WO2018092957A1 - Method, device and program for determining for re-learning with respect to input value in neural network model - Google Patents

Abstract

The present invention relates to a method, a device and a computer program for determining for re-learning with respect to unknown data in a neural network model. To this end, provided is a method, for determining for re-learning with respect to an input value in a neural network model, comprising: a training set clustering step for clustering with respect to a training set used for the learning of a learned neural network model; an input value plotting step for plotting an input value, which is input into the neural network model, to the clustering distribution of the training set; an input value classifying step for determining whether or not the input value is comprised in a particular cluster which is one of the clustering distribution of the training set; an input value re-learning determining step for determining for the input value re-learning or re-learning weight on the basis of the distance between the input value and a centroid value of the particular cluster or the probability of the input value being comprised in the particular cluster; and a neural network model re-learning step for performing the re-learning of the model on the basis of the input value re-learning or re-learning weight which have been determined in the input value re-learning determining step. Accordingly, the neural network model can be learned for each user even when data which is not refined in a time-series manner is input.

Description

Methods, devices, and programs for determining whether to relearn inputs in neural network models

The present invention relates to a method, apparatus and program for determining whether to relearn unknown data in neural networks.

The neural network model is a supervised learning algorithm conceived from neuronal structures in biology. The basic principle of neural network modeling is to interconnect several neurons to predict the optimal output for the input. From a statistical point of view, the neural network model can be viewed as a projective tracking regression that takes a nonlinear function on the linear combination of input variables.

1 is a schematic diagram showing the structure of a multilayer neural network model (deep learning or deep neural network model). As shown in FIG. 1, the multi-layer neural network model includes an input layer, a hidden layer, and an output layer. The input layer is composed of nodes corresponding to each input variable, and the number of nodes is equal to the number of input variables. The hidden layer processes the linear combination of variable values from the input layer as a nonlinear function such as a sigmoid function, and delivers it to the output layer or another hidden layer. (Recently, a chain rule is applied in the back propagation. The problem of vanishing gradient dilution occurs in ReLU instead of sigmoid function. The output layer is a node corresponding to the output variable. In the classification model, as many output nodes as the number of classes are generated.

2 is a schematic diagram showing a calculation process at a node. As shown in Figure 2, the computation actually occurs at each node, which is mathematically designed to simulate the processes occurring in the neurons that make up the human neural network. When a node receives a stimulus above a certain magnitude, it responds. The magnitude of the response is roughly proportional to the product of the input value plus the node's coefficients (or weights), except for the bias value. In general, a node receives multiple inputs and has as many coefficients as the number of inputs. Thus, by adjusting this coefficient, different weights can be given to different inputs. The final multiplications are added together and the sum is entered into the activation function. The result of the active function corresponds to the output of the node, which is ultimately used for classification or regression analysis. Each layer of the neural network model consists of at least one node, and whether each node is activated or deactivated depends on the input value. The input data is the input of the first layer (input layer), after which the output of each layer is again the input of the next layer. All coefficients change little by little over the course of the lesson, reflecting which input each node counts as important. And the 'training' of the neural network model is the process of updating this coefficient.

The most problematic in multilayer neural network models, namely deep learning, is the overfitting problem. Overfitting occurs when the complexity of the model is high compared to the amount of data given. Unfortunately, as the neural network deepens, the complexity of the model increases exponentially. Thus, despite the fact that everybody knows a good deep network for learning almost infinite phenotypes, the overfitting is so bad that neural network research was stopped by Marvin Minsky's 1969 Perceptrons theory. However, as a new initialization method to prevent overfitting was proposed in 2007-2008, deep learning research was actively conducted again, due to the proposed RBM (Restricted Boltzmann Machine) and CNN (Convolutional Neural Network).

In particular, RBM is used as a component of DBN (Deep Belief Network), and overfitting by effectively pre-training each layer of feedforward neural network to be learned through unsupervised Restricted Boltzmann machine (RBM). We've got a level of initialize point that prevents this from happening, and then we're learning in the form of using supervised back propagation.

However, unsupervised pretraining, such as RBM and DBN, was proposed in two papers by Professor Bengio's team published in NIPS 2006, Greedy layer-wise training of deep networks, and Hinton's team published in NIPS 2007. The methods are not used in recent practice. This is because it is now known that random initialization performs much better than initializing weights this way if there is enough data. As part of random initialization, the concept of drop-out was introduced, and recently, most of them use the drop-out method.

