WO2025041292A1 - Training device, inference device, training inference device, program, training inference system, training method, and inference method - Google Patents

Abstract

In the present invention, a pattern detection device (100) comprises: a training data generation unit (101) that generates training data in which a plurality of feature vectors corresponding respectively to a plurality of training target images are arranged in time series; a true value data construction unit (104) that constructs true value data in which a plurality of true value vectors, each indicating true/false for each of the plurality of training target images, are arranged so as to correspond to the time series of the plurality of feature vectors; and a reservoir circuit training unit (107) that uses the training data and the true value data to train a three-layer neural network using reservoir computing.

Description

Learning device, inference device, learning inference device, program, learning inference system, learning method, and inference method

This disclosure relates to a learning device, an inference device, a learning inference device, a program, a learning inference system, a learning method, and an inference method.

Machine learning is divided into CNN (Convolutional Neural Network) and RNN (Recurrent Neural Network).
CNN is a deep learning technology that has been improving the accuracy of static pattern detection. On the other hand, RNN has been used exclusively for time series signal learning because it has recurrent connections that CNN does not have.

Generally, CNN requires large-scale hardware resources compared to RNN, whereas reservoir computing, which is a technique for RNN, has the advantage of being able to learn with fewer resources compared to other CNN and RNN techniques. In addition, reservoir computing has the advantage that it does not require learning inside the intermediate layer, so the intermediate layer can be implemented as a circuit using ordinary physical media.

Reservoir computing has been widely applied to time-series signal learning, but the scope of application of reservoir computing can be expanded by applying reservoir computing to the learning and detection of static two-dimensional patterns. As an example of such technology, a method for implementing two-dimensional pattern detection learning using reservoir computing is shown in Non-Patent Document 1.

Paugam-Moisy et al. , “Delay learning and polychronization for reservoir computing”, Neurocomputing 71 (7-9), pp. 1143-1158, 2008.

　In conventional technology, it is necessary to learn the coupling constants between the intermediate layer units of reservoir computing, which means that the original feature of reservoir computing, that is, the ability to use ordinary physical media to construct the intermediate layer, is lost. For this reason, even if there were physical media that could be used to construct the intermediate layer, there is the problem that it would be more difficult to construct a learning network using them.

Therefore, one or more aspects of the present disclosure aim to make it possible to easily build a learning network using reservoir computing.

The learning device according to one aspect of the present disclosure is characterized by comprising: a learning data generation unit that generates learning data in which a plurality of feature vectors corresponding to a plurality of learning target images are arranged in a time series; a true-value data construction unit that constructs true-value data in which a plurality of true-value vectors indicating the truth or falsity of each of the plurality of learning target images are arranged so as to correspond to the time series of the plurality of feature vectors; and a learning unit that uses the learning data and the true-value data to train a three-layer neural network using reservoir computing.

The inference device according to one aspect of the present disclosure is characterized by comprising: an application data generation unit that generates a feature vector as application data from an inference target image, which is an image to be inferred; and a processing unit that uses training data in which a plurality of feature vectors corresponding to each of a plurality of training target images are arranged in a time series, and true value data in which a plurality of true value vectors indicating the authenticity of each of the plurality of training target images are arranged so as to correspond to the time series of the plurality of feature vectors, to train a three-layer neural network using reservoir computing, generates a trained three-layer neural network using the coupling constants of the three-layer neural network obtained as a result of the training, and inputs the application data into the trained three-layer neural network to infer the authenticity of the inference target image.

The learning and inference device according to one aspect of the present disclosure is characterized by comprising: a learning data generation unit that generates learning data in which a plurality of feature vectors corresponding to each of a plurality of learning target images are arranged in a time series; a true-value data construction unit that constructs true-value data in which a plurality of true-value vectors indicating the authenticity of each of the plurality of learning target images are arranged so as to correspond to the time series of the plurality of feature vectors; a learning unit that uses the learning data and the true-value data to train a three-layered neural network using reservoir computing and obtains coupling constants of the three-layered neural network as a result of the training; an application data generation unit that generates feature vectors as application data from an inference target image, which is an image that is the subject of inference; and a processing unit that uses the obtained coupling constants to generate a trained three-layered neural network and inputs the application data to the trained three-layered neural network to infer the authenticity of the inference target image.

The program according to the first aspect of the present disclosure causes a computer to function as a training data generation unit that generates training data in which a plurality of feature vectors corresponding to a plurality of training target images are arranged in a time series, a true-value data construction unit that constructs true-value data in which a plurality of true-value vectors indicating the truth or falsity of each of the plurality of training target images are arranged so as to correspond to the time series of the plurality of feature vectors, and a learning unit that uses the training data and the true-value data to train a three-layer neural network using reservoir computing.

The program according to the second aspect of the present disclosure causes a computer to function as an application data generation unit that generates a feature vector as application data from an inference target image, which is an image to be inferred, and a processing unit that uses training data in which a plurality of feature vectors corresponding to each of a plurality of training target images are arranged in a time series and true-value data in which a plurality of true-value vectors indicating the truth or falsehood of each of the plurality of training target images are arranged so as to correspond to the time series of the plurality of feature vectors, trains a three-layer neural network using reservoir computing, generates a trained three-layer neural network using the coupling constants of the three-layer neural network obtained as a result of the training, and inputs the application data into the trained three-layer neural network to infer the truth or falsehood of the inference target image.

The learning inference system according to one aspect of the present disclosure includes a learning data generation unit that generates learning data in which a plurality of feature vectors corresponding to each of a plurality of learning target images are arranged in a time series; a true-value data construction unit that constructs true-value data in which a plurality of true-value vectors indicating the authenticity of each of the plurality of learning target images are arranged so as to correspond to the time series of the plurality of feature vectors; a learning unit that uses the learning data and the true-value data to train a three-layer neural network using reservoir computing and obtains coupling constants of the three-layer neural network as a result of the training; an application data generation unit that generates feature vectors as application data from an inference target image, which is an image that is the subject of inference; and a processing unit that uses the obtained coupling constants to generate a trained three-layer neural network and inputs the application data to the trained three-layer neural network to infer the authenticity of the inference target image.

