KR102339061B1 - Neural waveform classification device and learning apparatus and method therefor - Google Patents

Abstract

The present invention receives a plurality of neural waveforms obtained in a predetermined manner from a neural signal sensed through at least one electrode, and extracts features for each of the plurality of neural waveforms applied according to a predetermined feature extraction technique. A clustering unit that clusters a plurality of extracted features into a plurality of clusters according to a predetermined clustering technique, determines the center point of each of the plurality of clusters, and determines the determined center point among the total number of features of each cluster based on the determined center point A template selector that selects, as a template feature, a number of features according to a predetermined template selection ratio adjacent as a reference, and a neural waveform corresponding to the selected template feature from among a plurality of applied neural waveforms are extracted as a template waveform, and added to the extracted template waveform A neural waveform classification apparatus capable of easily acquiring a large amount of accurately classified and labeled training data from a neural waveform, including a training data acquisition unit that obtains training data by labeling an identifier corresponding to each cluster with a label, and learning for the same Apparatus and methods may be provided.

Description

Neural waveform classification device and learning device and method therefor

The present invention relates to a neural waveform classification apparatus and a learning apparatus and method therefor, and to a neural waveform classification apparatus capable of efficiently acquiring learning data and a learning apparatus and method therefor.

The neural waveform test is a test for analyzing and discriminating brain activity by detecting an electrical change (current activity) that occurs when a signal is transmitted between neurons in the brain.

In general, the neural waveform test is performed by attaching a micro-electrode to the cortex of the brain, and sensing and amplifying the spontaneous electrical activity generated by the activity of nerve cells in the brain over a certain period of time through the attached electrode. An invasive method is used. However, since the size of the electrode is very large compared to the size of the nerve cell, in most cases, a very large number of neural waveforms are simultaneously sensed by one electrode.

Therefore, when examining a neural waveform, only a neural signal having an intensity greater than or equal to a predetermined reference intensity is extracted among a plurality of neural signals sensed through an electrode to obtain a neural waveform for a nerve cell that is performing a meaningful activity, but separately acquired In most cases, 2 to 10 neural waveforms are mixed and detected even in the neural waveform. Therefore, it is difficult to accurately classify a plurality of mixed and detected neural waveforms, which has a limitation in that it is not easy to analyze the neural waveforms.

In order to overcome this limitation, recently, research to classify neural waveforms using artificial neural networks is being conducted. However, in order to classify neural waveforms using the artificial neural network, the artificial neural network must be trained in advance, and learning data is required to train the artificial neural network. Here, the training data refers to data in which each of the different neural waveforms is pre-classified and labeled with a delimiter.

However, in order to acquire training data from which the artificial neural network can be effectively trained, it is necessary to accurately classify a number of neural waveforms, and to label each classified neural waveform with a separate identifier. In addition, since neural waveforms are acquired in large numbers (eg, about 4000) for a certain period of time during neural waveform testing, it is practically very difficult to label the classifiers by classifying the acquired neural waveforms. In particular, in order to train an artificial neural network, learning of a large number of neural waveforms of various patterns is required, so it is practically impossible to obtain learning data by labeling a separator for a large number of neural waveforms obtained in large numbers.

Korean Patent Publication No. 10-2015-0085007 (published on July 22, 2015)

It is an object of the present invention to provide a neural waveform classification apparatus capable of effectively learning an artificial neural network for neural waveform classification, and a learning apparatus and method therefor.

Another object of the present invention is to provide a neural waveform classification apparatus capable of easily acquiring accurately classified learning data, and a learning apparatus and method therefor.

In order to achieve the above object, a learning apparatus for a neural waveform classification apparatus according to an embodiment of the present invention receives a plurality of neural waveforms obtained in a predetermined manner from a neural signal sensed through at least one electrode, and a feature extraction unit for extracting features for each of a plurality of neural waveforms applied according to a feature extraction technique; a clustering unit for clustering the plurality of extracted features into a plurality of clusters according to a predetermined clustering technique; A template selection unit that determines the center point of each of a plurality of clusters, and selects, as a template feature, a number of features according to a predetermined template selection ratio adjacent to the determined center point among the total number of features of each cluster based on the determined center point ; and a learning data acquisition unit that extracts a neural waveform corresponding to a selected template feature from among a plurality of applied neural waveforms as a template waveform, and obtains learning data by labeling the extracted template waveform with an identifier corresponding to each cluster with a label. .