3 is a schematic diagram showing a drop-out method. As shown in FIG. 3, the drop-out method uses about 50% of the neurons instead of all the neurons in the hidden layer at each learning. It is known that in one deep learning, several small neural networks are ensembleed, and the ensemble is greatly reduced in overfitting. In addition, since neurons having similar weights are reduced, the neurons that make duplicate judgments are reduced, and thus the neurons can be efficiently used.

Recently, an activation function called ReLU (Rectified Linear Unit) has been used to solve slow learning time and overfitting problems. 4 is a graph showing the ReLU activity function. As shown in FIG. 4, the conventional sigmoid function causes an error to disappear every time the gradient descent is performed in several layers. As the extremes go through several layers, the sigmoid function becomes less Gradient, causing the weight not to update. However, when using the ReLU function as an active function, when the slope is learned to be 0 or 1, the error is propagated to 100% and this problem is solved. The ReLU function is not limited to [0,1] like sigmoid, and because of its unlimited range, it can be considered more expressive. In addition, there are many unnecessary values among the output values of each node. In this case, all values should be calculated when using sigmoid function. ReLU function can reduce the amount of computation and improve computing speed. Regularization can be improved by the ReLU function.

However, when back propagation is performed in ANN (BPN, back propagation neural network), if the unclassified unknown data is used as a training set again, the model is more likely to be distorted during parameter learning. Apart from the problem of not finding a global minimum and falling into a local minima, the weight of the model is incorrectly trained as the neural network model is overfitting on too random training data. This problem is particularly a problem when neural network models need to be trained for each user by inputting unrefined data in time series, and cannot be completely solved by RBM, CNN, Drop-out, and ReLU methods.

For example, when the user's sitting posture is determined by a pressure sensor distributed in a plane having a plurality of rows and columns, the posture is different for each user, so that specific unknown pressure distribution data can be clearly classified by the cut off value. Often, difficult input values (pressure distribution data) are displayed. In particular, the pressure distribution data is input to the neural network model in time series while the user is sitting, so that the pressure distribution data in the unrefined state may be significantly input and learned. In the case of back propagation up to such unknown pressure distribution data, the predictability of the model can be significantly reduced.

In addition, when the sitting posture of the user is determined by a pressure sensor distributed in a plane having a plurality of rows and columns, the posture information output from the pressure distribution data passing through the neural network model should be learned differently for each user. That is, even if the same pressure distribution data is displayed, the user A may be in a bent posture, and the user B may be in a correct posture. This is because the user's body type and posture are different.

Alternatively, back propagation is frequently used for image analysis mainly using CNN, natural language analysis mainly using Recurrent Neural Network (RNN), image caption, and automatic translation. However, input values that are difficult to classify are often used for relearning, which often hinders the performance of the model.

Accordingly, the present invention has been devised to remedy the problems presented above.

An object of the present invention is to use the input value in the neural network model to re-learn the neural network model to update the weight, and use the input for re-learning such as back propagation of the neural network model using unsupervised learning such as clustering. The present invention provides a method, apparatus and program for determining whether to re-learn input values in a neural network model that selects a value and updates a centroid of a cluster to which the corresponding input value belongs to improve the performance of the neural network model.

Hereinafter, specific means for achieving the object of the present invention will be described.

SUMMARY OF THE INVENTION An object of the present invention is a method for determining whether to relearn input values in a neural network model performed by a computing device, wherein the computing device performs clustering on a training set used for learning a previously learned neural network model. Training set clustering step; An input value plotting step of the computing device plotting an input value input to the neural network model to a clustering distribution of the training set; An input value classification step of determining, by the computing device, whether the input value is included in a specific cluster which is one of clustering distributions of the training set; An input value for determining, by the computing device, whether or not to re-learn the input value or the re-learning weight based on a distance between the input value and a centroid of the specific cluster or a probability that the input value is included in the specific cluster Determining whether to relearn; And a neural network model relearning step, wherein the computing device re-learns the neural network model based on the re-learning weight or the re-learning weight of the input value determined in the re-learning of the input value. This can be achieved by providing a method for determining whether to relearn input values in the model.

