neuron as the category of the input vector; Dominant selecting the category with the largest number of categories of firing neurons as the category of the input vector; Unanimity that determines the category of the input vector when all categories of firing neurons are the same; and Min consensus for determining the corresponding category as the category of the input vector only when the number of the same categories among the categories of the firing neurons is equal to or greater than the preset number.

The dual layer structure RBF algorithm includes: a first layer for deriving a first result value by recognizing a characteristic pattern using individual data as input data; and a second layer that integrates the first result value through a concatenation function and uses it as an input value.

The first result value may include: a distance value between the input vector and a firing neuron; and a category that is a prediction result for each input vector.

The simulator for learning the RBF neural network operated in the neuromorphic chip according to the embodiment of the present invention improves the suitability for multiple input data than the existing RBF algorithm through the RBF algorithm of the double layer structure, and the GUI module (Graphic User Interface) ), it is possible to improve user convenience by facilitating adjustment of performance evaluation options.

On the other hand, the effects obtainable in the present invention are not limited to the above-mentioned effects, and other effects not mentioned will be clearly understood by those of ordinary skill in the art to which the present invention belongs from the description below. will be able

1 is a view showing the appearance of a GUI module (Graphic User Interface) module according to an embodiment of the present invention.
FIG. 2 is a view showing only a status option box, a RBF Learning Option box, and a Classification Option box provided on the left side of the GUI module of FIG. 1 .
FIG. 3 is a view showing only the distribution per category box provided on the upper right side of the GUI module of FIG. 1 .
FIG. 4 is a view showing only the Learning results box provided at the lower right of the GUI module of FIG. 1 .
FIG. 5 is a view showing only the classification results box provided at the lower right side of the GUI module of FIG. 1 .
FIG. 6 is a view showing only a log box provided at the bottom of the GUI module of FIG. 1 .
FIG. 7 is a screen displayed when a learning result table (Knowledge content) in which information of a learned neuron can be viewed in detail in the view of FIG. 1 is selected.
FIG. 8 is a screen displayed when a classification result table (Dataset in detail) in which a classification result can be viewed in detail is selected in the view of FIG. 1 .
9 is a diagram illustrating an operation process of a simulator according to an embodiment of the present invention.
FIG. 10 is a diagram illustrating a process in which a decision space is defined to distinguish a category of an input vector, and then neurons in the decision space selectively store information of the input vector to generate an influence field.
11 is a diagram showing (a) Norm L1 calculation formula and graph, (b) Norm LSup calculation formula and graph, (c) Norm L1-based neuron influence field example, and (d) Norm LSup-based neuron influence field example .
12 is a diagram illustrating a process of classifying neurons generated in a decision space created after learning in response to an input vector entering an influence field.
13 is a diagram schematically illustrating influence field mapping according to the RCE/RBF classification method and the KNN classification method.
14 is a diagram showing the structure of an existing multi-neural RBF network.
15 is a diagram showing the structure of a dual layer structure RBF algorithm.
16 is a graph showing the pattern change according to the drowsiness state of four types of simulated data.
17 is a graph showing the accuracy (%) for 20 simulated driver data.
18 is a graph showing the number (dogs) of neurons used in the experiment.

Other advantages and features of the present invention, and a method of achieving them will become apparent with reference to the embodiments described below in detail in conjunction with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but may be implemented in various different forms, and only this embodiment serves to complete the disclosure of the present invention, and to obtain common knowledge in the technical field to which the present invention pertains. It is provided to fully inform the possessor of the scope of the invention, and the present invention is only defined by the scope of the claims.

Even if not defined, all terms (including technical or scientific terms) used herein have the same meaning as commonly accepted by common technology in the prior art to which this invention belongs. Terms defined by general dictionaries may be interpreted as having the same meaning as in the related description and/or in the text of the present application, and shall not be interpreted conceptually or excessively formally, even if not expressly defined herein. won't

The terminology used herein is for the purpose of describing the embodiments and is not intended to limit the present invention. In this specification, the singular also includes the plural, unless specifically stated otherwise in the phrase. As used in the specification, 'comprise' and/or the various conjugations of this verb, eg, 'comprising', 'comprising', 'comprising', 'comprising', etc., refer to the stated composition, ingredient, component, Steps, acts and/or elements do not exclude the presence or addition of one or more other compositions, components, components, steps, acts and/or elements. As used herein, the term 'and/or' refers to each of the listed components or various combinations thereof.