A learning method according to one aspect of the present disclosure is characterized in that it generates learning data in which a plurality of feature vectors corresponding to a plurality of training target images are arranged in a time series, constructs true-value data in which a plurality of true-value vectors indicating the truth or falsehood of each of the plurality of training target images are arranged to correspond to the time series of the plurality of feature vectors, and uses the learning data and the true-value data to train a three-layered neural network using reservoir computing.

The inference method according to one aspect of the present disclosure is characterized in that it generates a feature vector as application data from an inference target image, which is an image that is the subject of inference, and trains a three-layered neural network using reservoir computing using training data in which a plurality of feature vectors corresponding to each of a plurality of training target images are arranged in time series, and true-value data in which a plurality of true-value vectors indicating the authenticity of each of the plurality of training target images are arranged so as to correspond to the time series of the plurality of feature vectors, generates a trained three-layered neural network using the coupling constants of the three-layered neural network obtained as a result of the training, and inputs the application data into the trained three-layered neural network to infer the authenticity of the inference target image.

According to one or more aspects of the present disclosure, a learning network can be easily constructed using reservoir computing.

1 is a block diagram illustrating a schematic configuration of a pattern detection device as a learning and inference device according to a first embodiment. 1 is a schematic diagram showing an example of a learning image that is a sample image including a pattern detection target; 10 is a schematic diagram showing an example of extracting a plurality of partial images for determining whether the image is a genuine image or a false image from a learning image; FIG. 11 is a schematic diagram for explaining an example of a learning image set. FIG. 4 is a schematic diagram for explaining processing in a learning phase in the first embodiment. FIG. 2 is a schematic diagram for explaining an example of learning a plurality of learning images in the first embodiment. FIG. 11 is a schematic diagram for explaining a process in an application phase in the first embodiment. FIG. 1 is a block diagram showing a schematic configuration of a computer. FIG. 11 is a block diagram illustrating a schematic configuration of a pattern detection device as a learning and inference device according to a second embodiment. 13 is a schematic diagram for explaining the processing in a learning input data designation unit and an orthogonal filter application unit in the learning phase in embodiment 2. FIG. FIG. 2 is a schematic diagram illustrating an example of a two-dimensional Walsh-Hadamard filter with rank N=8. FIG. 11 is a schematic diagram for explaining processing in an application phase in the second embodiment. FIG. 11 is a block diagram illustrating a schematic configuration of a pattern detection device according to a third embodiment. 13 is a schematic diagram for explaining the processing in a learning input data designation unit, an orthogonal filter application unit, and a range conversion unit in the learning phase in embodiment 3. FIG. 1 is a graph showing a hyperbolic tangent function. FIG. 13 is a schematic diagram for explaining processing in an application phase in the third embodiment.

Embodiment 1.
FIG. 1 is a block diagram illustrating a schematic functional configuration of a pattern detection device 100 serving as a learning and inference device according to the first embodiment.
The pattern detection device 100 includes a learning data generating unit 101 , a true value data constructing unit 104 , a reservoir circuit learning unit 107 , an application data input unit 111 , a reservoir circuit processing unit 112 , and an application data output unit 113 .

The pattern detection device 100 operates in two phases, a learning phase and an application phase, for a processing process that uses an Echo State Network, which is a discrete-time implementation of reservoir computing.

The learning data generating unit 101 generates learning data in which a plurality of feature vectors corresponding to a plurality of learning target images are arranged in time series.
In embodiment 1, the training data generation unit 101 extracts pixel values in a predetermined order from each of a plurality of training target images, which are two-dimensional images, and generates each of a plurality of feature vectors by arranging the extracted pixel values in the order in which they were extracted.
The learning data generating unit 101 includes a learning input data specifying unit 102 and a learning feature vector storage unit 103 .

In the learning phase, the learning input data designation unit 102 transforms one learning image to generate a feature vector. Note that each partial image included in the learning image is also referred to as a learning target image. The learning input data designation unit 102 transforms multiple learning target images to generate multiple feature vectors. Then, in the learning phase, the learning input data designation unit 102 configures these multiple feature vectors to have a fixed time length, thereby generating learning feature vectors as learning data.

The learning feature vector storage unit 103 stores the learning feature vectors generated by the learning input data designation unit 102 during the learning phase.

The true value data construction unit 104 constructs true value data in which a plurality of true value vectors indicating the true or false status of each of the plurality of training target images are arranged to correspond to a time series of a plurality of feature vectors generated from the plurality of training target images.
The true-value data constructing unit 104 includes a learning output data specifying unit 105 and a learning true-value vector storing unit 106 .

In the learning phase, the learning output data specification unit 105 generates a learning true value vector that indicates true value data corresponding to each feature vector included in the learning feature vector, so that the true value vector corresponds one-to-one to the true values of each feature vector included in the learning feature vector generated by the learning input data specification unit 102.

In the learning phase, the training true value vector storage unit 106 stores the individual true value vectors included in the training true value vector generated by the training output data designation unit 105 so that the individual true value vectors correspond to the individual feature vectors included in the training feature vector stored in the training feature vector storage unit 103.
Then, the learning true-value vector storage unit 106 inputs the learning true-value vector as true-value data to the reservoir circuit learning unit 107 .

The reservoir circuit learning unit 107 functions as a learning unit that uses the learning data generated by the learning data generation unit 101 and the true value data constructed by the true value data construction unit 104 to train the three-layer neural network using reservoir computing.

In the learning phase, the reservoir circuit learning unit 107 applies an Echo State Network learning algorithm to the three-layer neural network based on the learning feature vectors stored in the learning feature vector storage unit 103 and the learning true-value vectors stored in the learning true-value vector storage unit 106.
Then, the reservoir circuit learning unit 107 stores the coupling constants of the three-layered neural network as the learning result obtained by the Echo State Network learning algorithm.

The application data input unit 111 in the first embodiment functions as an application data generation unit that generates, as application data, a feature vector from an inference target image, which is an image to be inferred.
In embodiment 1, the application data input unit 111 extracts pixel values in a predetermined order from the inference target image, which is a two-dimensional image, and generates application data by arranging the extracted pixel values in the order in which they were extracted.