In addition, after the neural network model relearning step, the computing device may determine a centroid of the particular cluster based on relearning or relearning weight of the input value determined in the re-learning of the input value. The cluster center value updating step of updating may be further included.

Further, in the training set clustering step and the input value floating step, clustering may be performed by performing linear mapping through K-means, spectral clustering or N-cut or non-linear mapping based on an auto encoder.

Another object of the present invention, the receiving unit for receiving an input value; A neural network model computing processor that receives the input value from the receiver, includes a neural network model that has been learned, and calculates an output value by applying the received input value to the neural network model; A clustering processor configured to generate clustering information by receiving the input value from the receiving unit and clustering a training set used for learning the neural network model among the input values; The clustering processor receives the clustering information, and re-learning the input value based on a probability of the input value received at the receiving unit being included in a specific cluster or a distance from a centeroid of a specific cluster of the clustering information. A re-learning determination processor for determining whether or not a re-learning weight applied to the input value; And re-learning whether the input value is re-learned or the re-learning weights, and re-learning the neural network model of the neural network model computing processor to update the neural network model of the neural network model computing processor. It can be achieved by providing an apparatus for determining re-learning of input values in a neural network model including a learning processor.

The relearning processor may update a centeroid of a specific cluster of the clustering information.

In addition, the clustering processor may be characterized by performing clustering by performing linear mapping or auto encoder based non-linear mapping through K-means, spectral clustering or N-cut.

Another object of the present invention is to perform clustering on a training set used for learning a neural network model that has been learned in a program stored in a recording medium that performs a re-learning method for input values in a neural network model. Training set clustering step; Plotting an input value inputted to the neural network model to a clustering distribution of the training set; An input value classification step of determining whether the input value is included in a specific cluster which is one of clustering distributions of the training set; Determining whether to re-learn the input value or re-learning weight based on the distance between the input value and the centroid of the specific cluster or the probability that the input value is included in the specific cluster. ; And a neural network model relearning step of re-learning the neural network model based on re-learning or re-learning weight of the input value determined in the re-learning whether the input value is re-learned. It can be achieved by providing

As described above, the present invention has the following effects.

First, according to an embodiment of the present invention, even when data that is not purified in time series is input, an effect of learning a neural network model for each user is generated.

Second, according to one embodiment of the present invention, the neural network model is overfitting on too random training data, thereby preventing the weight itself of the model from being incorrectly learned.

Third, according to an embodiment of the present invention, it is possible to improve the performance of the neural network model to an individual by updating a centroid of a cluster to which an input value determined to be relearned belongs.

The following drawings, which are attached to this specification, illustrate preferred embodiments of the present invention, and together with the detailed description thereof, serve to further understand the technical spirit of the present invention. It should not be interpreted.

1 is a schematic diagram showing the structure of a multilayer neural network model (deep learning or deep neural network model),

2 is a schematic diagram showing a calculation process at a node;

3 is a schematic diagram showing a drop-out method;

4 is a graph showing the ReLU activity function,

5 is a diagram illustrating a re-learning method for input values in a neural network model according to an embodiment of the present invention.

6 illustrates an example of determining whether to relearn an input value in step S14 according to an embodiment of the present invention.

7 illustrates an example of cluster-centric updating of a training set in step S15 according to an embodiment of the present invention.

8 is a schematic diagram showing an apparatus for determining re-learning of input values in a neural network model;

9 illustrates a posture classification system using pressure distribution information according to an embodiment of the present invention.

10 is an example of a screen displayed on the user client of the posture classification system according to an embodiment of the present invention,

11 is a flowchart illustrating a posture classification method using pressure distribution information according to an embodiment of the present invention;

12 is a graph of an experimental result of a posture classification method using pressure distribution information according to an embodiment of the present invention;

13 is a graph showing the results of an experiment on a posture classification method using pressure distribution information of a conventional relearning method;

14 is a graph comparing average accuracy of a posture classification method using a pressure distribution information and a conventional method according to an embodiment of the present invention.

DETAILED DESCRIPTION Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. However, in describing in detail the principle of operation of the preferred embodiment of the present invention, if it is determined that the detailed description of the related known function or configuration may unnecessarily obscure the subject matter of the present invention, the detailed description thereof will be omitted.

