JP6825770B2 - Machine learning equipment, classification equipment and methods and programs - Google Patents

Description

The present technology relates to machine learning devices, classification devices and methods and programs.

It is thought that nerve cells regulate the functions of various organs in living organisms, and in the central nervous system, they perform various information processing by forming a network of nerves and activating them. At present, with the development of cell culture technology, it has become possible to culture nerve cells extracted from a living body or nerve cells differentiated from pluripotent stem cells in a culture medium for a long period of time.

The activity of nerve cells is to transmit information using electrical signals, such as the mechanism of electrical signal generation in nerve cells and nerve fibers, the biochemical properties of the membrane surrounding nerve cells, and the response to nerve cells and nerve fibers. In order to elucidate the pharmacological action of a substance, the potential difference between the inside and outside of the membrane of this nerve cell is measured.

When cultured nerve cells are measured, as methods for measuring the action potential of nerve cells, (1) intracellular potential recording method, (2) patch electrode recording method, and (3) extracellular potential recording method can be mentioned. .. Since (1) and (2) are for inserting glass microelectrodes into cells or applying them to the cell surface for recording, it is possible to directly measure the membrane potential, but the procedure is difficult. It is difficult to perform long-term measurements such as damaging the cells themselves. On the other hand, the extracellular potential recording method (3) is to measure the potential change occurring around the cell by arranging the electrode near the cell, which is technically relatively easy and enables long-term measurement. Yes, and its application range is wide. In particular, the extracellular potential recording method using a multi-point planar electrode (MEA: Multiple Electrode Array) is used as a method capable of stably and simultaneously measuring from a plurality of electrodes and a plurality of samples.

Conventionally, as a measuring device used for the extracellular potential recording method, one cell is used so as not to generate noise due to electromagnetic induction generated between signal lines connected to different microelectrodes arranged in different cells or cell groups. Alternatively, the signal line connected to the microelectrodes arranged in the cell group and the signal line connected to the microelectrodes arranged in other cells or cell groups are substantially separated and connected to the amplifier. Those have been proposed (for example, Patent Document 1).

Japanese Unexamined Patent Publication No. 2014-124133

In order to analyze the nerve activity pattern from the action potential data obtained by measuring the change in the cell potential of the nerve cell, noise is excluded from the measured activity waveform data of the nerve cell, and firing indicating the activity of the nerve cell is performed. Is detected.

However, at present, in order to detect the firing of nerve cells such as spikes representing the activity of nerve cells and bursts in which continuous spikes occur, from the measurement data of such extracellular potential changes of nerve cells, the measurer is required. A method of visually detecting based on experience and a method of detecting ignition by setting the threshold value to a constant multiple of the standard deviation are used. However, classification by setting a simple threshold causes false detection of ignition, and it is necessary for humans to visually correct it in order to prevent it. Since the results of human judgment also vary, it cannot be said that the accuracy and work efficiency are good.

Therefore, an object of the present invention is to provide a technique for improving the accuracy and efficiency of the work of detecting firing from the extracellular potential data of nerve cells.

The machine learning device according to the present invention uses the value of the potential measured in a predetermined period and the teacher value indicating the occurrence of ignition in the period for the data representing the time course of the extracellular potential of the nerve cell. Learn the relationship between the presence or absence of ignition and the potential value. Further, the machine learning device uses a feature amount calculation unit that calculates the feature amount based on the potential value and a predetermined function, a value based on the feature amount calculated by the feature amount calculation unit, and a teacher value. It is provided with a learning processing unit that modifies the coefficients used in the function.

In this way, for example, machine learning using a neural network can be performed. Whether or not firing has occurred in the extracellular potential data cannot be classified only by setting a simple threshold value, but it becomes possible to learn classification criteria that are difficult to formulate. That is, it is possible to improve the accuracy and efficiency of the work of detecting firing from the extracellular potential data of nerve cells.

In addition, the feature amount calculation unit takes the potential value or the feature amount based on the potential value as an input, and is based on a fully connected layer in which all the nodes on the input side are connected to all the nodes on the output side and a predetermined coefficient associated with each connection. The converted feature amount may be calculated. For example, in this way, a configuration including a so-called fully connected layer can be adopted.

In addition, the feature amount calculation unit is a function that takes the potential value or the feature amount based on it as an input and uses the value corresponding to the processing result of the previous stage node for the processing of the subsequent stage node for a series of nodes on the input side. The feature amount after conversion may be calculated by a recurrent neural network that performs processing based on the above. For example, in this way, a configuration including a so-called recurrent layer may be adopted. Note that the processing by the recurrent neural network may be to learn the temporal features in the forward direction along with the passage of time, or may be to learn the temporal features in the reverse direction in the time series. Further, both may be executed in parallel.

Further, the feature amount calculation unit may input the potential value or the feature amount based on the potential value or calculate the feature amount after conversion by a predetermined convolutional neural network. For example, it is possible to capture the characteristics of potential values that change over time.

Further, the value of the extracellular potential is measured via a plurality of electrodes arranged so as to be in contact with the nerve cell, and the feature amount calculation unit is a value of a plurality of potentials measured at the same time by different electrodes or a value thereof. The feature amount after conversion may be calculated by using the feature amount based on the feature amount. In this way, it is possible to learn the synchronic characteristics of firing patterns that occurred at the same time in at least a part of the cells.

