Restoration feasibility determination method, restoration feasibility determination device and program

Abstract

In the present invention, a computer executes a determination procedure in which an abnormality detection technique is applied to compressed data generated by compressing input data using an encoder of a learned auto-encoder, and it is thereby determined whether the compressed data can be decompressed. As a result of executing the determination procedure, it is possible to determine decompression feasibility on the basis of the compressed data.

Description

[Technical field]

[0001] The present invention relates to a restoration feasibility determination method, a restoration feasibility determination device, and a program. [Background technology]

[0002] In recent years, sensor nodes used in sensor networks have become energy-efficient and have not only sensing and communication functions, but also some information processing functions, such as data compression and data classification / event detection functions. This field is called edge computing (for example, Non-Patent Document 1, etc.).

[0003] In addition, if the observation sensing data obtained by the sensor node is sent to the center as is, an increase in the number of sensor nodes will increase the amount of communication to the center, which may exceed the limit of the available communication amount (bandwidth). Therefore, it has been considered to compress the sensor data at the sensor node to reduce the amount of communication information to the center.

[0004] Communications based on data classification / event detection do not involve communicating the sensor data itself, so this serves as a form of data compression.

[0005] The following three types of communication can be considered between the sensor node and the center. (1) The sensor node identifies the sensor data and sends only the identification results to the center. (2) The sensor node obtains the features of the sensor data and sends only the features to the center. (3) The sensor data is compressed on the sensor node side and the compressed data is sent to the center.

[0006] In the case of (1), the system works if the identification is performed correctly. However, since it is often desired to function even when the sensor data is non-stationary, it is difficult to identify unexpected inputs.

[0007] In the case of (2), identification is performed using the features transmitted to the center, but when the sensor data exceeds the expected identification, the required features are different, making it difficult to identify unexpected inputs.

[0008] In the case of (3), a data-dependent compression / decompression method can be used depending on the type of sensor data to be compressed. Different compression / decompression methods are often used depending on whether the sensor data is an image, sound, or other time series. There are also lossless compression and lossy compression, and lossy compression generally has a higher compression ratio than lossy compression.

[0009] Here, we consider the case of (3), in which an autoencoder (AE) is used as the decompressor, and data compressed by an encoder at the sensor node is sent to a center, where it is decompressed by a decoder at the center. The AE can use any sensor data, but it needs to be trained in advance using the sensor data (Non-Patent Documents 2 and 3). In the AE, the decompressor is created by learning, so it cannot decompress anything other than the assumed learning data and data close to it, and the output of the AE will differ from the input. In the case of sensor data that exceeds such expectations, the AE will no longer function as a decompressor.

[0010] In addition, in general, when edge computing is performed, the computational power and energy consumption on the sensor node side are limited compared to the center side. [Prior art documents] [Non-patent literature]

[0011] [Non-Patent Document 1] Mobile, Ubiquitous, and Intelligent Computing, [online], Internet<URL:https: / / rd.springer.com / content / pdf / 10.1007%2F978-3-642-40675-1.pdf> [Non-Patent Document 2] Diederik P Kingma, Max Welling, Auto-Encoding Variational Bayes, [online], Internet<URL:https: / / arxiv.org / abs / 1312.6114> [Non-Patent Document 3] Carl Doersch, Tutorial on Variational Autoencoders, [online], Internet<URL:https: / / arxiv.org / abs / 1606.05908> [Non-Patent Document 4] Kohei Komukai, Shin Mizutani, Yasue Kishino, Yutaka Yanagisawa, Yoshinari Shirai, Takayuki Suyama, Ren Omura, Hiroshi Sawada, Futoshi Naya,"A Communication Method Using Autoencoders for Distributed Recognition Sensor Networks with Continuous Update of Classifiers," Transactions of Information Processing Society of Japan, Vol. 60, No. 10, pp. 1780-1795, (Oct. 2019) Summary of the Invention [Problem to be solved by the invention]

[0012] In the sensor network described above, when communication between the sensor nodes and the center is performed using compressed data and lossy compression is used, an autoencoder may be used as a compression / decompression device (Non-Patent Document 4).

[0013] An autoencoder consists of an encoder and a decoder, where the encoder compresses input data and the decoder restores the compressed data. An autoencoder is a type of neural network that is trained so that the input data to the encoder is output directly from the decoder. Data used as training data and similar data can be compressed and restored through learning and generalization. On the other hand, it is unclear whether data other than the above can be compressed and restored, and it is unclear whether the input sensor data can be restored on the decoder side.

[0014] In an autoencoder, it is determined that restoration is possible when the difference between input and output is less than a certain threshold. However, in the above-mentioned sensor network, the input data is in the sensor node and the output data is in the center, so in order to compare the input data and the output data, it is necessary to transmit the input data or output data to the center or the sensor node.

[0015] In addition, since autoencoders are expected to generalize through learning, the training data itself can be compressed and restored, and data close to the training data can also be compressed and restored. Therefore, if the data obtained by the sensor nodes can be limited to the training data and its close equivalents, the sensor network can function and be operated.

[0016] However, in actual observations, the sensor node may obtain data equivalent to unlearned data, and it is difficult to determine whether the input data has been restored using only the restored data on the center side. Therefore, it is necessary to be able to determine whether the data can be restored based on the compressed data that is the transmitted data.

