JP6961527B2 - Information processing equipment, learning methods, and programs - Google Patents

Description

The present invention relates to an information processing device, a learning method, and a program, and particularly to a machine learning technique using a neural network.

In recent years, speeding up of CPU (Central Processing Unit) and GPU (Graphics Processing Unit), increasing the capacity of memory, and machine learning technology using a neural network have been rapidly advancing. For this reason, machine learning using learning data on the order of hundreds of thousands to one million has become possible, and highly accurate identification technology and classification technology are being established (see Non-Patent Document 1).

Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick, Sergio Guadarrama, and Trevor Darrell. Caffe: Convolutional architecture for fast feature embedding. In Proceedings of the 22nd ACM international conference on Multimedia (pp. 675-678) ). ACM.

A large amount of computational cost is required to perform machine learning based on a large amount of learning data. In addition, a huge amount of labor is required for preparing a large amount of learning data and for preprocessing to use the prepared learning data for machine learning. Model creators who generate learning models are also beginning to make a profit by building a learning model created at a high cost on the cloud and letting users use the learning model via a communication network. ing.

When the input data and output data of the learning model include privacy information, such information is transmitted and received via the communication network. This can leak privacy information.

The present invention has been made in view of these points, and an object of the present invention is to provide a technique for suppressing information leakage when using a learning model via a communication network.

The first aspect of the present invention is an information processing device. This device is an information processing device that generates weights for each layer of a neural network including a plurality of layers for detecting a target task. This device has an input conversion unit that converts input data input to the neural network by a predetermined conversion procedure to generate conversion input data, and a conversion input data generated by the input conversion unit as input data of the neural network. It includes a weight update unit that updates the weight of each layer. The predetermined conversion procedure executed by the input conversion unit is a conversion procedure in which the distance between arbitrary two points in the space before conversion and the distance between the two points in the space after conversion have a constant multiple relationship. be.

In the information processing device, the output vector output by the output layer of the neural network has a distance between arbitrary two points in the space where the output vector exists and a distance between the two points in the space after conversion. A conversion procedure that has a constant multiple relationship may further include an output conversion unit that converts the output vector to generate a conversion output vector.

The target task may be an image recognition task for a color image, and the input conversion unit may perform the predetermined task in a space whose coordinate axes are a plurality of variables for specifying the characteristics of each pixel of the color image. You may perform the conversion procedure of.

The predetermined conversion procedure may be any one conversion of expansion conversion, reduction conversion, equal length conversion, and parallel movement conversion, or a composite conversion of two or more conversions.

A second aspect of the present invention is also an information processing device that generates weights for each layer of a neural network including a plurality of layers for detecting a target task. This device uses the target task learning data as input data to update the weights of each layer of the neural network, and the output vector output by the output layer of the neural network can be arbitrarily set in the space where the output vector exists. An output conversion unit that converts the output vector to generate a conversion output vector in a conversion procedure in which the distance between the two points and the distance between the two points in the space after conversion have a constant multiple relationship. To be equipped.

A third aspect of the present invention is a learning method executed by a processor of an information processing apparatus that generates weights for each layer of a neural network including a plurality of layers for detecting a target task. In this method, the processor converts the input data input to the neural network by a predetermined conversion procedure to generate the conversion input data, and the generated conversion input data is used as input data for each layer of the neural network. Perform the steps to update the weights and. The predetermined conversion procedure in the generation step is a conversion procedure in which the distance between arbitrary two points in the space before conversion and the distance between the two points in the space after conversion have a constant multiple relationship.

A fourth aspect of the present invention is a program. This program converts the input data input to the neural network into an information processing device that generates weights for each layer of the neural network having a plurality of layers for detecting the target task by a predetermined conversion procedure, and converts the input data. And a function of updating the weight of each layer of the neural network by using the generated conversion input data as input data. The predetermined conversion procedure in the function to be generated is a conversion procedure in which the distance between arbitrary two points in the space before conversion and the distance between the two points in the space after conversion have a constant multiple relationship.

According to the present invention, the present invention has been made in view of these points, and it is possible to suppress information leakage when using a learning model via a communication network.

It is a figure which shows typically the structure of the neural network which concerns on embodiment. It is a figure which shows typically the outline of the learning model execution system which concerns on embodiment. It is a figure which shows typically the functional structure of the information processing apparatus which concerns on embodiment. It is a figure for demonstrating the conversion procedure carried out by the input conversion unit and the output conversion unit which concerns on embodiment. It is a figure for demonstrating another example of the conversion procedure carried out by the input conversion part which concerns on embodiment. It is a flowchart for demonstrating the flow of the learning process executed by the information processing apparatus which concerns on embodiment.