The learning data acquisition unit may apply the template waveform as an input of an artificial neural network to be learned to receive and classify a plurality of neural waveforms.

The learning apparatus further includes an error backpropagating unit for calculating an error by comparing the identifier obtained by classifying each of a plurality of neural waveforms in the artificial neural network with the identifier of the learning data, and backpropagating the calculated error to the artificial neural network for learning can do.

The feature extractor may extract features having a dimension lower than that of the neural waveforms from the plurality of neural waveforms according to a principal component analysis technique.

The clustering unit may cluster the plurality of features into K clusters according to a K-mean clustering technique.

The template selection ratio may be designated as a single value within the range of 5 to 20%.

A neural waveform classification apparatus according to another embodiment of the present invention for achieving the above object includes: a signal detector for sampling an analog low-level neural signal sensed through at least one electrode and converting it to a digital signal to obtain the neural signal; a preprocessor for extracting a neural signal having an intensity greater than or equal to a preset reference intensity from the neural signal and arranging it in a preset manner to obtain a plurality of neural waveforms; a neural waveform classification unit implemented as an artificial neural network, classifying a plurality of neural waveforms according to a pre-learned pattern estimation method, and outputting an identifier corresponding to each of the classified neural waveforms; and a plurality of neural waveforms coupled during learning of the neural waveform classifier, extracting features for each of the plurality of applied neural waveforms according to a predetermined feature extraction technique, and a clustering technique using the plurality of extracted features By clustering into a plurality of clusters according to by selecting as a template feature, extracting a neural waveform corresponding to the selected template feature among a plurality of applied neural waveforms as a template waveform, and labeling the extracted template waveform with an identifier corresponding to each cluster as a label to obtain training data Includes study.

In a learning method for a neural waveform classification apparatus according to another embodiment of the present invention for achieving the above object, a plurality of neural waveforms obtained in a predetermined manner from a neural signal sensed through at least one electrode are applied, and a predetermined extracting features for each of a plurality of applied neural waveforms according to a feature extraction technique; clustering the extracted plurality of features into a plurality of clusters according to a predetermined clustering technique; determining a center point of each of a plurality of clusters; selecting, as template features, a number of features according to a predetermined template selection ratio adjacent to the determined central point from among the total number of features of each cluster based on the determined central point; extracting a neural waveform corresponding to a selected template feature from among a plurality of applied neural waveforms as a template waveform; and labeling the extracted template waveform with an identifier corresponding to each cluster with a label to obtain training data.

Therefore, the learning apparatus and method for classifying a neural waveform according to an embodiment of the present invention can easily obtain a large amount of accurately classified and labeled learning data from a neural waveform, and perform neural waveform classification based on the acquired learning data. It can effectively train artificial neural networks for

1 shows a schematic structure of a neural waveform classification apparatus including a learning apparatus for neural waveform classification according to an embodiment of the present invention.
FIG. 2 shows signals obtained in each configuration of the learning apparatus for classifying neural waveforms of FIG. 1 .
FIG. 3 is a diagram for explaining the concept of extracting features from a plurality of neural waveforms by the feature extraction unit of FIG. 1 .
FIG. 4 is a diagram for explaining the configuration of the neural waveform classification unit of FIG. 1 and the concept of learning the same.
5 shows a learning method for neural waveform classification according to an embodiment of the present invention.
6 shows a simulation result of learning accuracy according to a template feature selection ratio.
7 shows a simulation result of classification accuracy according to the number of clusters and noise.

In order to fully understand the present invention, the operational advantages of the present invention, and the objects achieved by the practice of the present invention, reference should be made to the accompanying drawings illustrating preferred embodiments of the present invention and the contents described in the accompanying drawings.

Hereinafter, the present invention will be described in detail by describing preferred embodiments of the present invention with reference to the accompanying drawings. However, the present invention may be embodied in various different forms, and is not limited to the described embodiments. In addition, in order to clearly explain the present invention, parts irrelevant to the description are omitted, and the same reference numerals in the drawings indicate the same members.