Further, the value of the extracellular potential is measured via one or a plurality of electrodes arranged so as to be in contact with the nerve cell, and the feature amount calculation unit is measured by the same electrode at different time points within a predetermined period. The converted feature quantity may be calculated by using a plurality of potential values or a feature quantity based on the potential value. In this way, it is possible to learn the characteristics of the change over time in the ignition pattern detected at each electrode.

The classification device of the present invention classifies whether or not ignition has occurred by using the coefficient corrected by the learning processing unit included in the machine learning device and the potential value. In this way, it is possible to classify whether or not firing has occurred from the extracellular potential data based on the characteristics learned by the machine learning device.

The contents described in the means for solving the problems can be combined as much as possible without departing from the problems and technical ideas of the present invention. Further, the content of the means for solving the problem can be provided as a device such as a computer or a system including a plurality of devices, a method executed by the computer, or a program executed by the computer. The program can also be run on the network. Further, a recording medium for holding the program may be provided.

According to the present invention, it is possible to improve the accuracy and efficiency of the work of detecting firing from the extracellular potential data of nerve cells.

It is a perspective view which shows an example of a measuring apparatus. It is a top view which shows an example of a chip. It is a functional block diagram which shows an example of the machine learning apparatus. It is a graph which shows an example of the value of the measured extracellular potential. It is a device block diagram which shows an example of a computer. It is a processing flow diagram which shows an example of the machine learning processing. It is a figure which shows an example of the structure of a neural network. It is a figure for demonstrating an example of the structure of a convolution layer. It is a figure for demonstrating an example of the structure of a fully connected layer. It is a figure for demonstrating an example of the structure of a recurrent layer. It is a figure for demonstrating an example of the detailed structure of a recurrent layer. It is a figure for demonstrating the integrated processing of a feature quantity.

Hereinafter, embodiments of the present invention will be described with reference to the drawings. The embodiment is an example, and the present invention is not limited to the following configuration.

<Device configuration>
FIG. 1 is a perspective view showing an example of a measuring device according to the present embodiment. The measuring device 1 includes a chip holder 11, a potential measuring device 12, a monitor 13, and a machine learning device 14. On the chip holder 11, a chip 111 having a plurality of electrodes for measuring the extracellular potential from the nerve cell group is placed. The potential measuring device 12 amplifies the potential measured via the electrode of the chip 111 by the built-in amplifier, and measures the difference between the amplified potential and a predetermined reference potential. Further, the potential value measured by the potential measuring device 12 is output to the machine learning device 14, and the machine learning device 14 records the measured potential value. At this time, the potentials measured at each of the plurality of electrodes are recorded, for example, in association with the measurement time so that they can be processed in synchronization. Further, the machine learning device 14 learns the characteristics of ignition from the measurement data by using DCNN (Deep Convolutional Neural Network), and detects the ignition from the measurement data by using the learned model. The monitor 13 appropriately displays measurement data, information on the detected ignition, and the like. It should be noted that the firing includes a single firing of a nerve cell, a spike measured as a rapid rise and fall of an electric potential occurring in a short period of time, and a burst in which continuous firing occurs intermittently.

FIG. 2 is a plan view showing an example of the chip (probe) 111. Tissue sections are placed or cells are cultured in the central region of chip 111. Further, 16 measuring electrodes 1111 (also simply referred to as “electrodes”) are provided in the central region of the chip 111, and the potential generated in the contact region of the cell group can be measured. It should be noted that the measurement electrode 1111 can also perform current stimulation according to the output from the machine learning device 14. Further, a reference electrode 1112 for measuring a reference value of the potential is provided around the region where the cell group is placed. The measuring electrode 1111 and the reference electrode 1112 are connected to the potential measuring device 12 via the wiring 1113. A plurality of chips 111 as shown in FIG. 2 may be placed on the chip holder 11 so that potential data can be measured from a plurality of cell groups. The number of electrodes may be one or more, and the shape and the like related to the arrangement when a plurality of electrodes are provided are not limited to the example of FIG.

FIG. 3 is a functional block diagram showing an example of the machine learning device 14. The machine learning device 14 of FIG. 3 includes a potential data storage unit 141, a machine learning unit 142, a classification model storage unit 143, a classification processing unit 144, a classification result storage unit 145, and a verification processing unit 146. The machine learning device 14 divides the continuously measured potential input signal sequence into predetermined periods (sections), learns the occurrence of ignition for each section, and for potential data whose classification result is unknown. The presence or absence of ignition is classified for each predetermined section.

FIG. 4 is a graph showing an example of the value of the extracellular potential measured by the plurality of measuring electrodes 1111 of the measuring device 1. In the example of FIG. 4, the passage of time is shown in the direction of the arrow along the horizontal axis, and in the vertical direction, a graph showing the potential values detected by the eight measurement electrodes 1111 shows the measurement timing. It is shown to synchronize. Specifically, the potential value is recorded as a sequence of numerical values that are continuously measured at predetermined sampling intervals. The extracellular potential data used in this embodiment and the examples described later were provided by Ikuo Suzuki Laboratory, Department of Intelligent Electronics, Tohoku Institute of Technology.