Meanwhile, terms such as '~ unit', '~ group', '~ block', and '~ module' used throughout this specification may mean a unit that processes at least one function or operation. For example, it can mean software, hardware components such as FPGAs or ASICs. However, '~ part', '~ group', '~ block', and '~ module' are not meant to be limited to software or hardware. '~ unit', '~ group', '~ block', and '~ module' may be configured to be in an addressable storage medium or configured to regenerate one or more processors.

Accordingly, as an example, '~ part', '~ group', '~ block', and '~ module' are components such as software components, object-oriented software components, class components, and task components. fields, processes, functions, properties, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuitry, data, databases, data structures, tables, arrays and include variables. The functions provided within the components and '~part', '~gi', '~block', and '~module' are smaller than the number of components and '~bu', '~gi', '~block' ', '~modules' or may be further separated into additional components and '~parts', '~groups', '~blocks', and '~modules'.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings of the present specification.

The simulator 10 for learning the RBF (Radial Basis Function) neural network operated in the neuromorphic chip of the present invention first verifies the algorithm to be trained in the RBF neural network of the neuromorphic chip through the simulator 10 on the computer. By doing so, it is possible to reduce the waste of time and money that occurs when the algorithm is directly trained on the neuromorphic chip and verified without such a simulation process.

In addition, in this process, the dual layer structure RBF algorithm 20 is additionally provided to improve the suitability for multiple input data as well as provide a GUI module (Graphic User Interface, 100) that can improve user convenience. This makes it easier to adjust options and check results in the learning and classification process.

Hereinafter, each configuration of the simulator 10 for learning the RBF neural network operated in the neuromorphic chip and the dual-layer structure RBF algorithm that can be used in the simulation process will be described in more detail.

A GUI module-based RBF algorithm simulator 10 for simulating an RBF algorithm that can be operated in a neuromorphic chip according to an embodiment of the present invention operates a simulation module 200 that simulates the RBF algorithm and the simulation module. and a GUI module 100 for inputting options for doing and outputting simulation results.

1 is a diagram illustrating a graphical user interface (GUI) 100 according to an embodiment of the present invention.

Referring to FIG. 1 , the GUI module 100 is largely divided into both sides around the center. In the left area, you can adjust the options related to learning and classification, including the status option box (Status box, 110), the learning option box (RBF Learning Option box, 120), and the Classification Option box (130). this is provided

In addition, in the right area, the results according to learning and classification, including the distribution per category box (140), the learning results box (151) and the classification results box (152), and the An area is provided in which the following graph can be checked.

Additionally, a log box 160 in which information on selected options is displayed is provided in the lower area of the GUI module 100 .

As such, the GUI module 100 of the present invention improves readability by grouping each function into one box according to the purpose of each configuration.

2 shows only a status option box (Status box, 110), a learning option box (RBF Learning Option box, 120) and a classification option box (130) provided on the left side of the GUI module 100 of FIG. It is a drawing.

Referring to FIG. 2 , a status option box 110 for adjusting the status of the simulation module is provided at the top of the GUI module.

In the status option box (Status box, 110), the name of the data used for the simulation (Data Filename, 114) and the number of data samples (#_samples, 115) are displayed. In addition, in the platform area, a distance calculation method control unit 111 for selecting a distance calculation method between an input vector and a neuron, a chip number selection unit (SimuNMnK, 112) for selecting the number of neuromorphic chips used for simulation, and learning It includes a learning information delete button (Forget All, 113) that erases the information of the old neurons, and the total number of neurons (#_neurons, 116) according to the number of neuromorphic chips and the number of neurons generated after learning (#_Used neurons) , 117) appears.

In more detail, a distance calculation method control unit that can select an RBF distance calculation method is provided in the Platform area. When input vectors are input for learning of the simulator, each neuron included in the simulator generates an influence field by calculating the distance between the input vector and the center of the neuron. There are two methods, Norm L1 and Norm LSup, for calculating the distance between the input vector and the neuron when generating such an influence field, and a detailed description of each method will be described later.

The chip number selection unit SimuNMnK 112 is used to adjust the number of neuromorphic chips used for learning and classification. In the simulator provided for algorithm verification in the prior art, the algorithm was verified using one neuromorphic chip, but in the present invention, the algorithm is learned using a plurality of neuromorphic chips through the chip number selection unit (SimuNMnK, 112). and categorization.