Throughout the specification, when a part "includes" a certain component, it does not exclude other components, unless otherwise stated, meaning that other components may be further included. In addition, terms such as "...unit", "...group", "module", and "block" described in the specification mean a unit that processes at least one function or operation, which is hardware, software, or hardware. and a combination of software.

1 shows a schematic structure of a neural waveform classification apparatus including a learning apparatus for neural waveform classification according to an embodiment of the present invention, and FIG. 2 is obtained in each configuration of the neural waveform classification learning apparatus of FIG. indicates a signal. 3 is a diagram for explaining the concept of extracting features from a plurality of neural waveforms by the feature extraction unit of FIG. 1, and FIG. 4 is a diagram for explaining the configuration of the neural waveform classification unit of FIG. 1 and the concept of learning it.

Referring to FIG. 1 , the apparatus for classifying a neural waveform according to the present embodiment includes a signal detecting unit 100 , a preprocessing unit 200 , a neural waveform classifying unit 300 , and a learning unit 400 .

First, the signal detector 100 acquires a plurality of neural signals by being applied through at least one electrode that detects an electrical signal generated by the activity of a nerve cell.

The signal detection unit 100 obtains a raw level neural signal that is repeatedly applied for a predetermined period through each of at least one electrode, and samples the obtained low level neural signal to convert it into a digital neural signal . At this time, since the strength of the electrical signal generated from the nerve cells is very weak, the signal detection unit 100 amplifies the obtained low-level nerve signal, and converts the amplified analog low-level signal to a digital nerve signal using an analog-to-digital converter. can be converted into a signal.

The preprocessor 200 extracts and aligns the neural signal having an intensity equal to or greater than a predetermined reference intensity from among the neural signals acquired by the signal detector 100 to obtain a neural waveform. Here, the preprocessing unit 200 extracts the neural signal of the reference intensity or higher in order to extract the neural signal generated from the nerve cell that is performing a meaningful activity among the nerve cells in which the nerve signal is detected.

As shown in (a) of FIG. 2 , a neural signal obtained by sensing an electrical signal generated in a nerve cell may be divided into a plurality of signals having a small amplitude and a signal having a large amplitude appearing in the form of spikes. Here, the signal component in the form of a spike is an electrical signal generated as the nerve cell performs a significant activity. However, nerve cells can generate microscopic electrical signals even when they are not performing any significant activity. In addition, noise may be included in the process in which the signal detector 100 detects and amplifies an electrical signal of a very weak strength generated from a nerve cell. Therefore, the neural signal is acquired together with a signal with a small amplitude and a spike signal with a large amplitude.

Accordingly, the preprocessor 200 extracts only the spike signal of a predetermined reference intensity or higher, arranges it in a predetermined manner, and obtains a neural waveform as shown in FIG. make it possible

The signal detecting unit 100 and the preprocessing unit 200 may be integrated into the neural waveform obtaining unit. In addition, since the signal detector 100 detects the neural signal and the preprocessor 200 acquires the neural waveform from the neural signal is the same as before, a detailed description will be omitted here.

The neural waveform classification unit 300 may receive a plurality of neural waveforms obtained by preprocessing in the preprocessing unit 200, classify each of the plurality of neural waveforms according to a pattern according to a pre-learned pattern estimation method, and label the identifier. . The neural waveform classification unit 300 classifies a plurality of neural waveforms applied as shown in (b) of FIG. 2 into three neural waveforms as shown in (c) of FIG. can be labeled.

The neural waveform classifying unit 300 may be implemented as an artificial neural network, extracting features of each of a plurality of neural waveforms acquired in the preprocessing unit 200 according to a pre-learned pattern estimation method, and applying the pre-learned pattern classification method. By classifying the extracted features, it is possible to distinguish and identify multiple neural waveforms. That is, it is possible to distinguish and classify the same neural waveform from a different neural waveform among a plurality of neural waveforms. In addition, it is possible to label and output a label assigned to each of the classified neural waveforms.

A detailed description of the neural waveform classification unit 300 will be described later.