[0017] The present invention has been made in consideration of the above-mentioned points, and has an object to make it possible to determine whether or not restoration is possible based on compressed data. [Means for solving the problem]

[0018] In order to solve the above problem, a determination procedure is provided in which an anomaly detection technique is applied to compressed data generated by compressing input data using an encoder of a trained autoencoder, and the determination procedure determines whether the compressed data is restorable before restoring the compressed data. and , a learning procedure in which the autoencoder is retrained using the input data when it is determined in the determination procedure that the data cannot be restored; The computer executes the above. Effect of the Invention

[0019] It is possible to determine whether or not restoration is possible based on the compressed data. [Brief description of the drawings]

[0020] [Figure 1] 1 is a diagram illustrating an example of a configuration of a sensor network according to an embodiment of the present invention. [Diagram 2] 2 is a diagram illustrating an example of a hardware configuration of a center 10 according to an embodiment of the present invention. [Diagram 3] FIG. 2 is a diagram illustrating an example of a functional configuration of a sensor network according to a first embodiment. [Figure 4] FIG. 2 is a diagram showing a schematic diagram of a compression unit 21 and a restoration unit 13 according to the embodiment of the present invention. [Diagram 5] FIG. 1 is a diagram illustrating an autoencoder. [Figure 6] FIG. 13 is a diagram showing an example of AE by two-dimensional CNN used in image data. [Figure 7] FIG. 1 is a diagram illustrating an example of CNN+FC used in an autoencoder. [Figure 8] FIG. 1 is a diagram for explaining a learning data set used for this embodiment. [Figure 9] FIG. 1 is a diagram for explaining an unlearned data set used for this embodiment. [Figure 10] FIG. 13 is a diagram showing the variation in the latent space of trained / untrained data points when the number of dimensions of the latent variable Z is two. [Figure 11] FIG. 13 is a diagram showing the variation in the latent space of trained / untrained data points when the number of dimensions of the latent variable Z is 16. [Figure 12] FIG. 13 is a diagram showing the variation in the latent space of trained / untrained data points when the number of dimensions of the latent variable Z is 64. [Figure 13] FIG. 13 shows a histogram of squared Mahalanobis distance (anomaly score function) of trained data points. [Figure 14] FIG. 1 is a first diagram showing thresholds in squared Mahalanobis distance. [Figure 15] FIG. 2 is a second diagram showing thresholds in squared Mahalanobis distance. [Figure 16] FIG. 13 is a diagram showing thresholds for −ln probability values. [Figure 17] FIG. 13 is a diagram showing thresholds according to LOF based on the nearest neighbor method. [Figure 18] FIG. 13 is a diagram showing thresholds according to One-class SVM. [Figure 19] FIG. 13 is a diagram illustrating thresholds according to Isolation Forest. [Figure 20] FIG. 13 is a diagram showing evaluation results of the recognition error rates of cases 1 to 3 and cases 5 to 7. [Figure 21] 10 is a flowchart illustrating an example of a processing procedure when communicating compressed data. [Figure 22] FIG. 11 is a diagram illustrating an example of a functional configuration of a sensor network according to a second embodiment. [Diagram 23] FIG. 13 is a diagram showing evaluation results of compression and decompression using multiple types of AE. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0021] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. Fig. 1 is a diagram showing an example of the configuration of a sensor network in an embodiment of the present invention. In the sensor network shown in Fig. 1, one or more sensor nodes 20 are connected to a center 10 via a network such as the Internet.

[0022] The sensor node 20 is a computer that is connected to a sensor (or has a sensor), compresses data sensed by the sensor (hereinafter referred to as “sensor data”), and transmits the compressed data (hereinafter referred to as “compressed data”) to the center 10.

[0023] The center 10 is one or more computers having a function of receiving compressed data transmitted from the sensor node 20 and decompressing the compressed data.

[0024] In this embodiment, for convenience, the sensor data is assumed to be image data. Details of image data will be described later. However, data to which this embodiment can be applied is not limited to a specific format.

[0025] Fig. 2 is a diagram showing an example of a hardware configuration of the center 10 in the embodiment of the present invention. The center 10 in Fig. 2 includes a drive device 100, an auxiliary storage device 102, a memory device 103, a processor 104, and an interface device 105, which are all connected to each other via a bus B.

[0026] A program for implementing the processing at center 10 is provided by a recording medium 101 such as a CD-ROM. When recording medium 101 storing the program is set in drive device 100, the program is installed from recording medium 101 to auxiliary storage device 102 via drive device 100. However, the program does not necessarily have to be installed from recording medium 101, but may be downloaded from another computer via a network. Auxiliary storage device 102 stores the installed program as well as necessary files, data, etc.

[0027] When an instruction to start a program is received, the memory device 103 reads out and stores the program from the auxiliary storage device 102. The processor 104 is a CPU or a GPU (Graphics Processing Unit), or a CPU and a GPU, and executes functions related to the center 10 in accordance with the program stored in the memory device 103. The interface device 105 is used as an interface for connecting to a network.

[0028] The sensor node 20 may have the same hardware configuration as the center 10. However, the hardware performance of the sensor node 20 may be lower than that of the center 10.