For example, in the application phase, the application data input unit 111 accepts a new input of a two-dimensional image and converts the two-dimensional image to obtain its feature vector. This two-dimensional image becomes the inference target image.

The reservoir circuit processing unit 112 generates a trained three-layer neural network using the coupling constants of the three-layer neural network obtained as a result of learning in the reservoir circuit learning unit 107, and functions as a processing unit that infers the authenticity of the inference target image by inputting application data from the application data input unit 111 to the trained three-layer neural network.

For example, in the application phase, the reservoir circuit processing unit 112 constructs a trained three-layer neural network using the coupling constants stored in the reservoir circuit learning unit 107. The reservoir circuit processing unit 112 applies a trained Echo State Network to the trained three-layer network. As shown in FIG. 7, which will be described later, the trained Echo State Network includes an input layer 112a, an intermediate layer 112b, and an output layer 112c.
Then, the reservoir circuit processing unit 112 inputs the feature vector acquired by the application data input unit 111 to the input layer 112a as input data.

The application data output unit 113 obtains the output data from the output layer 112c and outputs it as the judgment result.

In this embodiment, a two-dimensional image will be taken as an example of a pattern detection target.
A set of partial images extracted from a learning image is defined as a learning input image set, and each of the partial images is a learning target image.
Of the multiple partial images, the partial image containing the detection target is defined as a true image, and the other partial images are defined as false images. If the partial image is a true image, the true vector (1,0) ^T (T is the transpose of the vector) is used. If the partial image is a false image, the true vector (0,1) ^T is used. These are used as true data for two output layer units, and correspond to the multiple partial images included in the learning input image set, forming a learning output data set.
Here, an example of operation when a learning image set consisting of a learning input image set and a learning output data set is learned by an Echo State Network will be described, divided into a learning phase and an application phase.

First, the learning phase will be described.
FIG. 2 is a schematic diagram showing an example of a learning image, which is a sample image including a pattern detection target.
In this example, six apples present in a learning image are used as an example of a detection target.

For this reason, for example, as shown in FIG. 3, in order to determine whether an image is genuine or fake from the learning image IM, multiple rectangular frames are specified, and the images within the multiple rectangular frames are extracted as partial images P01 to P12.

FIG. 4 is a schematic diagram for explaining an example of a learning image set.
In the learning image set, positive images that are the detection targets are indicated as "true," and false images that are not the detection targets are indicated as "false."

The normal image shown in Fig. 4 is a partial image extracted from a rectangular frame including six apples present in the learning image IM shown in Fig. 3. In order to learn this normal image as a detection target, a true value vector (1, 0) ^T is given as true value data.

The fake image shown in Fig. 4 is a partial image extracted from a randomly specified rectangular frame in an area not including six apples present in the learning image IM shown in Fig. 3. To learn this fake image as a detection target, a true value vector (0, 1) ^T is given as true value data.

Here, we will use Figure 5 to explain the process of learning a partial image from a learning image set, which has a true or false image as its true value, using the Echo State Network.

FIG. 5 is a schematic diagram for explaining the processing in the learning phase.
First, two-dimensional image data showing one partial image used for learning the Echo State Network is stored in the learning image input buffer 102 a of the learning input data designation unit 102 .

Next, in order to convert the partial image represented by the two-dimensional image data stored in the training image input buffer 102a into a feature vector, the training input data designation unit 102 converts the resolution of the partial image and stores the converted data in the feature vector conversion image buffer 102b.

Next, the learning input data specification unit 102 converts the resolution-converted converted partial image into a feature vector and stores it in the feature vector buffer 102c. To vectorize this converted partial image, the learning input data specification unit 102 may, for example, start with the upper left corner pixel of the converted partial image and read out the pixel values one pixel at a time in the leftward direction, proceed from the top row to the bottom row, and so on down to the bottom right pixel, thereby vectorizing the image.

Next, the true value of one partial image input to the learning input data specification unit 102, in other words, the true value vector associated with a positive image if the image contains the detection target, or a false image if not, is stored in the true value vector storage buffer 105a of the learning output data specification unit 105.

Here, the number of elements of the true vector may be specified as a number equal to or greater than 1. For example, the number of elements of the true vector may be set to 2, and learning may be performed by associating a column vector (1,0) ^T with a positive image and a column vector (0,1) ^T with a false image.

As described above, by performing appropriate network initial settings for the feature vectors obtained from partial images and stored in the feature vector buffer 102c, the true-value vectors stored in the true-value vector storage buffer 105a, and the Echo State Network of the reservoir circuit learning unit 107, it is possible to learn two images, a positive image and a false image, in the Echo State Network using the learning algorithm described in Reference 1 below.

Here, the learning of the Echo State Network requires learning only the coupling constants from the intermediate layer to the output layer, and a specific calculation method for this will be described.
First, in the reservoir circuit learning unit 107, for the connections mutually connecting the input layer 107a, the intermediate layer 107b, and the output layer 107c of the Echo State Network, the coupling constant matrix from the input layer 107a to the intermediate layer 107b is W _in , the mutual coupling constant matrix of the internal units of the intermediate layer 107b is W _res , and the coupling constant matrix from the output layer 107c to the intermediate layer 107b is W _fb . In addition, when the input signal to the input layer 107a at a discrete time k is u[k], the state of the internal units of the intermediate layer 107b is x[k], and the output signal of the output layer 107c is y[k], the relationship of the following formula (1) holds.

In equation (1), λ is called the leaking rate, and indicates the extent to which the activity of x[k] from one time point before is reflected in the activity of x[k+1], in other words, how much of the past internal state history of the hidden layer is dragged along the time axis. When λ=1, x[k] is not reflected at all in the update of x[k+1], in other words, the internal state history is not dragged along the time axis at all.

Furthermore, when the coupling constant matrix from the intermediate layer 107b to the output layer 107c _{is Wout} , the matrix in which u[k] and x[k] from time 1 to T are organized as column vectors is X, and the matrix in which y[k], which is the true value data, is organized as a column vector is _Ytarget , there is a relationship between X and _Ytarget as shown in the following equation (2).

When equation (1) is solved for W _out using the Ridge Regression described in the above-mentioned Reference 1, the following equation (3) is obtained.