The learning unit 400 is a learning apparatus for neural waveform classification according to the present embodiment, and is coupled only when learning of the neural waveform classification apparatus, and may be removed from the neural waveform classification apparatus after learning is completed.

The learner 400 acquires learning data from a plurality of neural waveforms acquired by the preprocessor 200 , and trains the neural waveform classifier 300 using the acquired learning data.

The learning unit 400 may include a feature extraction unit 410 , a clustering unit 420 , a template selection unit 430 , a training data acquisition unit 440 , and an error backpropagation unit 450 .

The feature extraction unit 410 receives a plurality of neural waveforms obtained from the preprocessor 200 and extracts features in a predetermined manner. The feature extraction unit 410 receives a plurality of acquired neural waveforms as shown in (b) of FIG. 2, extracts features of each of the plurality of applied neural waveforms, and has a predetermined dimension as shown in (d) of FIG. It can be expressed as a plurality of feature points in the feature space.

The feature extractor 410 may extract feature points using various techniques, but here, as an example, the feature extractor 410 extracts the features of a neural waveform according to a Principal Component Analysis (PCA) technique. Assume The PCA technique projects a number of data distributed in a multidimensional space to a lower dimension, and transforms the data into lower dimensional (or order) data while minimizing information loss of each data, thereby facilitating data analysis. is a technique

As shown in FIG. 3( a ), when a plurality of data is applied, the PCA technique acquires at least one principal component PC1 and PC2 in the order of increasing the variance according to the distribution of the applied data. Here, each of the at least one principal component PC1 and PC2 is a feature of neural waveforms, and the number of principal components obtained may be adjusted according to the number of features of the neural waveform. If two principal components are obtained, as shown in FIG. 3B, by projecting the data on the feature space with the two principal components PC1 and PC2 as each axis, as shown in FIG. 3C, A feature with a lower order can be obtained.

For example, in the present embodiment, each of a plurality of neural waveforms may be acquired as 32-dimensional data, and when the PCA technique is applied, the plurality of neural waveforms may be extracted as 3-dimensional features.

When a plurality of feature points corresponding to each of the plurality of linear waveforms are extracted by the feature extraction unit 410 , the clustering unit 420 clusters the plurality of extracted features in a predetermined manner. The clustering unit 420 classifies features of a plurality of corresponding linear waveforms into a plurality of clusters as shown in FIG. 2E by clustering the plurality of extracted features. In FIG. 2E , the clustering unit 420 classifies a plurality of features into three clusters, and each cluster is displayed in a different color for classification.

The clustering unit 420 includes a K-means clustering algorithm, a Gaussian mixture model (GMM) algorithm, and a density-based spatial clustering of applications using noise (DBSCAN). Various algorithms such as algorithms can be used, but it is assumed here that a plurality of features are clustered into K predetermined numbers according to a K-Means clustering technique as an example.

As the feature extraction unit 410 extracts the 3D feature from the 32D neural waveform, the clustering performance of the clustering unit 420 may be greatly improved.

When a plurality of features are clustered by the clustering unit 420 , the template selector 430 selects a template feature from each of the plurality of clustered clusters. The template selection unit 430 determines a center point of each of the plurality of clusters, and selects a number of adjacent features according to a predetermined ratio based on the determined center points in each of the plurality of clusters as the template features.

As an example, the template selector 430 may select a feature according to a 10% ratio adjacent to a plurality of feature center points clustered in each cluster as a template feature. Here, the template selection unit 430 selects the number of features according to a predetermined ratio adjacent to the central point as template features from the neural waveform classified according to the features clustered by the feature extraction unit 410 and the clustering unit 420 . This is to further improve classification reliability.

As described above, since the neural waveform classification unit 300 implemented as an artificial neural network classifies the neural waveform according to the pattern estimation method learned in advance, the training data for learning the neural waveform classification unit 300 includes the classification result. It should be labeled. And at this time, the label labeled on the training data requires very high reliability. If there is an error in the label of the training data, the neural waveform classifying unit 300 is not normally trained, and thus it is impossible to accurately classify the neural waveform.