[0029] Fig. 3 is a diagram showing an example of the functional configuration of a sensor network in the first embodiment. In Fig. 3, the sensor node 20 has a compression unit 21 and a transmission unit 22. These units are realized by processing that one or more programs installed in the sensor node 20 cause a processor (e.g., a CPU) of the sensor node 20 to execute.

[0030] The compression unit 21 compresses the sensor data to generate compressed data. In this embodiment, lossy compression is performed. The transmission unit 22 transmits the compressed data to the center 10.

[0031] Meanwhile, the center 10 has a receiving unit 11, a determining unit 12, a restoring unit 13, a classifying unit 14, and a learning unit 15. Each of these units is realized by a process in which one or more programs installed in the center 10 are executed by a processor 104. The center 10 also uses a data storage unit 16. The data storage unit 16 can be realized by using, for example, an auxiliary storage device 102, or a storage device connectable to the center 10 via a network.

[0032] The receiving unit 11 receives compressed data transmitted from the sensor node 20. The determining unit 12 performs a restorable determination for the compressed data. The restorable determination means a determination of whether restoration is possible or not. In this embodiment, restorable means a state in which the difference between the input and output of the compression / recovery device (the compression unit 21 and the restoration unit 13) is less than a predetermined threshold, and non-restorable means a state in which the difference is equal to or greater than the threshold. The restoration unit 13 generates restored data by restoring the compressed data determined to be restorable. The classification unit 14 performs classification of the restored data, which will be described later. The learning unit 15 performs learning for the neural network constituting the compression unit 21 and the restoration unit 13, by using a data group stored in the data storage unit 16. In this embodiment, the compression and restoration (encoding and decoding) of the sensor data are performed using an autoencoder (AE).

[0033] Fig. 4 is a diagram showing a schematic diagram of the compression unit 21 and the restoration unit 13 in the embodiment of the present invention. As shown in Fig. 4, the compression unit 21 is realized by the encoder of the restoration unit 13, and the restoration unit 13 is realized by the decoder of the AE, and the information with the smallest number of units in the intermediate layer of the AE is communicated as compressed data between the sensor node 20 and the center 10, thereby reducing the amount of communication.

[0034] As shown in Fig. 5, AE is a type of layered neural network, and includes a compression coder (encoder) and a restoration decoder (decoder) (Non-Patent Documents 2 and 3). In Fig. 5, the white circles represent neural network units, and the lines connecting the units represent weights (links) between the units. Fig. 5 shows an AE with a five-layer structure, in which the input is five-dimensional, compressed to two dimensions, and the input is reproduced on the output side.

[0035] In the encoder, the number of units gradually decreases as the processing proceeds from left to right in each layer to compress the dimensionality of the input data. In the decoder's output layer, the number of units that was reduced in the intermediate layer increases to the same number as in the encoder's input layer, restoring the input information.

[0036] That is, a normal AE has multiple layers of encoders and decoders, with the middle layer having the fewest number of units, forming an hourglass shape. The information from this part with the fewest number of units is used as compressed data for communication between the sensor node 20 and the center 10, thereby reducing the amount of communication. Note that AE is trained by supervised learning so that input and output are the same. The loss function used in training varies depending on the data set used, but may be MSE (Mean Square Error), BCE (Binary Cross Entropy), or CCE (Categorical Cross Entropy), etc.

[0037] As described above, in this embodiment, the sensor data is image data. The number of images in the image data is 28×28, and each pixel is 8 bits. An AE suitable for such image data is shown in FIG. 6.

[0038] The AE in Figure 6 compresses and decompresses image data with a 28x28 matrix as 8-bit input, and typically uses a Convolutional Neural Network (CNN) (https: / / qiita.com / icoxfog417 / items / 5fd55fad152231d706c2). In a CNN, the dimensionality of the information is reduced while preserving the spatial position information, and it is compressed into information such as features. The CNN in Figure 6 has a 9-layer structure, and the input and output represent 28x28 dimensional vectors. N and M in each hidden layer, represented as a rectangular parallelepiped, represent the number of unit faces (types of filters), and the shaded rectangle represents the range of connections from the previous layer.

[0039] In addition, as shown in Figure 7, a neural network with a fully connected layer (FC (Fully Connected) layer) sandwiched between CNNs is also suitable for image data.

[0040] In the above examples, we considered a neural network with a left-right symmetric structure centered on the middle hidden layer, but any type of neural network can be used as long as the input and output have the same dimension (number of units), and any neural network can be used as a compressor / decompressor as long as the input data can be restored in the output layer.

[0041] Next, the determination unit 12 will be described in more detail. As described above, if the difference between the input and output of the compression / decompression unit (compression unit 21 and restoration unit 13) is defined as being less than a predetermined threshold value as being restorable, and the difference is defined as being equal to or greater than the threshold value as being unrestorable, in order to strictly execute the restorability determination according to such a definition, the sensor data before compression and the restored data are required. In that case, the sensor data needs to be constantly transmitted from the sensor node 20, and the significance of compressing the sensor data is lost. Therefore, in this embodiment, the restorability determination is performed by determining whether the compressed sensor data is learned data (unlearned data) for the autoencoder as the compression unit 21 and restoration unit 13 (hereinafter referred to as "learned determination"). In other words, the determination result that the data is learned indicates that the data is restorable, and the determination result that the data is unlearned indicates that the data is unrestorable.