Further, in the present embodiment, a teacher value (label) indicating whether or not ignition has occurred is used for learning. The teacher value is the result of the expert classifying whether or not the potential data has been ignited, and is stored for each predetermined section. For example, the scale of the broken line that divides the waveform of FIG. 4 for each predetermined period on the time axis represents an example of the boundary that divides the graph into predetermined sections. The potential sampling interval measured by the measurement electrode 1111 does not have to coincide with a predetermined section. For example, a plurality of measured values may be acquired within a section divided by the scale in FIG. Good. Further, among the potential data, a data group having a teacher value is called a labeled data set, and a data group without a teacher value is called an unlabeled data set. In the labeled data set, information indicating whether or not ignition has occurred is retained in association with each of the sections separated by the scale of FIG. Specifically, information indicating whether or not spikes have occurred, and if spikes have occurred, whether or not bursts have occurred may be retained, or spikes may occur. Only the information indicating whether or not the spike has occurred or only the information indicating whether or not the burst has occurred may be retained. Further, among the labeled data sets, the one used for generating the classification model described later is called a training data set, and the one used for evaluating the classification model is called a test data set. By using the generated classification model, it becomes possible to classify whether or not an ignition has occurred in an unlabeled dataset.

The potential data storage unit 141 stores the potential data as shown in FIG. 4 and the teacher value indicating whether or not ignition has occurred for each predetermined section of the potential data. In this embodiment, it is assumed that the training data set and the test data set are stored in advance.

The machine learning unit 142 performs machine learning using a neural network (NN: Neural Network) and a training data set. Specifically, the machine learning unit 142 has a feature amount calculation unit 1421 and a learning processing unit 1422. The feature amount calculation unit 1421 performs a process based on the NN having a predetermined configuration on the input potential data, and performs a process of calculating the feature amount of the change in the ignition pattern. Specifically, so-called convolutional layers (also called "convolutional layers" and "convolutional NNs"), recurrent layers (also called "recurrent layers" and "recurrent NNs"), and fully connected (FC: Full Connection) layers ("" Fully connected layer ”and“ fully connected NN ”) are used. Then, the characteristic amount of the change with time of the potential data measured by one electrode, the potential data of each electrode measured by a plurality of electrodes, or the potential data measured by at least a part of the plurality of electrodes can be calculated. , Calculate the characteristic amount of ignition that occurs at the same time, which is measured at least a part of a plurality of electrodes. The learning processing unit 1422 updates the value of the coefficient applied in each layer of the NN based on the difference between the teacher value held by the potential data storage unit 141 and the feature amount output by the NN.

The classification model storage unit 143 stores parameters such as the configuration of the NN adopted in the present embodiment and the weighting coefficient generated by the machine learning unit 142. The NN configuration, parameters, etc. shall be referred to as a "classification model". The classification processing unit 144 uses the test data and unlabeled data stored in the potential data storage unit 141 and the classification model stored in the classification model storage unit 143 to ignite each predetermined section of the potential data. Classify whether or not it has occurred. The classification result storage unit 145 stores the results classified by the classification processing unit 144. In addition, the verification processing unit 146 compares the result of the classification processing using the test data with the teacher value and evaluates the performance of the classification processing.

The machine learning unit 142 is a so-called computer, and when the processor executes a predetermined program, it functions as the machine learning unit 142, the classification processing unit 144, and the verification processing unit 146 described above. Further, the potential data storage unit 141, the classification model storage unit 143, and the classification result storage unit 145 are specifically realized by a main storage device or an auxiliary storage device provided in a computer.

<Device configuration>
FIG. 5 is a device configuration diagram showing an example of a computer. The machine learning device 14 is, for example, a computer as shown in FIG. The computer 1000 shown in FIG. 5 includes a CPU (Central Processing Unit) 1001, a main storage device 1002, an auxiliary storage device 1003, a communication IF (Interface) 1004, an input / output IF (Interface) 1005, a drive device 1006, and a communication bus 1007. ing. The CPU 1001 performs processing and the like according to the present embodiment by executing a program (also referred to as "software" or "application"). The main storage device 1002 caches the programs and data read by the CPU 1001 and expands the work area of the CPU. Specifically, the main storage device is a RAM (Random Access Memory), a ROM (Read Only Memory), or the like. The auxiliary storage device 1003 stores a program executed by the CPU 1001 and setting information used in the present embodiment. Specifically, the auxiliary storage device 1003 is an HDD (Hard-disk Drive), an SSD (Solid State Drive), a flash memory, or the like. The main storage device 1002 and the auxiliary storage device 1003 function as a potential data storage unit 141, a classification model storage unit 143, and a classification result storage unit 145. Although a plurality of storage units are shown in FIG. 3 for convenience of explanation, one storage device or a plurality of storage devices may be physically used. The communication IF1004 transmits / receives data to / from another computer. The machine learning device 14 may receive potential data from a computer (not shown) connected via the communication IF 1004. Specifically, the communication IF1004 is a wired or wireless network card or the like. The input / output IF1005 is connected to an input / output device, receives input from the user, and outputs information to the user. Specifically, the input / output device is a keyboard, a mouse, a display, a touch panel, or the like. The drive device 1006 reads data recorded on a storage medium such as a magnetic disk, a magneto-optical disk, or an optical disk, or writes data to the storage medium. Then, the above-mentioned components are connected by the communication bus 1007. It should be noted that a plurality of these components may be provided, or some components (for example, the drive device 1006) may not be provided. Further, the input / output device may be integrally configured with the computer. Further, a program executed in the present embodiment is provided via a portable storage medium that can be read by the drive device 1006, a portable auxiliary storage device 1003 such as a flash memory, a communication IF 1004, and the like. It may be. Then, the CPU 1001 executes the program to operate the computer as described above as the machine learning device 14 shown in FIG.