In addition, it provides a function to erase the information of the learned neurons through the learning information delete button (Forget All, 113).

Referring back to FIG. 2 , a learning option box (RBF Learning Option box, 120) is provided at the bottom of the GUI module 100 to adjust options necessary for learning on the simulator.

In the learning option box (RBF Learning Option box, 120), there is a learning number control unit (Iterative, 121) that adjusts the number of repetitions of learning, and a dual-layer learning algorithm selection unit (T) to learn through the double-layer structure RBF algorithm (20). -RBF, 122), the entire data recording selection unit (Write all samples, 123) that selects whether to record all information of the training data to the neuron, and the maximum (MaxIF) and minimum values (MinIF) of the influence field generated by the neuron A setting influence field value selection unit 124 and a learning start button (Learn, 125) for instructing to start learning with the set options are included.

In more detail, the number of learning control unit 121 that adjusts the number of repetitions of learning is an option for adjusting the number of repetitions, and through this, it is possible to adjust the maximum number of repetitions of learning. The number of iterations can be adjusted by adjusting Max_Iter.

The double-layer learning algorithm selection unit (T-RBF, 122) may be selected when learning is performed using the double-layer structure RBF algorithm 20, and a detailed description of the double-layer structure RBF algorithm 20 will be described later. .

The write all samples 123 for selecting whether to record all information of the learning data into the neurons is an algorithm for inputting all the data used for learning into the neurons. However, this option is not used as a learning algorithm.

Referring back to FIG. 2 , a classification option box 130 is provided at the bottom of the GUI module 100 to adjust options required for classification on the simulator.

In the Classification Option box 130, a classification algorithm selection unit 131 for selecting the simulation module to classify using any one of the RBF algorithm and the KNN algorithm, and a classification policy selection unit for selecting a classification policy (Category Out) , 132) and a classification start button (Classify, 133) for instructing to start classification.

More specifically, the selection of the algorithm used for classification can be one of the RBF algorithm and the KNN algorithm. At this time, KNN is used only when data is input to neurons through the Write all samples algorithm in the learning option box, and the K option control unit that can set the maximum number of firing neurons to be considered for classification is activated only when KNN is selected.

The classification policy selector (Category Out, 132) is an area in which a classification policy can be selected, and any one of Best Match, Dominant, Unanimity, or Min consensus can be selected.

In the case of Best Match, the category of the closest firing neuron, that is, the first firing neuron, is selected as the category of the input vector. At this time, if there is no firing neuron, the classification result of Unknown is selected, and if there is at least one firing neuron, the category of the first firing neuron storing the shortest distance is selected as the classification result. Unknown means the case where the input vector is not recognized by the neuron, that is, the case where classification is not performed. If the categories of firing neurons all match, it is classified as Identified, and if even one of the categories is different, it is classified as Uncertain.

In the case of the dominant, the category with the largest number of categories of firing neurons is selected as the category of the input vector. That is, it is a method of determining the number of categories by looking at the number of categories of firing neurons. For example, if the classification method is selected as Dominant and three firing neurons are generated, the first firing neuron category is 1, the second firing neuron category is 2, and the third firing neuron category is also 2, then the Dominant method is input. The category of the vector is determined as 2.

In the case of unanimity, it is a unanimous method, and if all categories of firing neurons are the same, it is determined as the category of the input vector. If even one is different, the classification result is determined as Unknown.

Min consensus is a method similar to Dominant, and basically follows the majority voting method, but only when there are at least N categories of firing neurons, it is determined as the category of the input vector. In this case, the value of N can be set by the user, and the default value is set to 2. If the number of identical categories among the categories of firing neurons is less than N, the input vector is classified as Unknown.

FIG. 3 is a view showing only the distribution per category box 140 provided at the upper right of the GUI module 100 of FIG. 1 .

Referring to FIG. 3 , a main graph is displayed in a distribution per category box 140 . In the main graph, three types of bar graphs are displayed for each category. From the left, the number of input data (Input), the number of generated neurons (Used neurons), and the number of classified data (Output) are indicated.

4 is a view showing only the learning results graph box (Learning results box, 151) provided on the lower right side of the GUI module 100 of FIG. 1,

FIG. 5 is a view showing only the classification results box 152 provided at the lower right of the GUI module 100 of FIG. 1 .