In addition, the same reference numerals are used for parts having similar functions and functions throughout the drawings. Throughout the specification, when a particular part is connected to another part, this includes not only the case where it is directly connected, but also the case where it is indirectly connected with another element in between. In addition, the inclusion of a specific component does not exclude other components unless specifically stated otherwise, it means that may further include other components.

In the following detailed description, it is assumed that computing is performed by putting input values into a neural network model in a user's client. However, the scope of the present invention is not limited thereto. The present invention is configured to receive inputs from a user's client from a device for generating inputs (e.g., near field communication), send inputs from a user client to a computing server, and compute the inputs into a neural network model at the computing server. May be That is, since the subject of computing may be a user client or a computing server, the subject of computing is referred to as a computing device in the present invention.

Determining Whether to Relearn Inputs in Neural Network Models

Regarding the re-learning method for input values in the neural network model according to an embodiment of the present invention, FIG. 5 illustrates a re-learning method for the input value in the neural network model according to an embodiment of the present invention. will be. As shown in FIG. 5, in the neural network model according to an embodiment of the present invention, a re-learning method for input values may include a neural network model learning step, an input value input step, an output value output step, an input value classification step, and an input value. Re-learning determination step, input value re-learning step may be included. In the neural network model according to an embodiment of the present invention, the re-learning method for input values may be performed by a computing server or a client.

The neural network model training step (S10) is a step of learning weights of the neural network model using a training data set for which labeling or tagging of attribute values is completed. In this case, the neural network model may include not only an existing artificial neural network (ANN) but also CNN, RNN, and the like. In particular, the neural network model according to an embodiment of the present invention may be a classification model.

Input value input step S11 is a step of inputting a specific input value which is a test set or an actual data set to an input layer of the neural network model.

The output value output step S12 is a step in which a specific input value input in S11 to the neural network model passes through a hidden layer and is output from an output layer. In S12, the specific input value is weighted by the neural network model and output as a matrix or vector output value having specific property values for the predetermined properties.

The input value classification step S13 is a step of classifying the training set of S10 by clustering or the like and plotting the input value of S11. At this time, prior to clustering, PCA analysis is performed with two or three attributes on the training set and input values, or K-means, spectral clustering, linear mapping through N-cut, and non-linear mapping based on Auto encoder. Can be clustered by

The input value classification step S13 may be specifically performed as a training set clustering step, an input value floating step, and an input value classification step. The training set clustering step is a step in which a computing device such as a user client or a server performs clustering on a training set used for learning a neural network model that has been previously learned. The input plotting step is a step in which a computing device such as a user client or a server plots an input value input to the neural network model into a clustering distribution of the training set. The input value classification step is a step in which a computing device such as a user client or a server determines whether the input value is included in a specific cluster which is one of the clustering distributions of the training set.

Clustering of a training set may be performed based on distance or probability (or probability distribution). K-means clustering is representative of distance-based clustering that can be used in S13. K-means assumes that data belonging to the same cluster is 'close' to each other. At this time, there is one 'center' for each cluster, and 'how close to each center' is defined as cost. K-means is an algorithm that finds clusters that reduce the cost as defined. If the formula is written, the following equation is obtained.

In Equation 1, there are n data and k clusters. In this case, b _j means 'center' of the j th cluster, and w _ij is a binary variable having 1 or 0 if the i th data belongs to the j th cluster. Also, the following condition is a constraint that must be assigned to one cluster for each data. The K-means algorithm is an algorithm that repeats the calculation until b and w no longer change while continuously updating b and w. Unfortunately, K-means does not converge to the global optimum, but converges to the local optimum, which is very vulnerable to initialization. So k-means clustering tends not to be used for data preprocessing of neural network models. However, in the method of determining whether to re-learn an input value according to an embodiment of the present invention, clustering is used to determine whether or not to re-learn the input value instead of using clustering as a process of preprocessing the input value by unsupervised learning. Therefore, local minimum can help to improve neural network model performance.