A part of the functional block illustrated in FIG. 3 may be shared by a plurality of computers, or a part of a data set to be processed may be processed in parallel by a plurality of computers. Further, the computer may provide a so-called cloud service on the network.

<Machine learning process>
FIG. 6 is a processing flow diagram showing an example of machine learning processing executed by the measuring device 1. First, the measuring device 1 performs a process in the learning stage using the training data among the potential data (S1). In this step, the probability that the ignition occurs for each predetermined section is calculated for the potential data using the NN and the predetermined coefficient having the configuration described later, and the process of updating the coefficient based on the teacher value is repeated.

After that, the measuring device 1 performs processing in the model evaluation stage using the test data among the potential data (S2). In this step, the presence or absence of ignition is classified for each predetermined section with respect to the potential data using the coefficients obtained in NN and S1 having the configuration described later, and it is evaluated whether or not the values sufficiently match the teacher values. According to the classification model evaluated to be sufficiently consistent with the teacher value in this step, it can be expected that the occurrence of ignition can be discerned with the same performance even for unknown potential data.

Further, the measuring device 1 uses the unlabeled data of the potential data to perform processing in the operation stage (S3). In this step, the presence or absence of ignition is classified for each predetermined section with respect to the potential data by using the coefficients obtained by NN and S1 having the configuration described later. Note that S1 to S3 do not need to be executed as continuous processing, and the measuring device 1 (this) having a classification model (that is, NN and coefficient) created in advance in S1 and performing only processing in the operation stage of S3. The "classification device") according to the invention may be provided.

<Construction of neural network>
FIG. 7 is a diagram showing an example of the configuration of an NN including a plurality of layers (also referred to as “layers”). In the NN of FIG. 7, a plurality of conversion processes are performed between the input layer into which the measured potential value is input and the output layer in which the classification result is shown. The configuration and processing of N1 to N5 existing between the input layer and the output layer is called an intermediate layer. Further, the machine learning unit 142 and the classification processing unit 144 of the machine learning device 14 shown in FIG. 3 repeatedly use the NN shown in FIG. 7 in each of S1 to S3 shown in FIG. 6 to perform processing.

First, when the feature amount calculation unit 1421 of the machine learning device 14 receives the input of the potential data divided for each predetermined section, the feature amount calculation unit 1421 uses a predetermined number of potential data that are continuous in time series for each electrode and each section. The feature amount is calculated by the so-called convolution layer (N1). It should be noted that one section includes a plurality of potential data. Further, the range representing the predetermined number continuous in the time series described above is referred to as a "window" in N1. In N1, for example, the window is slid for each potential data to change the input value, and the convolution process is performed.

FIG. 8 is a diagram for explaining a convolution layer. As described above, in the convolution layer, a predetermined number of continuous potential data (FIG. 8: x) selected by the window is multiplied by a weighting coefficient W represented by a matrix of a predetermined size, and a feature amount (FIG. 8). : Y) is obtained. It is assumed that the horizontal arrows in FIG. 8 represent the flow of time in one section, and each of the scales shown on the horizontal arrows represents one potential data. In the convolution layer, a window containing a predetermined number of continuous potential data is slid backward for each potential data, for example, input data to be processed is selected, and the calculation of the feature amount is repeated. Further, the calculation of the feature amount by the convolution layer in several stages may be repeated with respect to the calculated feature amount. Further, for the feature amount calculated by the convolution layer, for example, an average pooling may be taken by so-called pooling to improve the generalization performance against time fluctuations. In the present embodiment, the measured values are processed by the convolution layer for each measurement electrode 1111 and for each section, so that the slope of the waveform, the size of the peak, etc. are within the section of the ignition pattern detected by each electrode. The feature amount of change over time can be calculated.

In addition, the feature amount calculation unit 1421 calculates a new feature amount (output value in N2) by the fully connected layer with respect to the feature amount (input value in N2) calculated in N1 of FIG. 7 (FIG. 7 :). N2). In N2, a new feature amount is calculated from the input value included for each electrode and each section based on a predetermined function.