Referring to FIGS. 4 and 5 , a bar graph indicating the number of neurons generated for each category and a line graph indicating the neurons generated for each repetition number are displayed in a learning results box 151 .

In addition, a bar graph indicating the number of status codes (UNK, ID, UNC) classified by category and a pie graph indicating the total amount of status codes are displayed in the classification results box 152 .

In this case, the Status code is displayed as one of UNK (unknown), ID (identified), UNC_c (uncertain_correct), and UNC_i (uncertain_incorrect). If there is no firing neuron, it is classified as UNK, if all firing neurons have the same category, it is ID, and if even one category of firing neurons is different, it is divided into UNC. UNK does not exist in KNN mode.

FIG. 6 is a view showing only the log box 160 provided at the bottom of the GUI module 100 of FIG. 1 .

Referring to FIG. 6 , a log box 160 located at the bottom of the simulator is composed of text, user input and modification are impossible, and information and data information of selected options are recorded.

Referring back to FIG. 1 , a File button and a View button are provided in the upper left menu of the GUI module 100 .

In File, you can select Load dataset that calls input data and Export dataset, which is a function that can save classification results as an Excel file. A viewable classification result table (Dataset in detail) can be selected.

FIG. 7 is a screen displayed when a learning result table (Knowledge content) in which information of a learned neuron can be viewed in detail in the view of FIG. 1 is selected.

Referring to FIG. 7 , the simulator of the present invention provides a learning result table (Knowledge content) that allows the detailed understanding of the learning results of neurons that are difficult to check only through graphs.

The learning result table refers to a table in which detailed information of neurons can be checked, and the simulator creates a table in which detailed information of neurons is recorded after the network performs learning or after loading Knowledge, which is a set of generated neurons. The learning result table includes a neuron number (NeuronID), a context (Context), a category (Category), a distance value of an influence field (InfluenceField), and a minimum value of an influence field (MinIF).

FIG. 8 is a screen displayed when a classification result table (Dataset in detail) in which a classification result can be viewed in detail is selected in the view of FIG. 1 .

Referring to FIG. 8 , the simulator of the present invention provides a classification result table (Detail Report) that can grasp in detail the classification result that is difficult to confirm only with a graph. Through this classification result table (Detail Report), the classification result of all data can be viewed in detail after classification is performed in the network.

The classification result table (Detail Report) contains the pattern number (PatternID) of each input vector, context information (Context), the ground truth category value (CatGT) if there is a correct category answer, Status (Status) indicating whether the determination result is correct, Information on firing neurons and category information (CatOut) selected by the category classification policy are included. In addition, Cat1 represents the category of the first firing neuron, Dist1 represents the distance value of the first firing neuron, and Nid1 represents the ID of the first firing neuron (NeuronID in the learning result table). In addition, it will be apparent that Cat2, Dist2 and Nid represent the category, distance value, and neuron ID of the second firing neuron. At this time, the firing neurons with the shortest distance are stored from number 1, and the firing neurons with the longest distance are stored at the last number.

The simulator according to an embodiment of the present invention is largely divided into two structures: a learning processor and a classification processor, and operates.

9 is a diagram illustrating an operation process of a simulator according to an embodiment of the present invention.

Referring to FIG. 9 , the simulator operates by being divided into two processors: learning and classification. There are two cases of the learning process. There is a process of learning by setting a desired option for data to be used for learning (Case 1), and a process of calling a set of already learned neurons and using it for testing (Case 2). The test process selects a dataset and the user selects a classification option to check the results after classification.

If the result is not satisfactory in each process, the process can be repeated until the desired result is obtained by returning to the option setting or changing the data set.

In the simulator according to an embodiment of the present invention, the RBF algorithm 20 having a dual layer structure may be applied in the learning and classification process. In the learning process, the RBF algorithm 20 with a double layer structure can be applied by selecting T-RBF in the RBF Learning Option 120 box. applied by selection.

The RBF algorithm 20 of the double layer structure is designed based on the Reduced Coulomb Energy (RCE) network, which is an algorithm built into the CM1K chip. The RCE network is a type of RBF for identifying patterns and consists of 1024 neurons that can store and recognize vectors up to 256 bytes long. All neurons operate in parallel and cooperate with each other via a bidirectional neuron bus.

FIG. 10 is a diagram illustrating a process in which a decision space is defined to distinguish a category of an input vector, and then neurons in the decision space selectively store information of the input vector to generate an influence field.