[Revision 16.12.2016 under Rule 91]
A Gaussian Mixture Model (GMM) is typical for probability or probability distribution based clustering that can be used in S13. Gaussian Mixture Model, Mixture of Gaussian, GMM, or MoG assumes that the data consists of 'Mixture' of 'Gaussian'. Usually called GMM, it means the algorithm to find the most optimal GMM. In other words, if the data consists of k Gaussians, the algorithm finds the k averages and covariances that best describe the data.

[Revision 16.12.2016 under Rule 91]

GMM's performance measure is a log likelihood function. In other words, the goal is to find a parameter that gives the largest probability of data X for a given parameter. The log likelihood is defined as ln p (X | θ). The parameter θ consists of the average μj of each Gaussian, the covariance Σj, and the probability πj that each data belongs to each Gaussian. Therefore, if the multinomial Gaussian distribution of xi for a given μj, Σj is defined as N (xi | μj, Σj), the log likelihood function can be described as the following equation.

The most popular algorithm used to solve GMM is EM algorithm. The EM algorithm for solving the GMM first calculates the expectation by updating the information (e.g. z) on which cluster each data belongs to, and then uses the updated information to determine the most log likelihood of the remaining values. Find the parameters to maximize (μ, Σ, π). If you run the algorithm, it will find a better value each time iteration passes.

The re-learning of input value determination step (S14) determines whether or not an input value can be included in each cluster classified in S13, and if the input value is included in a specific cluster, inputs it back into the training set, and input value If it is not included in this particular cluster, it is vanished. That is, the step of determining whether to re-learn the corresponding input value based on the distance or the probability (or probability distribution) in the centroid of the specific cluster. In this case, in the case of re-learning, it is possible to determine whether the output value is included in a specific cluster by setting a specific cut off value based on distance in case of K-means clustering, and based on probability distribution in case of GMM (Gaussian Mixture Model). For example, whether the input value is included in a specific cluster may be determined based on whether or not it corresponds to a specific σ.

6 illustrates an example of determining whether to relearn an input value at step S14 according to an embodiment of the present invention. As shown in FIG. 6, whether the input value corresponds to a specific cluster of the clustered training set determines whether to relearn the input value. For example, in FIG. 6, the input value a may be classified as being included in the cluster A, and thus it is determined to relearn, and the input value b is not classified in the clusters A and B, so it is determined not to be relearned.

The input value relearning step (S15) is a step of re-learning the input value determined to be relearned in S 14 by a method such as back propagation on the neural network model. According to an embodiment of the present invention, the input determined to be relearned by S15 is included in the training set of S10, and the centroid of each cluster of S13 is updated. As the re-learning is repeated by the cluster-centric update through S14 and S15, an attribute of a specific cluster is enhanced (different from other users' data) for each user. This helps to build a customized neural network model in the following embodiments.

7 is a diagram illustrating cluster-based updating of a training set in step S15 according to an embodiment of the present invention. As shown in FIG. 7, the center of the cluster is updated by the input value to be relearned, and as the iteration is performed, the performance improvement of the neural network model may occur.

In the neural network model according to an embodiment of the present invention, the re-learning method for input values may be configured by a computer program stored in a recording medium to be performed by a computing device.

In the neural network model according to another embodiment of the present invention, the method of re-learning whether or not an input value belongs to a specific cluster according to the distance or probability with respect to the center value (centeroid) of the specific cluster of the input value in step S14. It can be performed by defining a value called re-learning weights for the target and re-learning by applying the re-learning weights to the input values. According to another exemplary embodiment of the present invention, re-learning weights are assigned to input values, and all of them are put in a training set to be relearned. Vanishing an input value that is not included in a particular cluster according to one embodiment of the present invention may correspond to giving a re-learning weight of zero to the input value in another embodiment of the present invention.

Device for determining re-learning of input value in neural network model

Regarding the apparatus for determining re-learning of input values in a neural network model according to an embodiment of the present invention, FIG. 8 is a schematic diagram illustrating an apparatus for determining re-learning of input values in a neural network model. As shown in FIG. 8, the apparatus for determining re-learning of an input value in the neural network model may include a receiver 100 and a processor.

An apparatus for determining re-learning of input values in a neural network model according to an embodiment of the present invention may include various types of computing devices. For example, in the neural network model according to an embodiment of the present invention, the apparatus for determining re-learning of input values may be a server or a client. In this case, the input value received as a training set can be trained using ANN including CNN, RNN, etc. Based on this, the input value received later as a test set is applied to the neural network model to improve the performance of the model. Can be verified Subsequently, when an input value, which is an actual data set, is applied to the neural network model, an output value corresponding to the same is output.