As a result, W _out can be calculated from the feature vector derived from the partial image, X configured from the internal state of the intermediate layer 107b, and Y _target configured from the true value vector.

It should be noted that equation (3) shows the calculation process of Ridge Regression, and β is an optimization parameter in Ridge Regression.
If Ridge Regression is not used for equation (2) and W _out is directly calculated using the Moore-Penrose pseudoinverse matrix, the solution is obtained by the following equation (4).

However, since the amount of calculation required for the pseudo-inverse matrix in equation (4) increases when the dimension of X is large, Ridge Regression is used to avoid the pseudo-inverse matrix calculation and reduce the amount of calculation. The optimization calculation using Ridge Regression in this case is given by the following equation (5).

In formula (5), if the term β∥W _out ∥ ² is not present, when the norm size of W _out becomes large, the output of the Echo State Network becomes unstable. For this reason, the term β∥W _out ∥ ² is included in the calculation process of formula (5) to perform optimization calculations and obtain W _out with an adjusted norm size.

Above, we have explained an example of training the Echo State Network on one partial image, but here we will use Figure 6 to explain an example of training multiple learning images with true values of either positive or false images assigned to them using the Echo State Network.

In FIG. 6, the feature vectors converted by the learning input data designation unit 102 are used as the learning images, which are multiple two-dimensional images that are trained by the Echo State Network. The process of converting each two-dimensional image into a feature vector can be performed using the procedure described above with reference to FIG. 5.

6, n (n is an integer of 2 or more) true images _T1 , _T2 , ..., _Tn and m (m is an integer of 2 or more) false images _F1 , _F2 , ..., _Fm are stored in chronological order in the form of feature vectors in the learning input data designation unit 102. Furthermore, l consecutive feature vectors are stored for each of the true and false images.

In other words, a total of l×(n+m) feature vectors are stored in the learning input data designation unit 102 in chronological order, constituting the column vectors of the matrix X in the above formula (2). Corresponding to this chronological order, a total of l×(n+m) true value vectors associated with each of the positive and false images are stored in the learning output data designation unit 105 in chronological order, constituting the column vectors of the matrix Y _target in the above formula (2).

As a specific true vector, as described above, a column vector (1,0) ^T may be associated with a true image, and a column vector (0,1) ^T may be associated with a false image.

Next, the application phase in the first embodiment will be described with reference to FIG.
First, two-dimensional image data representing one partial image is stored in the application image input buffer 111 a of the application data input unit 111 .

Next, in order to convert the partial image represented by the two-dimensional image data stored in the application image input buffer 111a into a feature vector, the application data input unit 111 converts the resolution of the partial image and stores the converted data in the image buffer for feature vector conversion 111b.

Next, the application data input unit 111 converts the resolution-converted converted partial image into a feature vector and stores it in the feature vector buffer 111c. To vectorize this converted partial image, the application data input unit 111 may, for example, start with the upper left corner pixel of the converted partial image and read out the pixel values one pixel at a time in the leftward direction, proceed from the top row to the bottom row, and so on down to the bottom right pixel, thereby vectorizing the image.

The reservoir circuit processing unit 112 generates a trained three-layered neural network using the coupling constants of the three-layered neural network acquired as a result of training by the reservoir circuit training unit 107. The reservoir circuit processing unit 112 applies a trained Echo State Network to the trained three-layered neural network. The reservoir circuit processing unit 112 having the trained Echo State Network accepts input of a feature vector constructed by the application data input unit 111 with a new two-dimensional image as a column vector.
The trained Echo State Network includes an input layer 112 a, an intermediate layer 112 b, and an output layer 112 c. The reservoir circuit processing unit 112 calculates an output vector using the trained Echo State Network, and stores the output vector in an output vector storage buffer 113 a of the application data output unit 113.

Then, in the application data output unit 113, the judgment circuit 113b judges whether the output vector is a positive image, meaning that it is a detection target, or a false image, meaning that it is not a detection target. In order to judge whether the image is positive or false, the judgment circuit 113b may, for example, apply a threshold judgment to each element of an output vector having two elements, or may use linear separation or other machine learning methods to make the judgment.

As a result, a learning network using reservoir computing can be easily constructed. Then, it becomes possible to detect a specific object contained in a static two-dimensional image as a rectangular position of a partial image using the Echo State Network. In order to further improve this discrimination ability, a machine learning technique called ensemble learning described in Reference 2 below is known, and high discrimination performance can be ensured by combining multiple discrimination functions with low discrimination performance. Using this ensemble learning, a trained Echo State Network can be regarded as one discrimination function, and multiple Echo State Networks can be combined to construct a discrimination function with higher accuracy.
Reference 2: I. D. Mienye and Y. Sun, “A Survey of Ensemble Learning: Concepts, Algorithms, Applications, and Prospects”, IEEE Access Vol. 10, pp. 99129-99149, 2022.

In the first embodiment, the input pattern to the reservoir computing circuit is expressed as a column vector, each column vector has a certain time length and constitutes an input pattern. By associating the output data with a teacher signal corresponding to the classification category of the input pattern, the input pattern can be associated with the dynamic equilibrium point of the intermediate layer of the reservoir computing circuit during learning of the reservoir computing circuit. This makes it possible to realize pattern detection using the reservoir computing circuit.

The above-described pattern detection apparatus 100 can be realized by, for example, a computer 10 as shown in FIG.
The computer 10 includes a storage 11 such as a hard disk drive (HDD) or a solid state drive (SSD), a memory 12 , a processor 13 such as a central processing unit (CPU), a processing circuit 14 , and a reservoir circuit 15 .

The processing circuit 14 is composed of a single circuit, a composite circuit, a processor operated by a program, a parallel processor operated by a program, an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), etc.

For example, the learning input data designation unit 102 , the learning output data designation unit 105 and the application data input unit 111 can be configured by the processing circuitry 14 .
The learning feature vector storage unit 103 and the learning true value vector storage unit 106 can be configured by the memory 12 .
The reservoir circuit learning unit 107 and the reservoir circuit processing unit 112 can be configured by loading a program stored in the storage 11 into the memory 12, having the processor 13 execute the program, and having the processor 13 utilize the reservoir circuit 15.
The application data output unit 113 can be realized by loading a program stored in the storage 11 into the memory 12 and having the processor 13 execute the program.