However, as described above, even when the features of multiple neural waveforms are classified according to the PCA technique and the K-means clustering technique, it is difficult to completely trust the classification accuracy. In particular, when a plurality of clusters clustered according to the K-means clustering technique are adjacent to each other, or when the distribution of features included in each cluster is very discrete, that is, when the variance is large, the reliability of the clustering result of each feature is inevitably low. In particular, in each cluster, the reliability of the clustering result decreases as the characteristics of the position relatively far from the central point are.

Accordingly, in the present embodiment, including the template selector 430 , adjacent features are selected by a predetermined ratio from a central point that can be considered to be most reliably classified in each cluster, and acquired as template features. That is, features with high clustering reliability are selected from each of a plurality of clusters and obtained as template features.

The training data acquisition unit 440 acquires, as a template waveform, a neural waveform corresponding to each of a plurality of template features acquired from each of a plurality of clusters, and labels the cluster identifier separated by the clustering unit 420 on the acquired template waveform. Obtain training data by labeling. That is, a neural waveform classified with high reliability among a plurality of neural waveforms is extracted as a template waveform, and learning data is generated by labeling the extracted template waveform with an identifier.

Here, the ratio of the number of features that the template selector 430 selects as the template feature among the features of each cluster may be set in various ways. . That is, the possibility of errors in the training data increases. On the other hand, as the number ratio decreases, the amount of learning data obtainable from the neural signal decreases. Therefore, in this embodiment, the effective ratio was analyzed through an experiment, and as a result of the experiment, it was confirmed that it is desirable to select 5 to 20% of the number of features in each cluster as the template feature, and here, as an example, 10% is selected as the template feature. was assumed to be.

The learning data acquisition unit 440 transmits the acquired template waveform as an input of the neural waveform classification unit 300 .

The error backpropagation unit 450 compares the identifier, which is a result of the neural waveform classification unit 300 receiving and classifying a plurality of template waveforms, with the identifier labeled by the learning data acquisition unit 440 for each template waveform, and is a neural waveform classification unit. (300) to determine the error. Then, the neural waveform classification unit 300 is trained by backpropagating the determined error to the neural waveform classification unit.

That is, as shown in FIG. 4 , the learning unit 400 according to this embodiment extracts a feature of a predetermined ratio based on the central point in each of a plurality of clustered feature groups extracted from a plurality of neural waveforms as a template feature. and a template waveform, which is a plurality of neural waveforms corresponding to the extracted template features, is applied as an input of the neural waveform classification unit 300 . And identifiers (Neuron #1, Neuron #2, ..., Neuron #K) and the identifier acquired by the learning data acquisition unit 440 are compared and the error is backpropagated, so that the neural waveform classification unit 300 can be accurately trained.

The fully connected layer of the neural waveform classification unit may be implemented as a layer according to various active functions, but it is assumed here that it is implemented as a layer according to a softmax function as an example.

5 shows a learning method for neural waveform classification according to an embodiment of the present invention.

1 to 4, when explaining the learning method for the neural waveform classification of Fig. 5, first, the signal detector 100 is a preprocessor ( 200) extracts a neural signal having an intensity greater than or equal to a predetermined reference intensity and receives and acquires a plurality of sorted neural waveforms (S10)

When a plurality of neural waveforms are acquired, features of each of the acquired neural waveforms are extracted in a predetermined manner (S20). In this case, a plurality of features may be extracted according to the PCA technique and converted to a lower dimension than the neural waveform.

When a plurality of features are extracted, the plurality of extracted features are clustered in a predetermined manner (S30). Here, a plurality of features may be clustered into a plurality of predetermined clusters according to a K-means clustering technique.

When a plurality of features are clustered into K clusters, a center point of each of the plurality of clusters is determined (S40). Then, among the total number of features of each cluster, a number of features according to a predetermined ratio adjacent to the determined central point is selected as template features (S50).

When a template feature is selected, a neural waveform corresponding to the selected template feature from among a plurality of applied neural waveforms is extracted as a template waveform (S60). Thereafter, learning data is obtained by labeling the extracted template waveform with an identifier corresponding to each cluster with a label ( S70 ).