[0042] In this embodiment, an anomaly detection technique is used to determine whether learning has been completed. There are various conventional methods in the technical field of anomaly detection, and these conventional techniques are used in this embodiment. However, assuming that the learned data is provided in advance to the autoencoder as the compression / decompression device, but the unlearned data is not provided in advance, it is possible to use an anomaly detection technique in which only normal data is provided.

[0043] In addition, in this embodiment, the learned determination is performed by applying an anomaly detection technique to the compressed data. That is, as a result of applying the anomaly detection technique to the compressed data, if no anomaly is detected (if it is determined to be normal), it is determined to have been learned, and if it is determined to be abnormal, it is determined to have not been learned. However, as described above, in this embodiment, since an anomaly detection technique in which only normal data is given is used, it is necessary to be able to determine that the learned data has been learned based on the compressed data in which normal data (i.e., learned data) has been compressed.

[0044] Therefore, in this embodiment, among AEs, a VAE (Variational AutoEncoder) trained with a loss function that includes a regularization term that satisfies some constraint for compressed data obtained by compressing input data is used (Non-Patent Documents 2 and 3). Anomaly detection is performed based on how well this constraint is satisfied or violated. In particular, input / output data is called observed variables, and compressed data is called latent variables.

[0045] VAE is an AE that is trained by defining a loss function that adds a term to the loss function between the input and output of a normal AE so that the compressed data in the middle layer becomes a set probability density function. Therefore, the encoder of the VAE outputs the parameters of the probability density function. In other words, a VAE is an inference model in which the encoder of the AE determines the latent variable Z (information that cannot be obtained directly from observations such as classification labels and handwriting) from the observed variables (observed data), and a generator model in which the decoder determines the observed variables from the latent variable Z.

[0046] In VAE, the latent variable space in the middle layer, called Z, is usually a Gaussian distribution of data points N(0,I 2 ) The encoder calculates the mean and standard deviation of the Gaussian distribution for each dimension of the latent variable from the input, and the decoder is trained to reproduce the input to the encoder by using the latent variable Z sampled from these probability distributions as input.

[0047] The loss function used is the input-output loss function plus the Kullback-Leibler Divergence (KLD), which represents the distance between distributions so that the probability distribution becomes a set Gaussian distribution. This is the variational lower bound of the log-likelihood.

[0048] In conventional AE, the distribution of compressed data cannot be specified, and it is unclear where untrained data exists in the latent variable space, making it difficult to determine whether data has been trained in order to determine whether it can be restored.

[0049] Therefore, in this embodiment, a VAE is used as the compression unit 21 and the restoration unit 13. In addition, the AE shown in Fig. 6 and Fig. 7 can also be used as a VAE (can be trained). In this embodiment, since the sensor data is image data, a case where the AE in Fig. 7 is used as a VAE will be described. The VAE will be hereinafter referred to as a "target AE."

[0050] The target AE is made to learn in advance the sensor data observed by the sensor node 20 as input and output so that the input can be restored to output. The learning data for the target AE usually needs to be a sufficient number of data sampled from an assumed probability distribution of an event to capture the characteristics of the distribution. It is considered that data points other than the sample points can be interpolated by the generalization ability of the AE.

[0051] Next, the data set used for this embodiment will be described. In this embodiment, handwritten digit data (http: / / yann.lecun.com / exdb / mnist / ), which is often used in the field of supervised learning of machine learning and is called MNIST shown in FIG. 8, was used as the learning data set for the target AE. This data is a 28×28 matrix with 0-255 (2 8 It is handwritten numeric data from 0 to 9, to which a value of 1:8 bits is assigned and the numbers 0 to 9 are given as classification labels. The total number of data in MNIST is 60,000, and it is usually classified into 50,000 data for learning and 10,000 data for testing. In this embodiment, it is assumed that the observation data obtained by the sensor node 20 is 8-bit handwritten numeric data in a 28×28 matrix, and a classification label is required on the center 10 side. In other words, the input data is an image of the handwritten numeric data taken by a camera (sensor node 20), and the output data on the center 10 side is the classification.

[0052] Next, we use Fashion-MNIST (F-MNIST) (https: / / github.com / zalandoresearch / fashion-mnist) as untrained data that cannot be assumed. The data of F-MNIST is the same as MNIST, with each of the 28x28 matrices containing values ​​from 0 to 255 (2 8 The monochrome image data of fashion items, as shown in Figure 9, is assigned a value of 1:8 bits and the numbers 0-9 are given as classification labels. The total number of data in F-MNIST is 60,000, which is classified into 50,000 data for regular training and 10,000 data for testing.

[0053] The data value to be entered is the integer value of the brightness of the pixel of the image, 0-255 (2 8 :8bit) values ​​are normalized to [0,1] and used.

[0054] The target AE is a latent variable Z with a probability density function N(0,I 2 ), so trained data is likely to be close to this distribution, while untrained data is likely to deviate from it.

[0055] Here, we consider the problem of what the dimensionality of the latent variable Z should be to achieve both the learned judgment and the compression rate. That is, it is considered that the smaller the dimensionality of the latent variable Z, the higher the compression rate will be, but the accuracy of the learned judgment will be lower.