The receiver 100 may obtain an input value composed of a scalar value, a vector value, or the like as a training set, a test set, or an actual data set. The receiver 100 may be implemented by various communication technologies, and may receive an input value input as a neural network model.

The processor may include a sampling processor 101, a neural network model computing processor 102, a neural network model training processor 103, a data preprocessor 104, a clustering processor 105, a relearning determination processor 106, a relearning processor ( 107, a storage processor 108.

The sampling processor 101 is configured to sample at least some of the input values and divide the training set into a training set and a test set.

The neural network model computing processor 102 is configured to derive an output value by inputting input values divided into test sets in the sampling processor 101 to the neural network model that has been learned. In addition, the neural network model computing processor 102 is configured to calculate an output value by applying an input value to the neural network model that has been learned by the receiver 100 when an actual data set is input as an input value.

The neural network model training processor 103 uses a data set divided into a training set in the sampling processor 101 or a training set including input values determined to be relearned among input values input to the neural network model computing processor 102. It is configured to update and learn the weight of the neural network model.

When the input value is obtained, the data preprocessing processor 104 may perform a process of converting or converting the input value into matrix data.

The clustering processor 105 is a configuration for clustering a training set to determine whether an input value of a test set or an actual data set is included in a specific cluster.

Re-learning determination processor 106 determines to re-learn that input if it is included in a particular cluster, and not to relearn that input if it is not included in a particular cluster. Configuration. According to another embodiment of the present invention, the re-learning determination processor 106 may be configured to relearn the input value by weighting it according to how close the input value is to a specific cluster.

The relearning processor 107 inserts the input value determined to be relearned into the training set, updates the weight of the neural network model by a model learning method such as back propagation, and updates the center of the cluster including the input value. Configuration.

The storage processor 108 is configured to store input values determined not to be relearned in the server.

Hereinafter, an apparatus for determining re-learning of input values in a neural network model according to an embodiment of the present invention will be described based on a pressure sensing device and a method using pressure distribution information. The scope of the present invention is not limited to posture classification using pressure distribution information, and may include those applied to general applications solved by ANN such as image caption and search using natural image information, natural language analysis, and the like.

Posture classification system and device using pressure distribution information

9 illustrates a posture classification system using pressure distribution information according to an embodiment of the present invention. As shown in FIG. 9, the posture classification system using the pressure distribution information includes a pressure sensing device 1, a pressure distribution sensor, a transceiver 100, a power supply unit, a server 6, a DB 7, and a user client 8. ) May be included.

The pressure sensing device 1 is mounted at the user's seating position and is configured to sense and output a pressure distribution when the user sits on the pressure sensing device. The pressure sensing device 1 may include a pressure distribution sensor 2, a transmitter 3, a power supply 4, and a controller 5.

The pressure distribution sensor 2 may refer to a configuration in which a plurality of piezoelectric sensors or pressure sensors configured in a flat or curved surface at a user's seating position are distributed in a lattice form. In the pressure distribution sensor 2, pressure values generated by each piezoelectric sensor or pressure sensor may be converted to generate pressure distribution information formed of image or matrix data.

The transmitter 3 is configured to transmit the pressure distribution information generated by the pressure distribution sensor 2 to the server 6 or the user client 8 at a specific time interval requested by the controller 5. When the transmitter 3 transmits the pressure distribution information to the user client 8, it can transmit by near field communication, such as BLE.

The power supply unit 4 is configured to supply power to the pressure distribution sensor 2, the transmission unit 3, and the control unit 5.

The control part 5 is a structure which controls the sensing duration of the pressure distribution sensor 2, transmission of the pressure distribution information of the transmitter 3, etc. FIG.

The server 6 receives the pressure distribution information directly from the pressure sensing device 1 or through the user client 8, and receives the label information for the specific pressure distribution information from the user client 8 for training. It constructs a set, the neural network model is trained using the training set, and the posture information of the user is classified using the neural network model and outputted to the user client 8.