The above programs may be downloaded to the storage 11 from a recording medium (not shown) via a reader/writer (not shown) or from a network via a communication I/F (Interface) (not shown), and then loaded onto the memory 12 and executed by the processor 13. Alternatively, the programs may be directly loaded onto the memory 12 from a recording medium (not shown) via a reader/writer or from a network via a communication I/F (not shown), and executed by the processor 13.
In other words, the program may be provided by a program product such as a recording medium.
As described above, the pattern detection apparatus 100 can be realized by a processing circuit network.

Embodiment 2.
FIG. 9 is a block diagram showing a schematic configuration of a pattern detection apparatus 200 as a learning and inference apparatus according to the second embodiment.
The pattern detection device 200 includes a learning data generation unit 201 , a true value data construction unit 104 , a reservoir circuit learning unit 107 , an application data generation unit 210 , a reservoir circuit processing unit 112 , and an application data output unit 113 .
In the second embodiment, an orthogonal filter application unit 220 that functions as a part of the learning data generation unit 201 and the application data generation unit 210 is provided.

The true value data construction unit 104, the reservoir circuit learning unit 107, the reservoir circuit processing unit 112, and the application data output unit 113 of the pattern detection device 200 according to the second embodiment are similar to the true value data construction unit 104, the reservoir circuit learning unit 107, the reservoir circuit processing unit 112, and the application data output unit 113 of the pattern detection device 100 according to the first embodiment.

However, in the application phase, the reservoir circuit processing unit 112 of the pattern detection device 200 according to the second embodiment inputs the feature vector filtered by the orthogonal filter application unit 220 to the input layer 112a as input data.

The learning data generating unit 201 generates learning data in which a plurality of feature vectors corresponding to a plurality of learning target images are arranged in time series.
In embodiment 2, the training data generation unit 201 extracts pixel values in a predetermined order from a processed image, which is the result of applying an orthogonal filter to each of a plurality of training images, which are two-dimensional images, and generates each of a plurality of feature vectors by arranging the extracted pixel values in the order in which they were extracted.

The learning data generating unit 201 includes a learning input data specifying unit 202 , an orthogonal filter applying unit 220 , and a learning feature vector storage unit 103 .
The training feature vector storage unit 103 of the pattern detection device 200 according to the second embodiment is similar to the training feature vector storage unit 103 of the pattern detection device 100 according to the first embodiment.

In the learning phase, the learning input data specification unit 202 provides the orthogonal filter application unit 220 with an image in which the resolution of one learning image has been converted.

The application data generation unit 210 extracts pixel values in a predetermined order from a processed image, which is the result of applying an orthogonal filter to the inference target image, which is a two-dimensional image, and generates application data by arranging the extracted pixel values in the order in which they were extracted.
The application data generating unit 210 includes an application data input unit 211 and an orthogonal filter application unit 220 .

In the application phase, the application data input unit 211 accepts new input of a two-dimensional image and provides the two-dimensional image with its resolution converted to the orthogonal filter application unit 220.

The orthogonal filter application unit 220 generates a feature vector by applying an orthogonal filter to the image from the learning input data designation unit 202. The orthogonal filter application unit 220 then generates a learning feature vector by configuring a plurality of filtered feature vectors to have a certain time length, and provides the feature vector to the learning feature vector storage unit 103.
Moreover, the orthogonal filter application unit 220 applies an orthogonal filter to the image from the application data input unit 211 to generate a feature vector, and provides the feature vector to the reservoir circuit processing unit 112 .

FIG. 10 is a schematic diagram for explaining the processing in the learning input data designation unit 202 and the orthogonal filter application unit 220 in the learning phase in the second embodiment.
Here, an example is shown in which a Walsh-Hadamard filter having filter coefficients of two values, "1" and "-1", is used as the orthogonal filter used in the orthogonal filter application unit 220. Also, as in the first embodiment, among the operations for converting the input data held in the learning input data designation unit 202, a process in the learning phase when one image having a true image or a false image as a true value from the learning image set is learned by the Echo State Network will be described.

First, two-dimensional image data showing one partial image to be used for learning the Echo State Network is stored in the learning image input buffer 202a of the learning input data designation unit 202.

Next, in order to convert the partial image represented by the two-dimensional image data stored in the training image input buffer 202a into a feature vector, the training input data designation unit 202 converts the resolution of the partial image and stores the converted data in the feature vector conversion image buffer 202b.

Next, the orthogonal filter application processing unit 220a of the orthogonal filter application section 220 performs filter processing on the two-dimensional image stored in the feature vector conversion image buffer 202b, and stores the result in the orthogonal filter output buffer 220b.

Next, the orthogonal filter application unit 220 generates a feature vector from the data stored in the orthogonal filter output buffer 220b, in the same manner as in embodiment 1, and stores the feature vector in the feature vector buffer 220c.

Then, the learning phase processing by the Echo State Network is the same as in embodiment 1. Here, an example of an orthogonal filter used in the orthogonal filter application processing unit 220a of the orthogonal filter application section 220 is the Walsh-Hadamard filter described in Reference 3 below.

Reference 3: Y. Hel-Or and H. Hel-Or, “Real-Time Pattern Matching Using Projection Kernels”, IEEE Transactions on Pattern Analysis and Machine Intelligence, 27(9), pp. 1430-1445, 2005.

A Walsh-Hadamard filter is defined as one-dimensional, and a two-dimensional Walsh-Hadamard filter can be further defined as the product of a first dimension and a first dimension.

Figure 11 shows an example of a two-dimensional Walsh-Hadamard filter with rank N=8, where one element filter made up of 8x8 coefficients is arranged vertically and horizontally in a way that refines the spatial frequency, and can be applied to an image with 8x8 pixels vertically and horizontally.

Here, the rank N of the Walsh-Hadamard filter must be a power of 2, and can be N=2, 4, 8, 16, 32, . . .
11, white means a coefficient of "+1" and black means a coefficient of "-1." In other words, when applying this Walsh-Hadamard filter to an image of 8×8 pixels, no multiplication is required, and the filter calculation is performed only by addition and subtraction, enabling high-speed calculation of spatial frequency decomposition.