FIG. 9 is a diagram for explaining a fully connected layer. FIG. 9 shows the relationship between the input value x and the output value y included in one section. In the fully connected layer, the input side unit x is processed according to the weighting coefficient W, and the value of the output side unit y is calculated. In N2, the feature amount (FIG. 8: y) calculated in N1 is the input value (FIG. 9: x). Further, all the units on the input side and all the units on the output side are combined, and a new feature amount y is obtained from the input value x. x is a vector having a predetermined number of elements and can be represented by a matrix. Further, y is also a vector having a predetermined number of elements, and is obtained by the following equation (1).
y = Wx + b ... (1)
Here, W is a weighting coefficient and is represented by a matrix of the number of elements of x × the number of elements of y. Further, b is a bias term and can be represented by the same number of vectors as x. With N2 as well, it is possible to calculate the feature amount for each section of the ignition pattern detected at each electrode.

In addition, the feature amount calculation unit 1421 calculates the feature amount by the so-called recurrent layer with respect to the feature amount calculated in N2 of FIG. 7 (FIG. 7: N3). Also in the recurrent layer, the feature amount is calculated from the input value based on a predetermined function.

FIG. 10 is a diagram for explaining the recurrent layer. The recurrent layer has a loop inside, and for example, the information A used for processing the data in the previous stage on the time series is taken over and used in the processing of the data in the subsequent stage. In other words, in the processing of the data in the previous stage, the feature quantity y is output based on the input value x and the information A, and new information A determined according to the processing of the data in the previous stage is used for processing the data in the subsequent stage. Output. Each of the input value x and the output value y shown by the rounded squares in FIG. 10 represents a feature amount calculated based on the potential value measured in the predetermined section described above. In the recurrent layer, the value including the temporal feature output by N2 is used as the input value (FIG. 10: x), and the feature amount of the ignition pattern in a longer period is calculated. That is, according to N3, it is possible to calculate the feature amount over time between a plurality of sections for each electrode.

FIG. 11 is a diagram showing an example of a detailed configuration of the recurrent layer. The input value x and the output value y shown by the rounded squares in FIG. 10 correspond to the rounded rectangles shown by the broken lines in FIG. A plurality of circles (nodes) surrounded by one rounded rectangle represent a feature amount calculated based on the potential value measured in one section. Further, the recurrent layer of FIG. 11 has a "state" (also referred to as "internal state" or "state") represented by a broken line rectangle. The value of the state z _{t in} a certain section is calculated based on the input value x _t and a predetermined function, and is output to the output value y _{t + 1} and the state z _{t + 1} in the next section. Then, the output value y _{t + 1} of the recurrent layer of FIG. 11 is calculated based on the value of the state z _t , the input value x _{t + 1,} and a predetermined function.

For example, the output value y _{t + 1} of the recurrent layer may be calculated by using the following equation (2).
y _{t + 1} = Wx _{t + 1} + Uz _t + b ... (2)
In this example, the state z _{t + 1} can be calculated using the following equation (3).
z _{t + 1} = Wx _{t + 1} + Uz _t + b ... (3)
W and U are predetermined weighting coefficients, and b is a predetermined bias term.

In this way, the recurrent layer uses an existing function that uses the value corresponding to the processing result of the node in the previous stage for the processing of the node in the subsequent stage for a series of data. For example, for data having a context in a time series, a state generated by processing the data in the previous stage can be used for processing the data in the subsequent stage. The function used in the recurrent layer is not limited to the above equations (2) and (3), and may be calculated by a function different from y _t and z _t , or so-called LSTM (Long Short-). Term Memory (see the following paper) may be used.

Further, the feature amount calculation unit 1421 calculates the feature amount by the fully connected layer by inputting the feature amount calculated in N3 of FIG. 7 (FIG. 7: N4). In particular, when the chip 111 having a plurality of electrodes is used, the feature amount based on the fully connected layer may be calculated by inputting the feature value based on the potential data measured by some or all the electrodes.

The structure of the fully connected layer is the same as that in FIG. Further, in N4, the potential data measured by different electrodes is processed by inputting the characteristic value as a result of performing the above-mentioned N1 to N3 processing. That is, in N4, the input value x in FIG. 9 is a characteristic value of some electrodes or all electrodes in the same section. Therefore, N4 can be used to calculate the synchronic features of firing that occurred at the same time in at least a part of the cell group. Since the firing of nerve cells tends to occur all at once in a wide range of cell groups, the accuracy of classification can be improved by learning the characteristics of firing that occurred at the same time in multiple electrodes, such as all electrodes. ..

Further, using the output value y of N4, the probability that ignition has occurred in each section and the probability that ignition has not occurred in the section are obtained (FIG. 7: N5). When the chip 111 including one electrode is used, the above probability is calculated with respect to the potential data measured by the electrode. Further, when the chip 111 having a plurality of electrodes is used, the probability that one set of ignition has occurred and the probability that ignition has not occurred may be calculated for all the electrodes, and the ignition may occur for each electrode. The probability of occurrence and the probability of non-ignition may be calculated.

In N5, a process of taking a moving average is further performed using the probabilities of a predetermined number of sections before and after each as the probability of occurrence of ignition and the probability of no ignition calculated for each section. You may do so. In general, nerve cell bursts occur continuously for a certain period of time, so when the occurrence of bursts is detected, for example, a temporary decrease in firing frequency can be determined by taking a moving average. It is less likely to be classified as. That is, it functions as a kind of hysteresis in the threshold value, and in the case of classifying as the occurrence of a burst from the non-burst state and the case of classifying as the end of the burst from the state in which the burst is occurring, it is against temporary noise and potential fluctuation. It reduces responsiveness and enables stable classification.