[0056] 10 to 12 are diagrams showing the variation in the latent space of trained / untrained data points for various latent variable dimensionalities. FIG. 10 shows the variation when the dimensionality of the latent variable Z is 2. FIG. 11 shows the variation when the dimensionality of the latent variable Z is 16. FIG. 12 shows the variation when the dimensionality of the latent variable Z is 64. However, for FIG. 11 (dimensionality=16) and FIG. 12 (dimensionality=64), data points are two-dimensionalized using a method of embedding data points in a low dimension called t-SNE (t-Student-distributed Neighbor Embedding) (https: / / blog.albert2005.co.jp / 2015 / 12 / 02 / tsne / , etc.). In each figure, black dots (black circles) indicate trained data points (MNIST data points), and white dots (white circles) indicate untrained data points (F-MNIST data points). Note that a trained data point refers to each data point contained in MNIST in the latent space of a target AE that has been trained based on MNIST, and an untrained data point refers to each data point contained in F-MNIST in the latent space.

[0057] 10 to 12, it is considered that if an appropriate number of latent dimensions is used, it is possible to determine whether or not a learning has been completed based on the difference in the positions where the learned / unlearned data points exist. That is, according to Figs. 10 to 12, the positions where the learned / unlearned data points exist seem to be sufficiently separated in 16 dimensions (Fig. 11). Therefore, in this embodiment, the dimension of the latent variable Z is set to 16, and a case will be described in which various anomaly detection methods (anomaly detection technologies) are applied to such a latent variable Z.

[0058] [Case 1] First, based on the Hotelling theory, we consider that the trained data points follow a multivariate Gaussian distribution and the untrained data points deviate from it. If we define the anomaly score function as the squared Mahalanobis distance (squared distance normalized for variance in various dimensions) from the origin or the mean of the trained data points (anomaly score function = squared Mahalanobis distance ∝ -ln probability value of multivariate Gaussian distribution), then the histogram of the squared Mahalanobis distance of the trained data points is χ 2It is known that it follows a distribution (Ide Tsuyoshi, Sugiyama Masaru, "Anomaly Detection and Change Detection", pp.15-25, Kodansha). The integral of the tail part above the set threshold is the rate of misjudging "learned" as "unlearned". The judgment unit 12 uses this threshold to judge whether the new input data is "learned" based on the anomaly score function value.

[0059] FIG. 13 shows a histogram of the squared Mahalanobis distance (anomaly score function) of the trained data points. In FIG. 13, the open histogram shows the histogram of the trained data, and the curve shows the theoretical curve χ 2 The distribution is shown below. 2 The degree of freedom of the distribution is the number of dimensions of the latent variable Z (= 16), and the scale factor is 1. As shown in Figure 13, when the trained data follows a normal distribution, the square of the Mahalanobis distance between new data that follows the same distribution as the trained data and the average of the distribution by the trained data is χ 2 This χ 2 If the misrecognition rate is determined based on the distribution, the threshold value can be determined based on the distribution.

[0060] Assuming that the unlearned data is F-MNIST or the like, the intersection point between the histogram for the squared Mahalanobis distance in the unlearned dataset (hereinafter referred to as the "unlearned histogram") and the histogram in the trained dataset described above (hereinafter referred to as the "trained histogram") is the threshold value for the squared Mahalanobis distance that minimizes the misclassification rate. A misclassification refers to a case where a trained dataset is judged to be unlearned, or vice versa.

[0061] Fig. 14 is a first diagram showing thresholds in squared Mahalanobis distance. In Fig. 14, the X-axis corresponds to squared Mahalanobis distance. Moreover, open histograms indicate learned histograms, and filled histograms indicate unlearned histograms (this also applies to Figs. 15 to 19). Moreover, dashed line L1 indicates the intersection point between the unlearned histogram and the learned histogram, that is, the position of the threshold (this also applies to Figs. 15 to 19).

[0062] The determination unit 12 uses this threshold to determine whether the newly input data has been learned. Specifically, the determination unit 12 calculates the squared Mahalanobis distance of the latent variable Z (i.e., compressed data) of the newly input data, and determines that the data has not been learned if the calculation result is equal to or greater than the threshold, and determines that the data has been learned if the calculation result is less than the threshold.

[0063] [Case 2] The threshold is the intersection of the histogram of the squared Mahalanobis distance from the mean of the trained data points in N(μ,Σ) when fitting to a multivariate Gaussian distribution using all trained data points (hereinafter referred to as the "trained histogram") and the histogram of the squared Mahalanobis distance in the untrained dataset F-MNIST (hereinafter referred to as the "untrained histogram").

[0064] 15 is a second diagram showing the threshold value in the squared Mahalanobis distance. The way to read FIG. 15 is the same as FIG. 14. The method of determining whether learning has been completed is also the same as in Case 1.

[0065] [Case 3] Since an N(μ,Σ) distribution (mean μ, covariance Σ, where matrix Σ is a Gaussian distribution with values ​​only on the diagonal components and zeros elsewhere) is obtained for each piece of trained data, the (1 / N)ΣN(μ,Σ) distribution averaged using all N training data points is considered to be the trained data distribution, and the anomaly score function is defined as a -ln probability value from the probability value of this distribution. In this case, a trained histogram is constructed, and similarly, an untrained histogram is constructed from the anomaly score function in the untrained dataset F-MNIST, and the intersection of the trained histogram and the untrained histogram is set as the threshold.