The DB 7 is connected to the server 6 to store the pressure distribution information received from the pressure sensing device 1 in chronological order, and connect the label information input from the user client 8 to the corresponding pressure distribution information. It is a configuration to save.

The user client 8 is connected to the server 6 and receives the pressure distribution information from the server 6 or inputs label information (in which position it is sitting) to the server 6 for the specific pressure distribution information. It is configured to transmit and display the posture information of the user classified by the server 6. According to an embodiment of the present invention, when the user client 8 serves as an access point of the server 6 and the pressure sensing device 1, the pressure distribution information may be directly received by the pressure sensing device 1. In addition, the received pressure distribution information may be transmitted to the server 6 to classify and predict the posture information of the user using a neural network model. Alternatively, if the computing power of the user client 8 is sufficient, the neural network model can be computed directly at the user client 8.

10 illustrates an example of a screen displayed on a user client of a posture classification system according to an embodiment of the present invention. As shown in FIG. 10, the received pressure distribution information 30 may be encoded and displayed, or the posture information 31 of the classified user may be transmitted and displayed.

Posture classification method using pressure distribution information

11 is a flowchart illustrating a posture classification method using pressure distribution information according to an embodiment of the present invention. The flowchart of FIG. 11 shows a case where the computing of the neural network model is performed at the user client 8, and whether re-learning is determined and the cluster update is performed at the server 6. As shown in FIG. 11, the posture classification method using the pressure distribution information according to an embodiment of the present invention includes a pressure distribution information generation step S20, a pressure distribution information transmission step S21, and a posture information output step S22. ), The output posture information transmitting step (S23), pressure distribution information floating step (S24), re-learning whether step (S25), re-learning step (S26), weight update step (S27), cluster update step (S28) It may include.

Pressure distribution information generation step (S20) is a step of generating the pressure distribution information according to the user's seating in the pressure sensing device (1).

The pressure distribution information transmitting step S21 is a step in which the pressure sensing device 1 transmits the pressure distribution information to the user client 8 using short-range communication.

The posture information output step S22 is a step of applying the neural network model learned by using the pressure distribution information received from the user client 8 as a test set, and outputting the posture information.

The pressure distribution information transmitting step S23 is a step of transmitting the pressure distribution information to the server 6 in the user client 8.

Pressure distribution information plotting step (S24) is a step of plotting the pressure distribution information received in the cluster information of the training set clustered on the basis of distance or probability in the server (6).

The re-learning determination step (S25) is a step of determining whether or not re-learning through whether or not the pressure distribution information is included in the specific cluster in the server (6).

The re-learning step (S26) is a step of re-learning the weight of the neural network model by the back propagation method on the pressure distribution information determined to be re-learned by the server 6.

The weight updating step S27 is a step of transmitting the weight of the neural network model retrained by the server 6 to the user client 8.

The cluster update step (S28) is a step of updating the centroid of the cluster in the cluster information of the training set by using the pressure distribution information determined by the server 6 to relearn.

According to another embodiment of the present invention, the re-learning weight may be determined in the pressure distribution information in S25. The re-learning weight may be defined to determine how far (distance or probability) the corresponding pressure distribution information is from the centroid of a particular cluster. In addition, according to another embodiment of the present invention, the weight of the neural network model may be re-learned in the state where the re-learning weight is applied to the pressure distribution information in S26. For example, if a specific input is included in a specific cluster, a weight of 1 affects the relearning of the neural network model, and if a particular input is not included in a particular cluster, a weight of 0 is used to affect the neural network model relearning. Can be.

12 is a graph showing an experiment result of a posture classification method using pressure distribution information according to an embodiment of the present invention. In FIG. 12, the X axis represents evolution, and the Y axis represents accuracy (%). 12 is an experimental example using a model in which nine postures are learned in a chair provided with 64 pressure sensors, and after re-learning the model according to an embodiment of the present invention using an input value of evolution n, This is an experiment of evolution n + 1. As shown in FIG. 12, according to an embodiment of the present invention, it can be seen that classification accuracy of 80% or more comes out in all postures after evolution 3.

FIG. 13 is a graph of an experiment result of a posture classification method using pressure distribution information of a conventional relearning method. As shown in FIG. 13, when re-learning all data including noise without re-learning according to an embodiment of the present invention, it can be seen that classification accuracy of 80% or more appears in all postures after evolution 6. have.