Next, the application phase in the second embodiment will be described with reference to FIG.
First, two-dimensional image data representing one partial image is stored in the application image input buffer 211 a of the application data input unit 211 .

Next, in order to convert the partial image represented by the two-dimensional image data stored in the application image input buffer 211a into a feature vector, the application data input unit 211 converts the resolution of the partial image and stores the converted data in the image buffer for feature vector conversion 211b.

Next, the orthogonal filter application processing unit 220a of the orthogonal filter application section 220 performs filter processing on the two-dimensional image stored in the feature vector conversion image buffer 211b, and stores the result in the orthogonal filter output buffer 220b.

Next, the orthogonal filter application unit 220 generates a feature vector from the data stored in the orthogonal filter output buffer 220b, as in the first embodiment, and stores the feature vector in the feature vector buffer 220c.

The reservoir circuit processing unit 112 having the trained echo state network receives an input of a feature vector constructed by the orthogonal filter application unit 220 with a new two-dimensional image as a column vector.
The trained Echo State Network includes an input layer 112 a, an intermediate layer 112 b, and an output layer 112 c. The reservoir circuit processing unit 112 calculates an output vector using the trained Echo State Network, and stores the output vector in an output vector storage buffer 113 a of the application data output unit 113.

Then, in the application data output unit 113, the determination circuit 113b determines whether the output vector is a positive image, meaning that it is a detection target, or a false image, meaning that it is not a detection target.

In the second embodiment, similarly to the first embodiment, in the learning phase, a plurality of learning images may be stored in the learning input data designation unit 202, and a plurality of true value vectors linked to each learning image may be stored in the learning output data designation unit 105, and then the reservoir circuit learning unit 107 may learn the Echo State Network.

In the second embodiment, an example was described in which a Walsh-Hadamard filter was used as the orthogonal filter used in the orthogonal filter application unit 220, but another orthogonal filter may be used. For example, a two-dimensional image may be converted into a feature vector using a discrete cosine transform (DCT) or a fast Fourier transform (FFT) as introduced in the above reference 3.

In addition, in the second embodiment, when the learning patterns to be trained by the reservoir computing circuit are mapped into the feature space, the distance in the feature space of the learning patterns having different true values can be expanded using an orthogonal transformation, so that a learning network using reservoir computing can be easily constructed and the discrimination performance of pattern detection using the reservoir computing circuit can be improved.

The above-described pattern detection apparatus 200 can also be realized by a computer 10 as shown in FIG.
For example, the orthogonal filter application unit 220 can also be configured by loading a program stored in the storage 11 into the memory 12 and having the processor 13 execute the program, or by the processing circuit 14 .

Embodiment 3.
FIG. 13 is a block diagram illustrating a schematic configuration of a pattern detection apparatus 300 according to the third embodiment.
The pattern detection device 300 includes a learning data generation unit 301 , a true value data construction unit 104 , a reservoir circuit learning unit 107 , an application data generation unit 310 , a reservoir circuit processing unit 112 , and an application data output unit 113 .
In the third embodiment, an orthogonal filter application unit 320 and a range conversion unit 321 that function as part of the learning data generation unit 301 and the application data generation unit 310 are provided.

The true value data construction unit 104, the reservoir circuit learning unit 107, the reservoir circuit processing unit 112, and the application data output unit 113 of the pattern detection device 300 of embodiment 3 are similar to the true value data construction unit 104, the reservoir circuit learning unit 107, the reservoir circuit processing unit 112, and the application data output unit 113 of the pattern detection device 100 of embodiment 1.
However, in the application phase, the reservoir circuit processing unit 112 of the pattern detection device 300 according to the third embodiment inputs the feature vector converted by the range conversion unit 321 to the input layer 112a as input data.

The learning data generating unit 301 generates learning data in which a plurality of feature vectors corresponding to a plurality of learning target images are arranged in time series.
In embodiment 3, the training data generation unit 301 extracts pixel values in a predetermined order from a transformed image, which is the result of applying a function that transforms pixel values into a predetermined range, to a processed image, which is the result of applying an orthogonal filter to each of a plurality of training images, which are two-dimensional images, and generates each of a plurality of feature vectors by arranging the extracted pixel values in the order in which they were extracted.

The learning data generation unit 301 includes a learning input data specification unit 202 , an orthogonal filter application unit 320 , a range conversion unit 321 , and a learning feature vector storage unit 103 .
The learning input data specifying unit 202 of the pattern detection device 300 according to the third embodiment is similar to the learning input data specifying unit 202 of the pattern detection device 200 according to the second embodiment.
Moreover, the training feature vector storage unit 103 of the pattern detection device 300 according to the third embodiment is similar to the training feature vector storage unit 103 of the pattern detection device 300 according to the first embodiment.

The application data generating unit 310 generates, as application data, a feature vector from an inference target image, which is an image to be inferred.
In embodiment 3, the application data generation unit 310 extracts pixel values in a predetermined order from a transformed image, which is the result of applying a function that transforms pixel values into a predetermined value range, to a processed image, which is the result of applying an orthogonal filter to an inference target image, which is a two-dimensional image, and generates application data by arranging the extracted pixel values in the order in which they were extracted.

The application data generation unit 310 includes an application data input unit 211 , an orthogonal filter application unit 320 , and a range conversion unit 321 .
The application data input unit 211 of the pattern detection device 300 according to the third embodiment is similar to the application data input unit 211 of the pattern detection device 200 according to the third embodiment.

The orthogonal filter application unit 320 applies an orthogonal filter to the image from the learning input data designation unit 202 , and provides the image data after the filter processing to a value range conversion unit 321 .
Furthermore, the orthogonal filter application unit 320 applies an orthogonal filter to the image from the application data input unit 211 , and provides the image data after the filter processing to the value range conversion unit 321 .

In the learning phase, the value range conversion unit 321 applies a function that converts a value range to image data that has been filtered by the orthogonal filter application unit 320 for the image from the learning input data designation unit 202, thereby generating a feature vector. The value range conversion unit 321 then configures a plurality of feature vectors to which such a function has been applied so as to have a certain time length, thereby generating a learning feature vector, and provides the feature vector to the learning feature vector storage unit 103.
In addition, during the learning phase, the value range conversion unit 321 applies a function that converts the value range to the image data filtered by the orthogonal filter application unit 320 for the image from the application data input unit 211, generates a feature vector, and provides the feature vector to the reservoir circuit processing unit 112.