Further, in the learning process (S1) of FIG. 6, the learning process unit 1422 uses the coefficient of the function used in each process (for example, the weighting coefficient described above) based on the difference between the information output in N5 and the teacher value. Bias term value) is corrected. That is, the coefficient is modified so that the difference between the probability that the ignition has occurred and the probability that the ignition has not occurred and the teacher value is small. Existing techniques such as backpropagation can be used to modify the coefficients. In S1, the calculation of the probability using the NN of FIG. 7 and the correction of the coefficient are repeated to generate a coefficient for outputting the desired probability for the training data set.

Further, the classification model including the generated coefficients is used in S2 and S3 of FIG. In the model evaluation stage (S2) of FIG. 6, for example, using the generated classification model and the test data set, the classification processing unit 144 classifies whether or not ignition has occurred in each electrode and each section. The verification processing unit 146 evaluates the accuracy of the generated model based on the matching rate between the output information and the teacher value. Specifically, of the probabilities that ignition has occurred and the probabilities that ignition has not occurred, for example, the one with the larger probability, which is calculated by the same procedure as in S1, is adopted to determine whether or not ignition has occurred. Classify. Then, for example, the agreement rate between the classification result and the teacher value is calculated, and the accuracy of the generated classification model is evaluated. Further, in the operation stage (S3) of FIG. 6, whether or not the classification processing unit 144 has ignited in each electrode and each section using the classification model generated in S1 and the unlabeled data set. To classify. In S3 as well, of the probabilities that ignition has occurred and the probabilities that ignition has not occurred, for example, the one with the larger probability, which is calculated by the same procedure as in S1, is adopted to classify whether or not ignition has occurred. .. In addition, either the probability that the ignition has occurred or the probability that the ignition has not occurred is obtained, and when the predetermined threshold value is exceeded, it is classified as having the ignition or not having the ignition. May be good.

Further, in each layer of the NN, for example, non-linear separation can be performed by using an arbitrary activation function (transfer function) as appropriate.

<Effect>
According to the configuration of the above-described embodiment, the temporal characteristics of the firing of nerve cells measured by one electrode or the firing of nerve cells measured by a part or all of a plurality of electrodes and the firing of a plurality of electrodes. It becomes possible to learn the synchronic characteristics of firing measured at the same time in part or all, and to classify the presence or absence of firing based on the learned characteristics based on the data of extracellular potential.

<Example>
For the extracellular potential data of nerve cells, we prepared data labeled with burst sections. In addition, the extracellular potential data was divided into sections every 10 milliseconds, and continuous sections including the start point and end point of the burst were extracted as a data set. Then, the continuous section was divided into training data and test data in the order of appearance. The training data was adjusted so that the ratio of burst to non-burst was 1: 1. The 110-minute training data created in this way was trained by a general computer for 5 days (Fig. 6: S1). Then, using the trained model, each section of the test data was classified into burst or non-burst (FIG. 6: S2).

More than 90% of the locations where bursts were found matched the label. Since it is not possible to classify whether or not a burst has occurred by simply setting a threshold value, it can be said that the model after learning in this example was able to learn a classification standard that is difficult to formulate. In addition, there were some parts that did not match the label, where the classification in the classification process was more accurate than the label. In other words, it can be said that the classification errors that can occur in human classification can be reduced. Therefore, according to the classification model generated by the examples, it can be expected that the occurrence of bursts can be discriminated with the same performance even in the classification process (FIG. 6: S3) performed on unknown potential data. According to such a firing classification technique, it becomes possible to improve the efficiency and accuracy of evaluation of the toxicity and efficacy of a drug on nerve cells. In addition, it becomes possible to accurately observe the memory and learning processes of nerve cells.

<Modification example 1>
The structure of the NN is not limited to the example shown in FIG. The above-mentioned NN has one or more intermediate layers between the input layer and the output layer, but the number of layers, the number of units belonging to each layer, and the like can be appropriately changed. Further, among the layers shown in FIG. 7, the NN may be composed of some layers, or the order of the layers may be changed.

For example, the NN including the input layer, the fully connected layer, the probability calculation, and the output layer shown in FIG. 7 may be adopted. For example, in the fully coupled layer, the measured potential data is the input value x in FIG. 9 for each electrode and each section. Further, the probability that the ignition has occurred in the section and the probability that the ignition has not occurred in the section are obtained as the output value y. The value of the input layer, the processing of the probability calculation, and the value of the output layer are the same as those in the above-described embodiment. Further, the fully connected layer may have a configuration in which the configuration shown in FIG. 9 is repeated in a plurality of stages. Further, the fully coupled layer may calculate the converted feature amount using a plurality of potential values measured at the same time by different electrodes or a feature amount based on the potential value, or may be determined by the same electrode. The converted features may be calculated using the values of multiple potentials measured at different time points within the period of, or the features based on the values, or these layers may be repeated in multiple steps. Good.