14 is a graph comparing average accuracy of a posture classification method using a pressure distribution information and a conventional method according to an embodiment of the present invention. In FIG. 14, "noised" means an existing method, and "cleaned" means a method including a re-learning determining step according to an embodiment of the present invention. As shown in Figure 14, it can be seen that the model in accordance with an embodiment of the present invention shows a much faster rise in accuracy. In other words, it can be seen that the case in which the performance according to the embodiment of the present invention can produce a better performance with less evolutionary data than before. According to an embodiment of the present invention according to an embodiment of the present invention is guaranteed an average of 80% accuracy since evolution 1, and according to the conventional method can be confirmed that the average of 80% accuracy since evolution 3 is guaranteed.

As described above, those skilled in the art will appreciate that the present invention can be implemented in other specific forms without changing the technical spirit or essential features. Therefore, the above-described embodiments are to be understood in all respects as illustrative and not restrictive. The scope of the present invention is shown by the following claims rather than the detailed description, and it should be construed that all changes or modifications derived from the meaning and scope and equivalent concept of the claims are included in the scope of the present invention.

Claims (7)

In the neural network model performed by the computing device to determine whether to re-learning the input value, A training set clustering step, wherein the computing device performs clustering on a training set used for learning a previously learned neural network model; An input value plotting step of the computing device plotting an input value input to the neural network model to a clustering distribution of the training set; An input value classification step of determining, by the computing device, whether the input value is included in a specific cluster which is one of clustering distributions of the training set; An input value for determining, by the computing device, whether or not to re-learn the input value or the re-learning weight based on a distance between the input value and a centroid of the specific cluster or a probability that the input value is included in the specific cluster Determining whether to relearn; And A neural network model relearning step, wherein the computing device re-learns the neural network model based on the re-learning or re-learning weight of the input value determined in the re-learning of the input value; Including, How to determine whether to relearn inputs in neural network model. The method of claim 1, After the neural network model relearning step, A cluster center value updating step of updating, by the computing device, a centroid of the specific cluster based on whether the input value is re-learned or the re-learning weight determined in the step of determining whether to re-learn the input value; Further comprising, How to determine whether to relearn inputs in neural network model. The method of claim 1, In the training set clustering step and the input value floating step, clustering is performed by performing linear mapping or auto encoder based non-linear mapping through K-means, spectral clustering, or N-cut. How to relearn the value. Receiving unit for receiving an input value; A neural network model computing processor that receives the input value from the receiver, includes a neural network model that has been learned, and calculates an output value by applying the received input value to the neural network model; A clustering processor configured to generate clustering information by receiving the input value from the receiving unit and clustering a training set used for learning the neural network model among the input values; The clustering processor receives the clustering information, and re-learning the input value based on a probability of the input value received at the receiving unit being included in a specific cluster or a distance from a centeroid of a specific cluster of the clustering information. A re-learning determination processor for determining whether or not a re-learning weight applied to the input value; And The relearning determination processor receives the relearning of the input value or the relearning weights, and relearns the neural network model of the neural network model computing processor to update the neural network model of the neural network model computing processor A processor; Including, Re-learning device for input values in neural network model. The method of claim 4, wherein The re-learning processor, characterized in that for updating the center (centeroid) of a specific cluster of the clustering information, Re-learning device for input values in neural network model. The method of claim 4, wherein The clustering processor, Re-learning device for input values in a neural network model, characterized in that the clustering by performing a linear mapping through K-means, spectral clustering or N-cut or non-linear mapping based on Auto encoder. In a program stored in a recording medium that performs a re-learning method for input values in a neural network model, A training set clustering step of performing clustering on the training set used for learning the previously learned neural network model; Plotting an input value inputted to the neural network model to a clustering distribution of the training set; An input value classification step of determining whether the input value is included in a specific cluster which is one of clustering distributions of the training set; Determining whether to re-learn the input value or re-learning weight based on the distance between the input value and the centroid of the specific cluster or the probability that the input value is included in the specific cluster. ; And A neural network model relearning step of re-learning the neural network model based on whether the input value is re-learned or the re-learning weight determined in the input value re-learning determination step; A program stored in a recording medium that runs on a computer.

Images