<Outline of the embodiment>
FIG. 1 is a diagram schematically showing a configuration of a neural network according to an embodiment. Further, FIG. 2 is a diagram schematically showing an outline of the learning model execution system S according to the embodiment. Hereinafter, an outline of the embodiment will be described with reference to FIGS. 1 and 2.

The neural network used by the information processing apparatus according to the embodiment to generate a learning model has a general configuration including an input layer, an intermediate layer, and an output layer as a whole. As shown in FIG. 1, in the information processing apparatus according to the embodiment, the learning data input to the input layer propagates through the intermediate layer, and finally corresponds to the data string output by the output layer and the learning data. The error from the correct label is calculated using the loss function. The information processing apparatus updates the weight set in the intermediate layer by using the error back propagation method based on the calculated error.

Here, the information processing apparatus according to the embodiment has an "input conversion" that converts the learning data by a predetermined conversion procedure before inputting the learning data to the input layer, and a predetermined data output by the output layer. At least one of the "output conversion" to be converted by the conversion procedure is performed.

Note that FIG. 1 shows an example in which the information processing apparatus performs both input conversion and output conversion. Details will be described later, but when the information processing apparatus inputs and converts readable information, the converted information becomes difficult to read. The same applies to output conversion. In addition, the input conversion is designed so that the input conversion of the training data does not affect the learning of the neural network.

As shown in FIG. 2, the learning model execution system S according to the embodiment includes an information processing device 1 and a user terminal 2 that are connected to each other via a communication network N in a manner in which they can communicate with each other. The user terminal 2 receives an input API (Application Program Interface) and an output API from the information processing device 1. By using the input API, the user terminal 2 can transmit the data to be processed to the information processing device 1 via the communication network N. Further, the user terminal 2 can receive the execution result of the learning model from the information processing device 1 via the communication network N by using the output API.

The input API applies the same conversion as the input conversion performed when the information processing device 1 generates the learning model to the processing target data, and transmits the converted processing target data to the information processing device 1 via the communication network N. .. Since the data to be processed after conversion is not readable, it is possible to suppress the reading of information from the data even if the data is leaked during transmission to the information processing device 1 via the communication network N.

Further, when the output result of the learning model is transmitted to the user terminal 2 after the information processing apparatus 1 performs the output conversion, the output API performs the inverse conversion of the output conversion on the acquired output result. As a result, the user terminal 2 can return the output result communicated in a state where it is difficult to read to readable information.

As described above, the learning model generated by the information processing apparatus 1 according to the embodiment can suppress information leakage when the learning model is used via the communication network. Further, even if a third party who does not have an input API for input conversion inputs the data to be processed into the online learning model by some means, the correct processing result cannot be obtained. As a result, the learning model generated by the information processing apparatus 1 according to the embodiment can suppress unauthorized use of the learning model.

<Functional configuration of the information processing device 1 according to the embodiment>
FIG. 3 is a diagram schematically showing a functional configuration of the information processing device 1 according to the embodiment. The information processing device 1 is a device that generates weights for each layer of a neural network including a plurality of layers for detecting a target task, and includes a storage unit 10 and a control unit 11.

The storage unit 10 includes a ROM (Read Only Memory) for storing the BIOS (Basic Input Output System) of the computer that realizes the information processing device 1, a RAM (Random Access Memory) that serves as a work area for the information processing device 1, and an OS (OS). An operating system), an application program, and a large-capacity storage device such as an HDD (Hard Disk Drive) or SSD (Solid State Drive) that stores various information referred to when the application program is executed.

The control unit 11 is a processor such as a CPU or GPU of the information processing device 1, and functions as an input conversion unit 110, a weight update unit 111, and an output conversion unit 112 by executing a program stored in the storage unit 10.

Note that FIG. 3 shows an example in which the information processing device 1 is composed of a single device. However, the information processing device 1 may be realized by computing resources such as a plurality of processors and memories, such as a cloud computing system. In this case, each unit constituting the control unit 11 is realized by executing a program by at least one of a plurality of different processors.

The input conversion unit 110 converts the input data to be input to the neural network by a predetermined conversion procedure to generate the conversion input data. Specifically, in the conversion procedure executed by the input conversion unit 110, the distance between arbitrary two points in the space before conversion and the distance between the corresponding two points in the space after conversion are constant times (0 times). It is a conversion procedure that is related to).

FIG. 4 is a diagram for explaining a conversion procedure carried out by the input conversion unit 110 and the output conversion unit 112 according to the embodiment. In FIG. 4, the input data I shows the input data shown in FIG. The input data I can be represented by a vector, and in the example shown in FIG. 4, it is an n-dimensional vector composed of n elements from N1 to Nn.