[0066] Fig. 16 is a diagram showing thresholds at -ln probability values. In Fig. 16, the X-axis is the -ln probability value. In Fig. 16, the case where the probability value is 0 is excluded, and in this case, the anomaly score function is considered to be infinite.

[0067] The determination unit 12 performs a learning determination of the newly input data using the threshold value shown in Fig. 16. Specifically, the determination unit 12 calculates a -ln probability value of the latent variable Z (i.e., compressed data) of the newly input data, and determines that the data has not been learned if the calculation result is equal to or greater than the threshold value, and determines that the data has been learned if the calculation result is less than the threshold value.

[0068] [Case 4] In addition, various differences (anomaly score functions) can be considered as differences in quantification indices in the latent variable Z between newly input data and trained data.

[0069] For example, for one new input data, the encoder outputs μ and Σ, and one distribution N(μ,Σ) is obtained. The difference (anomaly score function) may be defined by the distance between the distributions in KLD, and the threshold may be determined by considering their histogram. In this case, various distributions may be considered as a representative distribution of the trained data, such as the distribution with the smallest distance in KLD, or the difference with the distribution obtained by averaging the distributions of all trained data. In addition, the difference may be considered by the squared Mahalanobis distance, as described above, using only the average μ obtained from the encoder. The difference (anomaly score function) may be defined using the purpose, measurement, and calculable quantities, and the trained determination may be performed using anomaly detection technology.

[0070] The above (cases 1 to 4) are methods that involve threshold processing using an error rate threshold in the tail of the distribution, or an intersection threshold in a histogram using trained / untrained data sets.

[0071] On the other hand, in the cases 5 to 7 described below, usually, only a trained dataset is given and a threshold is set, so the maximum value of the anomaly score function value of the trained dataset is used as the threshold. However, as in the previous cases, an optimal threshold having a minimum misclassification rate due to the intersection with an untrained dataset may be used.

[0072] [Case 5 Nearest Neighbor Method + LOF] Anomaly detection using LOF (local outlier factor) based on the nearest neighbor method is used to determine whether the learning has been completed (Tsuyoshi Ide, "Introduction to Anomaly Detection Using Machine Learning", pp.72-77, Corona Publishing; Tsuyoshi Ide and Masaru Sugiyama, "Anomaly Detection and Change Detection", pp.41-51, Kodansha). LOF sets the number of neighbors in the nearest neighbor method, and defines an anomaly score function based on the distance from the point of interest. Distance 0 is excluded in the definition of LOF.

[0073] Fig. 17 is a diagram showing a threshold value by LOF based on the nearest neighbor method. In Fig. 17, the horizontal axis is LOF. In Fig. 17, the position of dashed line L2, which is the right end of the histogram of the trained dataset, is the threshold value. However, as described above, the intersection (position of dashed line L1) in the histogram of the trained / untrained datasets may be used as the threshold value.

[0074] The judgment unit 12 uses any of the thresholds to judge whether the newly input data has been learned. Specifically, the judgment unit 12 calculates the LOF of the latent variable Z (i.e., compressed data) of the newly input data, and judges that the data has not been learned if the calculation result is equal to or greater than the threshold, and judges that the data has been learned if the calculation result is less than the threshold.

[0075] [Case 6 One-class SVM] One-class SVM is a method devised to apply a two-class classifier, SVM (Support Vector Machine), to either training data or other data (One-class SVM with non-linear kernel (RBF), https: / / scikit-learn.org / stable / auto_examples / svm / plot_oneclass.html#sphx-glr-auto-examples-svm-plot-oneclass-py, etc.). In Case 6, One-class SVM is used as an anomaly score function with the latent variable Z as input.

[0076] Fig. 18 is a diagram showing thresholds by One-class SVM. In Fig. 18, the horizontal axis is the score by One-class SVM. In Fig. 18, the position of dashed line L2, which is the right end of the histogram of the trained dataset, is the threshold. However, as described above, the intersection (position of dashed line L1) in the histogram of the trained / untrained datasets may be used as the threshold.

[0077] [Case 7: Isolation Forest] Isolation Forest (IF) is a method devised to apply Random Forest to anomaly detection (e.g., Isolation Forest example, https: / / scikit-learn.org / stable / auto_examples / ensemble / plot_isolation_forest.html). In Case 7, Isolation Forest is used as an anomaly score function with the latent variable Z as input.

[0078] Fig. 19 is a diagram showing thresholds by Isolation Forest. In Fig. 19, the horizontal axis is the score by Isolation Forest. In Fig. 19, the position of dashed line L2, which is the right end of the histogram of the trained dataset, is the threshold. However, as described above, the intersection (position of dashed line L1) in the histogram of the trained / untrained datasets may be used as the threshold.

[0079] [Case 8] For cases 5 to 7, there is no restriction that the latent variable Z follows a predetermined distribution. Therefore, instead of a VAE, a normal AE can be used as the target AE. The difference between the latent variable Z in the normal AE encoder and the latent variable Z obtained by mapping the restored variable again by the encoder is defined as an anomaly score function, and since the function value is large for data that cannot be restored, a threshold for determining whether the function value has been learned may be defined.