Referring to FIG. 10 , pattern recognition in the learning process begins with defining a decision space suitable for classifying categories of input data. Neurons in the decision space create influence fields while selectively storing information from input vectors.

The first input vector is the information stored in the neuron. From the second input vector, we check if there is a neuron that recognizes the input vector. If there is a recognizing neuron, it moves on to the next input vector without taking any action, and if there is no recognizing neuron, the information is stored in a new neuron. Even if the corresponding input vector is recognized by the neuron, if the category is different, the information is stored in the neuron. At this time, even if the corresponding input vector is recognized by the neuron, if the category is different, the information is stored in the neuron.

After that, learning is performed by repeatedly inputting the learning dataset several times, but if there is no more change in the neurons, the learning is terminated.

Neurons generated in this way generate an influence field by calculating the distance between the input vector and the neuron, and two calculation formulas are provided: Norm L1 (Manhattan distance) and Norm LSup.

11 is a diagram showing (a) Norm L1 calculation formula and graph, (b) Norm LSup calculation formula and graph, (c) Norm L1-based neuron influence field example, and (d) Norm LSup-based neuron influence field example .

Referring to FIG. 11 , a calculation method in which neurons generated by input vectors are used to generate an influence field is as follows.

(1)

(One)

(2)

(2)

는 입력 벡터,

은 뉴런,

은 벡터와 뉴런 사이 거리의 총합을 나타낸다.Equation (1) above is the formula for calculating the Norm L1 (Manhattan distance) distance, and Equation (2) is the formula for calculating the Norm LSup distance. At this time

is the input vector,

silver neurons,

is the sum of the distances between the vector and the neuron.

12 is a diagram illustrating a process of classifying neurons generated in a decision space created after learning in response to an input vector entering an influence field;

13 is a diagram schematically illustrating influence field mapping according to the RCE/RBF classification method and the KNN classification method.

12 and 13 , the difference between the RCE/RBF classification scheme and the KNN classification scheme can be seen.

Here, the RCE/RBF classification method is an algorithm applied when RBF is selected in the classification option box 130 of the GUI module 100 and refers to a classification method using the dual layer structure RBF algorithm 20 . In addition, the KNN classification method refers to a K-Nearest Neighbor classification method as an algorithm applied when KNN is selected in the Classification Option box 130 of the GUI module 100 .

First, the RCE/RBF classification method will be described as follows.

The first of the classification process is to check whether the input vector matches the context expressing the characteristics of the generated neuron. Through this, multiple data that are not related to each other can be recognized and processed at the same time. When an input vector with the same context as a neuron is broadcast and placed in the neuron's influence field, the neuron enters a firing state and is called a firing neuron.

The shortest distance between the input vector and the firing neuron is recorded in the first firing neuron. The process of recording the distance to the firing neuron continues until the second short distance is written to the second firing neuron and the distance is read as 0xFFFF.

The firing neuron is a criterion for determining the category of the input vector, and there may be none, one, or several. In this case, the user can designate how many firing neurons to be considered as a criterion for judging the vector, and the default value set on the simulator is set to 3.

There are Best match, Dominant, Unanimity, and Min consensus of N methods for judging the category of the input vector using the information of the fired neurons, and the details are omitted as they are the same as described above.

Next, the KNN classification method is as follows.

The KNN classification method refers to the K-Nearest Neighbor classification method. This is a classification method that considers the K nearest neighbors. Like RBF, a decision space is created and pattern recognition starts with the neuron's influence field determined. The point is that there is no maximum value of a field, and all decision spaces are mapped.

When the influence fields of neurons of different categories overlap, the influence fields are divided by the same distance. The distance calculation method used is also not used.

The classification process proceeds almost the same as RBF, except that the method of generating the influence field is different, so that when a new input vector comes in, neurons fire in all input vectors and are always recognized. The fact that the input vector is always recognized means that there is no unclassified vector, and that all input vectors have firing neurons. The policy for determining the category of the input vector based on the information of firing neurons is the same as in the case of the RCE/RBF algorithm described above.

In the case of classification using KNN, it is used to check the distance and category for all input vectors rather than for efficient classification of the classifier.

14 is a diagram showing the structure of an existing multi-neural RBF network,

15 is a diagram showing the structure of the dual-layer structure RBF algorithm 10. As shown in FIG.