<Modification 2>
Further, the convolutional NN including the input layer, the convolutional layer, the fully connected layer, the probability calculation, and the output layer shown in FIG. 7 may be adopted. In this case, for example, the convolution layer as shown in FIG. 8 is processed in the same section with the same electrode. Further, the output value of the convolution layer is set as the input value x for each electrode and each section, and the probability that ignition has occurred in the section and the probability that ignition has not occurred in the section are set as the output value y, and the fully coupled layer Is processed. The value of the input layer, the processing of the probability calculation, and the value of the output layer are the same as those in the above-described embodiment. Further, the convolution layer and the fully connected layer may have a configuration in which the configurations shown in FIGS. 8 and 9, respectively, are repeated in a plurality of steps. Further, for the convolution layer and the fully coupled layer, the converted feature amount may be calculated using a plurality of potential values measured at the same time by different electrodes or a feature amount based on the potential value, or the same feature amount may be calculated. The converted features may be calculated using the values of multiple potentials measured at different time points within a predetermined period by the electrodes or the features based on the potential values, or these layers may be repeated in multiple steps. It may be.

<Modification example 3>
Further, the recurrent NN including the input layer, the recurrent layer, the probability calculation, and the output layer shown in FIG. 7 may be adopted. Further, in the configuration shown in FIGS. 10 and 11, for example, a plurality of potential data are set as input values x for each electrode. The value of the input layer, the processing of the classification layer, and the value of the output layer are the same as those in the above-described embodiment. Further, the recurrent layer may have a configuration in which the configurations shown in FIGS. 10 and 11 are repeated in a plurality of stages. Further, the recurrent layer may calculate the converted feature amount using a plurality of potential values measured at the same time by different electrodes or a feature amount based on the potential value, or may be determined by the same electrode. The converted features may be calculated using the values of a plurality of potentials measured at different time points within the period or the features based on the potential values, or these layers may be repeated in a plurality of steps. ..

<Modification example 4>
Further, the convolution recurrent NN including the input layer, the convolution layer, the recurrent layer, the probability calculation, and the output layer shown in FIG. 7 may be adopted. In this case, for example, the convolution layer as shown in FIG. 8 is processed in the same section with the same electrode. Further, in the configuration shown in FIGS. 10 and 11, for example, a plurality of potential data are set as input values x for each electrode. The value of the input layer, the process of calculating the probability, and the value of the output layer are the same as those in the above-described embodiment. Further, the convolution layer and the recurrent layer may have a configuration in which the configurations shown in FIGS. 8, 10 and 11, respectively, are repeated in a plurality of stages. Further, for the convolution layer and the recurrent layer, the converted feature amount may be calculated using a plurality of potential values measured at the same time by different electrodes or a feature amount based on the potential value, or the same electrode may be used. The converted feature quantity may be calculated using the value of a plurality of potentials measured at different time points within a predetermined period or the feature quantity based on the potential value, or these layers may be repeated in a plurality of steps. You may.

<Modification 5>
Further, the configuration may have more layers than the NN shown in FIG. 7. For example, N4 in FIG. 7 may be followed by a recurrent layer N6 (not shown) and a fully connected layer N7 (not shown).

For example, in the recurrent layer N6, the feature amount calculation unit 1421 calculates the feature amount of the data before and after in the forward direction in the time series with respect to the feature amount calculated in N4 of FIG. 7, and the time series. The features are calculated by the recurrent layer that calculates the features of the data before and after in the upper and reverse directions. That is, in the recurrent layer that calculates the features of the data before and after in the forward direction, for example, the "state" shown in FIG. 11 is on the time series from the state of processing the features of the previous section on the time series. It is configured to take over the information toward the state of processing the features of the later section. Further, in the recurrent layer that calculates the features of the data before and after in the opposite direction, for example, the "state" shown in FIG. 11 is on the time series from the state of processing the features of the later section on the time series. It is configured to take over the information toward the state of processing the features of the previous section. With such a configuration, between the temporal characteristics of the long-term firing pattern between the multiple sections found in the past-to-future direction and the multiple sections found in the future-to-past direction. You will be able to learn the temporal characteristics of the long-term firing pattern of.

Further, in the fully connected layer N7, the feature amount calculation unit 1421 integrates the forward and reverse feature amounts calculated in N6 for each corresponding section, and calculates the feature amount by the fully connected layer.

FIG. 12 is a diagram for explaining the integration process in N7. The output of the forward recurrent layer, which is the output of N6, includes the feature amounts F11, F12 ... Of the first section, the feature amounts F21, F22 ... Of the second section, and the like. Further, the output of the reverse recurrent layer includes the feature amounts R11, R12 ... Of the first section, the feature amounts R21, R22 ... Of the second section, and the like. In this case, in the integration process in N7, the output of the forward recurrent layer and the output of the reverse recurrent layer are combined for each section, and the features F11, F12 ... R11, R12 ... Of the first section are combined. A matrix including the feature quantities F21, F22 ... R21, R22 ... Of the second section is output. In this way, in N7, the process of integrating the features in the forward direction and the features in the reverse direction is performed.