FIG. 14 is a schematic diagram for explaining the processing in the learning input data designation unit 202, the orthogonal filter application unit 320, and the range conversion unit 321 in the learning phase in the third embodiment.
Here, as in the second embodiment, an example is shown in which a Walsh-Hadamard filter having filter coefficients of two values, "1" and "-1", is used as the orthogonal filter used in orthogonal filter application section 320.
Next, the hyperbolic tangent function (tanh(x)) is used as a nonlinear function for range conversion used in the range conversion unit 321, and the output value of the Walsh-Hadamard filter output from the orthogonal filter application unit 320 is limited to a range between "-1" and "1" to construct a feature vector.

First, two-dimensional image data showing one partial image to be used for learning the Echo State Network is stored in the learning image input buffer 202a of the learning input data designation unit 202.

Next, in order to convert the partial image represented by the two-dimensional image data stored in the training image input buffer 202a into a feature vector, the training input data designation unit 202 converts the resolution of the partial image and stores the converted data in the feature vector conversion image buffer 202b.

Next, the orthogonal filter application processing unit 320a of the orthogonal filter application section 320 performs filtering on the two-dimensional image stored in the feature vector conversion image buffer 202b, and stores the result in the orthogonal filter output buffer 320b.

Next, the range conversion unit 321 converts the values of the filtered image data stored in the orthogonal filter output buffer 320b using the range conversion function possessed by the range conversion function application processing unit 321a, and data having the converted values is stored in the range conversion function output buffer 321b.

Then, the range conversion unit 321 generates a feature vector from the data stored in the range conversion function output buffer 321b, in the same manner as in embodiment 1, and stores the feature vector in the feature vector buffer 321c.

Here, an example of the range conversion function that the range conversion function application processing unit 321a has is a hyperbolic tangent function.
The hyperbolic tangent function tanh(x) is expressed by the following equation (6), and FIG. 15 shows a graph thereof.

In the second embodiment, the feature vector is constructed directly from each value of the orthogonal filter output buffer 220b, which is the calculation result of the orthogonal filter application unit 220. In the third embodiment, however, even if the distribution of the application values of the orthogonal filter to the input image varies greatly, the value range conversion unit 321 ensures that tanh(x) falls within the range -1<f(x)<1, so that the operation of the reservoir circuit learning unit 107 during learning can be stabilized.

Next, the application phase in the third embodiment will be described with reference to FIG.
First, two-dimensional image data representing one partial image is stored in the application image input buffer 211 a of the application data input unit 211 .

Next, in order to convert the partial image represented by the two-dimensional image data stored in the application image input buffer 211a into a feature vector, the application data input unit 211 converts the resolution of the partial image and stores the converted data in the image buffer for feature vector conversion 211b.

Next, the orthogonal filter application processing unit 320a of the orthogonal filter application section 320 performs filter processing on the two-dimensional image stored in the feature vector conversion image buffer 211b, and stores the result in the orthogonal filter output buffer 320b.

Next, the range conversion unit 321 converts the values of the filtered image data stored in the orthogonal filter output buffer 320b using the range conversion function possessed by the range conversion function application processing unit 321a, and data having the converted values is stored in the range conversion function output buffer 321b.

Then, the range conversion unit 321 generates a feature vector from the data stored in the range conversion function output buffer 321b, in the same manner as in embodiment 1, and stores the feature vector in the feature vector buffer 321c.

The reservoir circuit processing unit 112 having a trained echo state network receives an input of a feature vector constructed by the range conversion unit 321 with a new two-dimensional image as a column vector.
The trained Echo State Network includes an input layer 112 a, an intermediate layer 112 b, and an output layer 112 c. The reservoir circuit processing unit 112 calculates an output vector using the trained Echo State Network, and stores the output vector in an output vector storage buffer 113 a of the application data output unit 113.

Then, in the application data output unit 113, the determination circuit 113b determines whether the output vector is a positive image, meaning that it is a detection target, or a false image, meaning that it is not a detection target.

In the third embodiment, as in the first embodiment, in the learning phase, multiple learning images may be stored in the learning input data designation unit 202, and multiple true value vectors linked to each learning image may be stored in the learning output data designation unit 105, and then the reservoir circuit learning unit 107 may learn the Echo State Network.

As described above, according to the third embodiment, before inputting a feature vector to the reservoir circuit processing unit 112 having a trained Echo State Network, the value of each element constituting the feature vector can be kept between "-1" and "1" by the value range conversion unit 321. This makes it possible to easily construct a learning network using reservoir computing and stabilize the processing operation of the reservoir circuit processing unit 112.

In addition, in the third embodiment, by narrowing the range of values of the input data input to the reservoir computing circuit, the operation of the reservoir computing circuit becomes stable, and it becomes possible to improve the discrimination performance of pattern detection using the reservoir computing circuit.

In the above-described embodiments 1 to 3, the pattern detection devices 100 to 300 have been described as devices that perform processing in both the learning phase and the inference phase, but embodiments 1 to 3 are not limited to such devices.
For example, the first to third embodiments can be configured as a learning device that performs processing in the learning phase and an inference device that performs processing in the inference phase.
Furthermore, the processes performed by the pattern detection devices 100 to 300 according to the first to third embodiments may be distributed among a plurality of computers that are connected to a network and capable of transmitting and receiving data to and from each other. In other words, the first to third embodiments may be configured as a pattern detection system that serves as a learning inference system made up of a plurality of computers.

100, 200, 300 Pattern detection device; 101, 201, 301 Learning data generation unit; 102, 202 Learning input data specification unit; 103 Learning feature vector storage unit; 104 True value data construction unit; 105 Learning output data specification unit; 106 Learning true value vector storage unit; 107 Reservoir circuit learning unit; 210, 310 Application data generation unit; 111, 211 Application data input unit; 112 Reservoir circuit processing unit; 113 Application data output unit; 220, 320 Orthogonal filter application unit; 321 Value range conversion unit.