When the training data is acquired, the template waveform of the acquired training data is input to the neural waveform classification unit 300 implemented as an artificial neural network (S80). And by comparing the identifier that the neural waveform classification unit 300 classifies and outputs the template waveform with the identifier labeled on the template waveform in the training data, calculates an error, and back propagates the calculated error to the neural waveform classification unit 300 The neural waveform classifier 300 is trained (S90).

As a result, the learning apparatus and method for classifying neural waveforms according to the present embodiment very easily acquire learning data in which a neural waveform is classified from a neural signal detected in the activity of a neural cell, and an identifier according to the classified neural waveform is labeled. can do. In particular, since a separate manual operation is not required and the learning data labeled with the identifier can be obtained automatically from the neural signal, it is possible to collect a large amount of learning data very effectively, and to clearly As only a classifiable neural waveform is used for learning as a template waveform, the neural waveform classification apparatus can be trained with very high performance.

6 shows a simulation result of learning accuracy according to a template feature selection ratio.

6(a) shows the distribution of a plurality of features extracted by the feature extraction unit 410, and red dots in (a) indicate a plurality of features in the clustering unit 420 into five clusters (K = 5). indicates the position of the center point obtained from the clustering result. Looking at (a), it can be intuitively seen that the location of the center point of the lower two clusters among the five clusters is inappropriate. This is because, as described above, even if the PCA and K-mean clustering techniques are applied, an error may occur in the clustering result as shown in (a). And when learning is performed by extracting adjacent features based on the central point as template features while controlling the template feature selection ratio, as shown in (b), high accuracy at low template feature selection ratio despite the error of the clustering result. It can be seen that the neural waveform classification device can be trained with the learning performance of

7 shows a simulation result of classification accuracy according to the number of clusters and noise.

7, (a) to (d) show the performance of the neural waveform classification apparatus learned according to the learning method of this embodiment when the number of clusters, ie, the number of classifications, is 2 to 5, respectively, according to the existing Mean Squared Error (MSE). ), Correlation Coefficient (CC), and PCA + K mean clustering techniques compared to the case of classification are shown.

As shown in (a) to (d) of Figure 7, the artificial neural network learned by extracting template waveforms at a predetermined ratio based on the central point from the features classified according to PCA + K mean clustering according to the present embodiment. When used, it can be seen that neural waveforms can always be classified with high accuracy regardless of the number of clusters or the noise level. That is, it can be seen that the performance of the neural waveform classification apparatus is very high.

The method according to the present invention may be implemented as a computer program stored in a medium for execution by a computer. Here, the computer-readable medium may be any available medium that can be accessed by a computer, and may include all computer storage media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data, and read dedicated memory), RAM (Random Access Memory), CD (Compact Disk)-ROM, DVD (Digital Video Disk)-ROM, magnetic tape, floppy disk, optical data storage, and the like.

Although the present invention has been described with reference to the embodiment shown in the drawings, which is only exemplary, those skilled in the art will understand that various modifications and equivalent other embodiments are possible therefrom.

Accordingly, the true technical protection scope of the present invention should be defined by the technical spirit of the appended claims.

100: signal detection unit 200: pre-processing unit
300: neural waveform classification unit 400: learning unit
410: feature extraction unit 420: clustering unit
430: template selection unit 440: learning data acquisition unit
450: error back propagation unit

Claims (14)