In addition, for the output after integration, the feature amount by the fully connected layer is calculated. The structure of the fully connected layer is the same as that in FIG. Further, in N5 of FIG. 7, based on the feature amount in the same section calculated in N7, the probability that ignition has occurred in the section and the probability that ignition has not occurred in the section are calculated for each electrode. Ask. In the case of one electrode, the above probability is calculated with respect to the potential data measured by the electrode. Further, in the case of a plurality of electrodes, the probability that one set of ignition has occurred and the probability that ignition has not occurred may be calculated for all the electrodes, or the probability that ignition has occurred for each electrode. And the probability that no ignition has occurred may be calculated.

It should be noted that, of the plurality of continuous sections to be processed, the classification results near the front and rear ends do not sufficiently reflect the temporal characteristics in the forward or reverse direction. Therefore, the classification result may not be adopted for the sections near the front and rear ends.

<Others>
The present invention includes a computer program that executes the above-mentioned processing and a computer-readable recording medium that records the program. The recording medium on which the program is recorded can perform the above-mentioned processing by causing the computer to execute the program.

Here, the computer-readable recording medium means a recording medium in which information such as data and programs is stored by electrical, magnetic, optical, mechanical, or chemical action and can be read from a computer. Among such recording media, those that can be removed from a computer include flexible disks, magneto-optical disks, optical disks, magnetic tapes, memory cards, and the like. Further, as the recording medium fixed to the computer, there are HDD, SSD (Solid State Drive), ROM and the like.

1: Measuring device 11: Chip holder 111: Chip 1111: Measuring electrode 1112: Reference electrode 1113: Wiring 12: Potential measuring device 13: Monitor 14: Machine learning device 141: Potential data storage unit 142: Machine learning unit 1421: Feature quantity Calculation unit 1422: Learning processing unit 143: Classification model storage unit 144: Classification processing unit 145: Classification result storage unit 146: Verification processing unit

Claims (11)

Regarding the data representing the time course of the extracellular potential of a nerve cell, the value of the potential measured in a predetermined period and the occurrence of ignition during the period in which the user classifies whether or not the data has been ignited. It is a machine learning device that learns the relationship between the presence or absence of ignition and the value of the electric potential by using a teacher value representing.
A feature amount calculation unit that calculates a feature amount based on the potential value and a predetermined function,
A learning processing unit that modifies the coefficient used in the function by using the value based on the feature amount calculated by the feature amount calculation unit and the teacher value.
A machine learning device equipped with. The feature amount calculation unit takes the value of the potential or the feature amount based on the potential value as an input, and all the nodes on the input side are connected to all the nodes on the output side to the fully connected layer and a predetermined coefficient associated with each connection. The machine learning device according to claim 1, wherein the converted feature amount is calculated based on the above. The feature amount calculation unit is a function that takes the potential value or a feature amount based on the potential value as an input and uses a value corresponding to the processing result of the node in the previous stage for the processing of the node in the subsequent stage for a series of nodes on the input side. The machine learning device according to claim 1 or 2, wherein the feature amount after conversion is calculated by a recurrent neural network that performs processing based on. The machine learning according to any one of claims 1 to 3, wherein the feature amount calculation unit calculates the feature amount after conversion by a predetermined convolutional neural network by inputting the potential value or the feature amount based on the potential value. apparatus. The value of the extracellular potential is measured via a plurality of electrodes arranged so as to be in contact with the nerve cell, and the feature amount calculation unit is the value of the plurality of potentials measured at the same time by different electrodes or the value thereof. The machine learning apparatus according to any one of claims 1 to 4, wherein the converted feature amount is calculated by using the feature amount based on the feature amount. The value of the extracellular potential is measured via one or a plurality of electrodes arranged so as to be in contact with the nerve cell, and the feature amount calculation unit is measured by the same electrode at different time points within the predetermined period. The machine learning apparatus according to any one of claims 1 to 4, wherein the converted feature amount is calculated by using the plurality of potential values or the feature amount based on the potential value. A classification device that classifies whether or not the ignition has occurred by using the coefficient corrected by the learning processing unit included in the machine learning device according to any one of claims 1 to 6 and the value of the potential. .. Regarding the data representing the time course of the extracellular potential of a nerve cell, the value of the potential measured in a predetermined period and the occurrence of ignition during the period in which the user classifies whether or not the data has been ignited. It is a machine learning method in which the relationship between the presence or absence of ignition and the value of the electric potential is learned by a computer using a teacher value representing.
The feature amount is calculated based on the potential value and a predetermined function.
The coefficient used for the function is modified by using the calculated value based on the feature amount and the teacher value.
Machine learning method. A classification method for classifying whether or not the ignition has occurred by using the coefficient modified by the machine learning method according to claim 8 and the value of the electric potential. Regarding the data representing the time course of the extracellular potential of a nerve cell, the value of the potential measured in a predetermined period and the occurrence of ignition during the period in which the user classifies whether or not the data has been ignited. It is a machine learning program that causes a computer to learn the relationship between the presence or absence of ignition and the value of the electric potential by using a teacher value representing.
The feature amount is calculated based on the potential value and a predetermined function.
The coefficient used for the function is modified by using the calculated value based on the feature amount and the teacher value.
Machine learning program. A classification program for classifying whether or not the ignition has occurred by using the coefficient modified by the machine learning program according to claim 10 and the value of the electric potential.

Images