14 and 15 , the existing multi-neural RBF algorithm detects the results individually when multiple inputs are received, and uses a simple classification method such as best match for the result to obtain an integrated final result. was pulling out However, since this method does not consider the characteristics of the prediction results of multiple inputs during classification, the development of an algorithm that can effectively integrate the prediction results is required.

The dual-layer structure RBF algorithm 20 applied to the simulation module of the present invention is designed to have a structure as shown in FIG. 15 and consists of two layers to derive one prediction result by fusing the prediction results of multiple data. A first layer 21 that derives a first result value by recognizing a characteristic pattern using individual data as input data, and a second layer 22 that uses the first result value as an input value by integrating the first result value through a concatenation function ) is a structure composed of

In this case, the first result value includes a distance value between the input vector and the firing neuron and a category that is a prediction result for each input vector.

The basic operation method extracts prediction results and features for each data from the first layer 21 , and integrates the results into one using a ‘Concatenation’ function and uses them as an input for the second layer 22 .

There are two features used here. The distance value (distance) between the firing neuron and the input vector and the prediction result (category) for each data are displayed together. A value derived by calculating the distance between the input pattern and the most similar pattern among the learned patterns to be. Since data with different characteristics are combined, a new context number is assigned, and new neurons with characteristics matching the value are fired.

Experiment example

Hereinafter, the advantages of the dual layer structure RBF algorithm 20 will be described by comparing the accuracy of the dual laser structure RBF algorithm and the existing RBF algorithm when multiple input data is provided.

The data used in the experiment are drowsy driver data created by simulation.

For the simulation, data were collected from 20 people, and 4 types of data (electrocardiogram, heart rate, blinking, steering angle) were collected every minute under the assumption that they were driving for more than 2 hours.

16 is a graph showing the pattern change according to the drowsiness state of four types of simulated data.

Referring to FIG. 16 , it can be seen that the heart rate, blinking eyes, electrocardiogram, and steering angle patterns are different when the driver is in a drowsy state and in an awake state. A section indicated by a red dotted line in FIG. 16 is where a pattern is displayed when the driver is in a drowsy state.

The collected data show relatively clear characteristics when drowsy. Heart rate refers to the number of heartbeats per minute (bpm/beats per minute), and when drowsy, the heart rate remains lower than usual. The electrocardiogram expresses the electrical signal of cardiac activity in frequency units (hz), and it continuously decreases from 10 minutes before drowsiness actually begins, and remains low during drowsiness. Blinking is the number of blinks per minute, and blinks more frequently than usual when tired, and increases by about 4 times when in drowsy state. The steering angle is measured based on the position of the steering wheel in neutral, and when drowsy, the steering wheel movement becomes less and dull, and sometimes abrupt angle changes occur.

In order to use the data of 20 drivers for the analysis of the usability of the double-layered RBF algorithm (20), the data was used in a ratio of 7:3 for learning and testing, and for consistency, the training and test data were divided using a random function. .

17 is a graph showing the accuracy (%) for 20 simulated driver data;

18 is a graph showing the number (dogs) of neurons used in the experiment.

Referring to FIGS. 17 and 18 , the blue graph shows the RBF algorithm 20 having a double layer structure, and the red graph shows the RBF. In general, the accuracy of the RBF algorithm 20 having a double layer structure is high, and the number of neurons used is also higher than that of the RBF.

Accuracy increased by about 1.8% on average, and the number of neurons used increased by about 50 on average. As a result, it was possible to confirm the performance improvement of the dual structure RBF by increasing the accuracy compared to the previous study and obtaining a high value of 96.2%.

Looking at the detailed analysis results, in terms of accuracy, the RBF algorithm 20 having a double layer structure shows increased accuracy compared to the existing algorithm, but it can be seen that the deviation is particularly small. This means that the RBF algorithm 20 of the dual layer structure consistently exhibits high performance for various objects.

However, in terms of neurons used, more neurons were used compared to the existing algorithm, but it is judged that additional neurons were used to express various patterns. However, when comparing the ratio of neurons used, the existing algorithm is 11% and the proposed algorithm is 17%, showing less than 20% usage in both cases. This usage rate shows that although the number of neurons used has increased, it is still using a small number of neurons compared to the total number of neurons.

Therefore, the RBF algorithm 20 of this dual layer structure, which still uses a low number of neurons while providing consistent high accuracy, is judged to be an algorithm suitable for use in multi-input prediction.