a feature extraction unit for receiving a plurality of neural waveforms obtained in a predetermined manner from a neural signal sensed through at least one electrode, and extracting features for each of the plurality of neural waveforms applied according to a predetermined feature extraction technique;
a clustering unit for clustering the plurality of extracted features into a plurality of clusters according to a predetermined clustering technique;
A template selection unit that determines the center point of each of a plurality of clusters, and selects, as a template feature, a number of features according to a predetermined template selection ratio adjacent to the determined center point among the total number of features of each cluster based on the determined center point ; and
A training data acquisition unit that extracts a neural waveform corresponding to a selected template feature from among a plurality of applied neural waveforms as a template waveform, and labels the extracted template waveform with an identifier corresponding to each cluster with a label to acquire learning data,
The template selection ratio is
A learning device for classifying neural waveforms that is specified as a single value within the range of 5 to 20%.
According to claim 1, wherein the learning data acquisition unit
Applying the template waveform as an input of an artificial neural network that must be learned to receive and classify a plurality of neural waveforms,
the learning device
Neural waveform classification further comprising an error backpropagating unit for calculating an error by comparing the identifier obtained by classifying each of a plurality of neural waveforms in the artificial neural network with the identifier of the learning data, and backpropagating the calculated error to the artificial neural network for learning learning device for The method of claim 1, wherein the feature extraction unit
A learning apparatus for neural waveform classification for extracting features having a dimension lower than the dimension of the neural waveforms from the plurality of neural waveforms according to a principal component analysis technique. The method of claim 1, wherein the clustering unit
A learning apparatus for neural waveform classification that clusters the plurality of features into K clusters according to a K-mean clustering technique. delete a signal detection unit that samples an analog low-level neural signal sensed through at least one electrode and converts it into a digital signal to obtain the neural signal;
a preprocessing unit for extracting a neural signal having an intensity greater than or equal to a preset reference intensity from the neural signal and arranging it in a preset manner to obtain a plurality of neural waveforms;
a neural waveform classification unit implemented as an artificial neural network, classifying a plurality of neural waveforms according to a pre-learned pattern estimation method, and outputting an identifier corresponding to each of the classified neural waveforms; and
The neural waveform classifier is combined during learning to receive a plurality of neural waveforms, extract features for each of the plurality of applied neural waveforms according to a predetermined feature extraction technique, and apply the plurality of extracted features to a predetermined clustering technique By clustering into a plurality of clusters according to Learning to obtain training data by selecting as a template feature, extracting a neural waveform corresponding to the selected template feature from among a plurality of applied neural waveforms as a template waveform, and labeling the extracted template waveform with an identifier corresponding to each cluster as a label including wealth,
The template selection ratio is
A neural waveform classifier specified as a single value within the range of 5 to 20%. The method of claim 6, wherein the learning unit
Applying the template waveform to the neural waveform classification unit, calculating an error by comparing the identifier output from the neural waveform classification unit with the identifier of the training data, and backpropagating the calculated error to the artificial neural network to classify the neural waveform A neural waveform classification device further comprising an error backpropagation unit for learning a part. The method of claim 6, wherein the learning unit
A neural waveform classification apparatus for extracting features having a dimension lower than that of the neural waveforms from the plurality of neural waveforms according to a principal component analysis technique. The method of claim 6, wherein the learning unit
A neural waveform classification apparatus for clustering the plurality of features into K clusters according to a K-mean clustering technique. As a method performed in a learning device for neural waveform classification,
receiving a plurality of neural waveforms obtained in a predetermined manner from a neural signal sensed through at least one electrode, and extracting features for each of the plurality of applied neural waveforms according to a predetermined feature extraction technique;
clustering the extracted plurality of features into a plurality of clusters according to a predetermined clustering technique;
determining a center point of each of a plurality of clusters;
selecting, as template features, a number of features according to a predetermined template selection ratio adjacent to the determined central point from among the total number of features of each cluster based on the determined central point;
extracting a neural waveform corresponding to a selected template feature from among a plurality of applied neural waveforms as a template waveform; and
Labeling the extracted template waveform with an identifier corresponding to each cluster with a label to obtain training data,
The template selection ratio is
A learning method for classifying neural waveforms that is specified as a single value within the range of 5 to 20%. 11. The method of claim 10, wherein the learning method
applying the template waveform as an input of an artificial neural network to be learned to receive and classify a plurality of neural waveforms;
calculating an error by comparing an identifier obtained by classifying each of a plurality of neural waveforms in the artificial neural network with an identifier of the training data; and
The learning method for neural waveform classification further comprising the step of backpropagating the calculated error to the artificial neural network. The method of claim 10, wherein the extracting of the feature comprises:
A learning method for neural waveform classification for extracting features having a dimension lower than that of the neural waveforms according to a principal component analysis technique for the plurality of neural waveforms. The method of claim 10, wherein the clustering comprises:
A learning method for neural waveform classification in which the plurality of features are clustered into K clusters according to the K mean clustering technique. delete

Images