Claims (14)

a learning data generation unit that generates learning data in which a plurality of feature vectors corresponding to a plurality of learning target images are arranged in time series;
a true-value data constructing unit that constructs true-value data in which a plurality of true-value vectors indicating true or false of each of the plurality of learning target images are arranged in correspondence with a time series of the plurality of feature vectors;
a learning unit that uses the training data and the true value data to train a three-layer neural network using reservoir computing. each of the plurality of training images is a two-dimensional image;
2. The learning device according to claim 1, wherein the learning data generation unit generates each of the plurality of feature vectors by extracting pixel values in a predetermined order in the two-dimensional image and arranging the extracted pixel values in the extraction order. each of the plurality of training images is a two-dimensional image;
2. The learning device according to claim 1, wherein the learning data generation unit extracts pixel values in a predetermined order from a processed image that is a result of applying an orthogonal filter to the two-dimensional image, and generates each of the plurality of feature vectors by arranging the extracted pixel values in the extraction order. each of the plurality of training images is a two-dimensional image;
2. The learning device according to claim 1, wherein the learning data generation unit extracts pixel values in a predetermined order from a transformed image, which is a result of applying a function that transforms pixel values into a predetermined range, to a processed image, which is a result of applying an orthogonal filter to the two-dimensional image, and generates each of the plurality of feature vectors by arranging the extracted pixel values in the order in which they were extracted. an application data generation unit that generates a feature vector as application data from an inference target image that is an image to be inferred;
a processing unit that uses training data in which a plurality of feature vectors corresponding to each of a plurality of training target images are arranged in a time series and true-value data in which a plurality of true-value vectors indicating the authenticity of each of the plurality of training target images are arranged to correspond to the time series of the plurality of feature vectors, generates a trained three-layered neural network using a coupling constant of the three-layered neural network obtained as a result of the training, and infers the authenticity of the inference target image by inputting the application data to the trained three-layered neural network. The inference target image is a two-dimensional image,
The inference device according to claim 5 , wherein the application data generation unit generates the application data by extracting pixel values in a predetermined order in the two-dimensional image and arranging the extracted pixel values in the order in which they were extracted. The inference target image is a two-dimensional image,
The inference device according to claim 5, wherein the application data generation unit generates the application data by extracting pixel values in a predetermined order from a processed image that is a result of applying an orthogonal filter to the two-dimensional image, and arranging the extracted pixel values in the order in which they were extracted. The inference target image is a two-dimensional image,
The inference device according to claim 5, wherein the application data generation unit generates the application data by extracting pixel values in a predetermined order from a transformed image, which is a result of applying a function that transforms pixel values into a predetermined value range, to a processed image, which is a result of applying an orthogonal filter to the two-dimensional image, and arranging the extracted pixel values in the order in which they were extracted. a learning data generation unit that generates learning data in which a plurality of feature vectors corresponding to a plurality of learning target images are arranged in time series;
a true-value data constructing unit that constructs true-value data in which a plurality of true-value vectors indicating true or false of each of the plurality of learning target images are arranged in correspondence with a time series of the plurality of feature vectors;
A learning unit that uses the learning data and the true value data to learn a three-layered neural network using reservoir computing, and obtains a coupling constant of the three-layered neural network and a result of the learning;
an application data generation unit that generates a feature vector as application data from an inference target image that is an image to be inferred;
a processing unit that uses the acquired coupling constants to generate a trained three-layered neural network and inputs the application data into the trained three-layered neural network to infer whether the inference target image is true or false. Computer,
a learning data generation unit that generates learning data in which a plurality of feature vectors corresponding to a plurality of learning target images are arranged in time series;
a true-value data constructing unit that constructs true-value data in which a plurality of true-value vectors indicating true or false of each of the plurality of learning target images are arranged in correspondence with a time series of the plurality of feature vectors; and
a program that functions as a learning unit that uses the training data and the true value data to train a three-layer neural network using reservoir computing. Computer,
an application data generation unit that generates a feature vector as application data from an inference target image that is an image to be inferred;
a processing unit that performs training of a three-layered neural network using reservoir computing, using training data in which a plurality of feature vectors corresponding to each of a plurality of training target images are arranged in a time series, and true-value data in which a plurality of true-value vectors indicating the authenticity of each of the plurality of training target images are arranged so as to correspond to the time series of the plurality of feature vectors, generates a trained three-layered neural network using a coupling constant of the three-layered neural network obtained as a result of the training, and inputs the application data into the trained three-layered neural network, thereby functioning as a processing unit that infers the authenticity of the inference target image. a learning data generation unit that generates learning data in which a plurality of feature vectors corresponding to a plurality of learning target images are arranged in time series;
a true-value data constructing unit that constructs true-value data in which a plurality of true-value vectors indicating true or false of each of the plurality of learning target images are arranged in correspondence with a time series of the plurality of feature vectors;
A learning unit that uses the learning data and the true value data to learn a three-layered neural network using reservoir computing, and obtains a coupling constant of the three-layered neural network and a result of the learning;
an application data generation unit that generates a feature vector as application data from an inference target image that is an image to be inferred;
a processing unit that uses the acquired coupling constants to generate a trained three-layered neural network and inputs the application data into the trained three-layered neural network to infer whether the inference target image is true or false. generating training data in which a plurality of feature vectors corresponding to a plurality of training target images are arranged in a time series;
constructing true-value data in which a plurality of true-value vectors indicating the authenticity of each of the plurality of learning target images are arranged to correspond to a time series of the plurality of feature vectors;
A learning method comprising: learning a three-layer neural network using reservoir computing by using the learning data and the true value data. A feature vector is generated as application data from an inference target image, which is an image to be inferred;
training a three-layered neural network using reservoir computing, using training data in which a plurality of feature vectors corresponding to each of a plurality of training target images are arranged in a time series, and true-value data in which a plurality of true-value vectors indicating the authenticity of each of the plurality of training target images are arranged so as to correspond to the time series of the plurality of feature vectors; and generating a trained three-layered neural network using coupling constants of the three-layered neural network obtained as a result of the training;
and inputting the application data into the trained three-layer neural network to infer whether the inference target image is true or false.

Images