Although the present invention has been described above through examples, the above examples are merely for explaining the spirit of the present invention and are not limited thereto. Those skilled in the art will understand that various modifications may be made to the above-described embodiments. The scope of the present invention is determined only through interpretation of the appended claims.

10: GUI-based RBF Algorithm Simulator
100: GUI module 200: simulation module
110: status option box 120: learning option box
130: Classification Options Box 140: Consolidated Graph Box
150: result graph box 151: learning result graph box
152: classification result graph box 160: log box

Claims (11)

In the simulator for simulating the RBF algorithm operated in the neuromorphic chip,
a simulation module for simulating the RBF algorithm; and
and a GUI module for inputting options for manipulating the simulation module and outputting simulation results,
The GUI module is:
a learning option box for adjusting options required for learning on the simulation module;
a classification option box for adjusting options required for classification on the simulation module;
a learning result graph box displaying the number of neurons generated for each category and the number of neurons generated for each number of repeated learning; and
a classification result graph box displaying the number of Status codes classified by category and the total amount of Status codes;
providing a learning result table for identifying a learning result and a classification result table for identifying a classification result;
The learning options box is:
Includes a dual-layer learning algorithm selection unit to learn through a dual-layer structure RBF algorithm,
The dual layer structure RBF algorithm is:
a first layer for deriving a first result value by recognizing a characteristic pattern using individual data as input data; and
and a second layer that integrates the first result value through a concatenation function and uses it as an input value,
The first result is:
distance value between input vector and firing neuron; and
including a category that is a prediction result for each input vector,
The learning result graph box contains:
a graph displaying the number of neurons generated for each category;
A graph is displayed showing the neurons generated for each iteration,
The classification result graph box contains:
a graph displaying the number of status code codes classified by category;
A graph showing the total amount of Status codes is displayed,
The learning result table includes information about a neuron number (NeuronID), a context (Context), a category (Category), a distance value of an influence field (InfluenceField), and a minimum value of an influence field (MinIF),
The classification result table includes the pattern number (PatternID) of each input pattern, context (Context), ground truth category value (GatGT) if there is a category correct answer, Status indicating whether the discrimination result is correct, information and category classification of firing neurons Category information selected by the policy (CatOut), the category of the first firing neuron (Cat 1), the distance value of the first firing neuron (Dist 1), and information about the ID of the first firing neuron (Nid 1) , a GUI-based RBF algorithm simulator. According to claim 1,
The GUI module is:
a state option box for adjusting the state of the simulation module; and
A GUI-based RBF algorithm simulator, including an integrated graph box displaying the number of neurons and data used in the simulation. 3. The method of claim 2,
The status option box is:
a distance calculation method control unit for selecting a distance calculation method between the input vector and the neuron;
a chip number selection unit that selects the number of neuromorphic chips used for simulation; and
Includes a learning information delete button that erases the information of the learned neuron,
A GUI-based RBF algorithm simulator that displays the name of data used for simulation, the number of data samples, the total number of neurons according to the number of neuromorphic chips, and the number of neurons generated after learning. 4. The method of claim 3,
The learning options box is:
a learning number control unit for adjusting the number of learning repetitions;
a whole data recording/non-recording unit to record all information of the learning data in the neurons;
an influence field value selection unit for setting a maximum value and a minimum value of an influence field generated by neurons; and
A GUI-based RBF algorithm simulator that includes a Start Learning button that commands to start learning. 5. The method of claim 4,
The classification option box is:
a classification algorithm selection unit for selecting the simulation module to classify by using any one of an RBF algorithm and a KNN algorithm;
a classification policy selection unit for selecting a classification policy; and
A GUI based RBF algorithm simulator, including a Start Classification button that commands to start sorting. 6. The method of claim 5,
The integrated graph box displays the number of input data, the number of generated neurons, and the number of classified data, a GUI-based RBF algorithm simulator. delete delete 6. The method of claim 5,
The above classification policy is:
Best Match, which selects the category of the first firing neuron as the category of the input vector;
Dominant selecting the category with the largest number of categories of firing neurons as the category of the input vector;
Unanimity that determines the category of the input vector when all categories of firing neurons are the same; and
A GUI-based RBF algorithm simulator comprising at least one of Min consensus that determines a corresponding category as a category of an input vector only when the number of the same category among categories of firing neurons is equal to or greater than a preset number. delete delete

Images