[0080] [others] In addition to the above cases, anomaly detection methods that can be performed using only trained data (known data) can be used to determine whether a model has been trained (https: / / scikit-learn.org / stable / auto_examples / plot_anomaly_comparison.html#sphx-glr-auto-examples-plot-anomaly-comparison-py, Ide Tsuyoshi, Sugiyama Masaru, "Anomaly Detection and Change Detection", Kodansha, etc.).

[0081] 20 shows the evaluation results of the recognition error rate in the experiments conducted by the inventors of the present application for the above-mentioned cases 1 to 3 and cases 5 to 7. Note that for cases 5 to 7, the reason why there is no value in the "distribution" item is that the abnormal scores for cases 5 to 7 are not abnormal scores based on the distribution of the latent variable Z.

[0082] The following describes the processing procedure executed by the sensor node 20 and the center 10 in Fig. 3. Fig. 21 is a flowchart for explaining an example of the processing procedure when communicating compressed data.

[0083] When sensor data (hereinafter referred to as "target data") is generated in the sensor node 20 (Yes in S101), the compression unit 21 compresses the sensor data using the encoder of the target AE to generate compressed data (S102). The transmission unit 22 transmits the compressed data to the center 10.

[0084] When the receiver 11 of the center 10 receives the compressed data, the judger 12 compares a value (hereinafter referred to as an "anomaly score") obtained by applying the anomaly score function to the compressed data (i.e., the value of the latent variable Z) with a threshold corresponding to the anomaly score function, thereby executing a learned judgement (i.e., a restorable judgement) for the target data (S103). Note that the threshold is set in advance.

[0085] If the anomaly score is less than the threshold (i.e., if it is determined that the target data has been learned (can be restored)) (Yes in S103), the restoration unit 13 generates restored data by restoring the compressed data using a decoder of the target AE (S104). Next, the restoration unit 13 outputs the restored data to the classification unit 14 and stores it in the data storage unit 16 (S105). Therefore, a set of restored data for sensor data determined to have been learned (can be restored) is accumulated in the data storage unit 16.

[0086] Next, the classification unit 14 classifies the restored data (S106). If the sensor data is an image of handwritten numbers, classification is performed on what numbers are included in the restored data.

[0087] On the other hand, if the anomaly score is equal to or greater than the threshold (i.e., if the target data is determined to be unlearned (unrecoverable)) (No in S103), the receiver 11 acquires the target data from the sensor node 20 and stores the target data in the data storage unit 16 (S107). For example, the receiver 11 requests the sensor node 20 to transmit the target data. In response to the request, the transmitter 22 of the sensor node 20 transmits the target data to the center 10. The receiver 11 stores the received target data in the data storage unit 16. Therefore, a set of sensor data determined to be unlearned (unrecoverable) is accumulated in the data storage unit 16.

[0088] In order to enable the transfer of target data, it is necessary to prepare a communication path that can send the sensor data itself to the center 10, and a certain amount of memory for storing the target data in the sensor node 20 in preparation for a request to send the target data. However, the communication path can also be one used for transmitting normal compressed data. Also, if the learning completion judgment and the request to send the target data can be processed in real time, no particular memory is required.

[0089] On the other hand, for example, when the number of data stored in the data storage unit 16 satisfies a predetermined condition or at a predetermined timing such as input of an instruction by a system administrator (Yes in S108), the learning unit 15 additionally learns or re-learns the target AE using the data sets (reconstructed data set, unlearned data set) stored in the data storage unit 16 and the trained data set used to train the original target AE (S109). As a result, the performance of the target AE as a compression decompressor can be improved. Note that a known method may be used to train the target AE.

[0090] Next, the learning unit 15 executes a process for updating the target AE (S110). Specifically, the learning unit 15 updates the model parameters of the decoder serving as the restoration unit 13 to values ​​after additional learning or re-learning. The learning unit 15 also transmits model parameters of the encoder serving as the compression unit 21 to the sensor node 20. The compression unit 21 of the sensor node 20 updates the model parameters of the encoder with the received values.

[0091] By updating the model parameters of the target AE, the target AE can restore unlearned data and also restore the originally learned data.

[0092] Next, a second embodiment will be described. In the second embodiment, differences from the first embodiment will be described. Points not specifically mentioned in the second embodiment may be the same as those in the first embodiment.

[0093] Fig. 22 is a diagram showing an example of the functional configuration of a sensor network in the second embodiment. In Fig. 22, the same parts as those in Fig. 3 are given the same reference numerals, and the description thereof will be omitted.

[0094] 22, the sensor node 20 further includes a judgment unit 23. The judgment unit 23 performs learning determination in the same manner as the judgment unit 12. On the other hand, the center 10 does not include the judgment unit 12. That is, the second embodiment is an example in which the learning determination is performed in the sensor node 20.

[0095] The processing procedure of the second embodiment will also be described with reference to Fig. 21. In the second embodiment, the compressed data is not transmitted in step S102, and the determination unit 23 of the sensor node 20 determines whether learning has been completed in step S103. If it is determined that learning has been completed (Yes in S103), the transmission unit 22 transmits the compressed data to the center 10, and steps S104 and onward are executed.

[0096] On the other hand, if it is determined that the data has not been learned (No in S103), the transmitting unit 22 transmits the target data. Upon receiving the target data, the receiving unit 11 of the center 10 stores the target data in the data storage unit 16.

[0097] In the above embodiments, the target VAE has the structure shown in Fig. 7, but the structure of the target VAE is not limited to that shown in Fig. 7, and CNN does not have to be used. As the basis for this, the inventors' evaluation results on compression and decompression using multiple types of AEs are shown.

[0098] Fig. 23 is a diagram showing the evaluation results of compression and decompression using multiple types of AE. Fig. 23 shows input data, true values ​​of classification labels of the input data, decompressed data, and classification label outputs for five types of AE, namely, AE, CNN-AE, VAE, CVAE, and CNN+FC-VAE. Note that AE indicates a fully-connected AE, CNN-AE indicates an AE with a CNN structure, VAE indicates a fully-connected VAE, CVAE indicates a fully-connected conditional VAE, and CNN+FC-VAE is a target VAE that is a VAE that combines a CNN structure and fully-connected.

[0099] As is clear from Fig. 23, compression and decompression are performed appropriately to the extent that classification is possible in all AEs. Note that for the CVAE, which is a conditional VAE, the classification label output value is not required, so it is not shown.

[0100] In this way, the AE applicable to each of the above embodiments is not limited to a specific VAE. Also, any AE that can specify the distribution in the latent variable Z space, not limited to the VAE, can be used as a decompressor that can determine whether it has been learned. Furthermore, if the method is from Case 5 onwards, even an AE that cannot specify the distribution in the latent variable Z space can be used as a decompressor that can determine whether it has been learned.

[0101] As described above, according to each of the above embodiments, it is possible to determine whether or not restoration is possible based on the compressed data.

[0102] In addition, it is possible to collect unlearned data that has been determined to be unlearned (unrestorable) after performing a learned determination (restorability determination), and then use this data to add / re-learn the compression / decompressor so that the compression / decompressor can handle the unlearned data.

[0103] The above-described embodiments may be applied to various types of data other than the data transmitted from the sensor node 20 to the center 10, which are to be compressed.

[0104] In each of the above embodiments, the center 10 and the sensor node 20 are examples of a restoration feasibility determination device.

[0105] Although the embodiment of the present invention has been described in detail above, the present invention is not limited to such specific embodiment, and various modifications and variations are possible within the scope of the gist of the present invention described in the claims. [Explanation of symbols]

[0106] 10 Center 11 Receiving section 12 Judgment section 13 Restoration Section 14 Classification section 15 Learning Department 16 Data storage section 20 Sensor Nodes 21 Compression section 22 Transmitter 23 Judgment section 100 Drive device 101 Recording media 102 Auxiliary storage 103 Memory device 104 processors 105 Interface device B Bus

Claims

1. a determination step of applying an anomaly detection technique to compressed data generated by compressing input data using an encoder of the trained autoencoder, and determining whether the compressed data is restorable before restoring the compressed data; a learning procedure in which the autoencoder is retrained using the input data when it is determined in the determination procedure that the data cannot be restored; The restoration possibility determination method is characterized in that the restoration possibility determination method is executed by a computer.

2. a determination step of determining whether the compressed data is restorable before restoring the compressed data by comparing a value obtained by applying a function based on an anomaly detection technique to compressed data generated by compressing input data using an encoder of a trained autoencoder with a threshold value; The restoration possibility determination method is characterized in that the restoration possibility determination method is executed by a computer.

3. a determination step of determining whether or not the compressed data can be restored before restoring the compressed data, based on whether or not the compressed data generated by compressing input data using an encoder of a trained autoencoder conforms to a probability distribution conformed to by compressed data based on the input data used to train the autoencoder; a learning procedure in which the autoencoder is retrained using the input data when it is determined in the determination procedure that the data cannot be restored; The restoration possibility determination method is characterized in that the restoration possibility determination method is executed by a computer.

4. a learning procedure for re-learning the autoencoder using the input data when it is determined in the determination procedure that the data cannot be restored; 3. The restoration possibility determination method according to claim 2, wherein the restoration possibility determination method is executed by a computer.

5. a determination unit that applies an anomaly detection technique to compressed data generated by compressing input data using an encoder of the trained autoencoder, and determines whether the compressed data is restorable before restoring the compressed data; a learning unit that re-learns the autoencoder using the input data when the determination unit determines that the input data is not restorable; A restoration possibility determination device comprising:

6. a determination unit that determines whether or not the compressed data can be restored before restoring the compressed data by comparing a value obtained by applying a function based on an anomaly detection technology to compressed data generated by compressing input data using an encoder of the trained autoencoder with a threshold value; A restoration possibility determination device comprising:

7. a determination unit that determines whether the compressed data can be restored before restoring the compressed data, based on whether compressed data generated by compressing input data using an encoder of a trained autoencoder follows a probability distribution that compressed data based on input data used for training the autoencoder follows; a learning unit that re-learns the autoencoder using the input data when the determination unit determines that the input data is not restorable; A restoration possibility determination device comprising:

8. 5. A program for causing a computer to execute the restoration possibility determination method according to claim 1.

Images