An arbitrary n-dimensional vector shall be expressed as one point in an n-dimensional Cartesian coordinate system centered on n defined vectors in which the m (1 ≦ m ≦ n) th element is 1 and the other elements are 0. Can be done. In FIG. 4, the defined vector of the N1 axis is (1,0, ..., 0), and the defined vector of the N2 axis is (0,1,0, ..., 0). The same applies hereinafter.

Let O be the origin of the n-dimensional Cartesian coordinate system, and let two different points in the n-dimensional space be points A and B, respectively. Further, the points A and B after conversion by the input conversion unit 110 are designated as points A'and B', respectively. Assuming that the distance between the point A and the point B in the n-dimensional space is D (A, B), the conversion procedure carried out by the input conversion unit 110 and the output conversion unit 112 sets k to 0 for any point A and point B. When a real number other than is used, the following equation (1) holds.
D (A, B) = kD (A', B') (1)

An example of such a conversion procedure is a conversion of any one of expansion conversion, reduction conversion, equal length conversion (rotation or inversion), and parallel movement conversion, or a composite conversion of two or more conversions.

In FIG. 4, the length D (O, A) of the line segment OA is a, the length D (O, B) of the line segment OB is b, and the length D (A, B) of the line segment AB is c. .. Similarly, the length D (O, A') of the line segment OA'is a', the length D (O, B') of the line segment OB' is b', and the length D of the line segment A'B'( Let A', B') be c'. Further, the angle AOB is θ, and the angle A'OB'is θ'. At this time, from the law of cosines, the following equations (2) and (3) hold.

cosθ = (a ² + b ^2- c ² ) / (2ab) (2)
^{cosθ '= (a' 2 +} b '2 -c' 2) / (2a'b ') (3)
Is established.

In the conversion procedure carried out by the input conversion unit 110 and the output conversion unit 112, the equation (1) holds for any two points in the n-dimensional space, so that a'= ka, b'= kb, and c'= kc hold. Substituting these into equation (3) gives the following equation (4).

cosθ'= ((ka) ² + (kb) ^2- (kc) ² ) / (2kakb)
= (A ² + b ^2- c ² ) / (2ab)
= Cosθ (4)

Equation (4) indicates that the angle is preserved in the conversion procedure performed by the input conversion unit 110 and the output conversion unit 112. Here, the weight updating unit 111 updates the weights of each layer of the neural network by using the conversion input data generated by the input conversion unit 110 as input data and using a known error back propagation method.

Although details are omitted because it is a known technique, the error back propagation method belongs to the problem of finding the extremum of the loss function based on the gradient of the loss function in the n-dimensional space. The gradient of the loss function indicates, so to speak, the "tilt angle" of the loss function in the n-dimensional space. The error back propagation method can be said to be an algorithm designed to reach the extreme value by selecting the direction in which the "tilt angle" of the loss function is large and advancing the "hill" of the loss function in the n-dimensional space.

As shown in the equation (4), the angle is preserved in the conversion procedure carried out by the input conversion unit 110 and the output conversion unit 112 according to the embodiment. This means that the "tilt angle" of the loss function does not change even with the conversion procedure performed by the input conversion unit 110 and the output conversion unit 112. Therefore, even if the conversion procedure performed by the input conversion unit 110 according to the embodiment is applied to the input data, the weight update unit 111 can learn in the same manner as when the conversion procedure is not performed.

Further, even if the output conversion unit 112 according to the embodiment performs a conversion procedure on the output data, it is guaranteed that the shape of the output data is not distorted by the conversion itself. Therefore, the output conversion unit 112 sets the output vector output by the output layer of the neural network as a constant between an arbitrary two points in the space where the output vector exists and a distance between the two points in the converted space. The output vector may be converted to generate a converted output vector by a conversion procedure having a double relationship. As a result, even if the output result is intercepted by a third party in the communication network N, it is possible to prevent a third party who does not have the output API from extracting information from the converted output vector.

5 (a)-(b) is a diagram for explaining another example of the conversion procedure carried out by the input conversion unit 110 according to the embodiment. Specifically, FIG. 5A is a diagram schematically showing a pixel structure of a color image, and FIG. 5B is a diagram schematically showing a conversion procedure in a color space.

As shown in FIG. 5A, most digital color images are composed of one pixel in three colors of red (R), green (G), and blue (B). Although not shown, one pixel is composed of a plurality of variables indicating a plurality of signal components even when the expression difference signal such as YCbCr is used.

For example, it is assumed that each pixel of a color image is represented by a combination of three RGB colors. In this case, each pixel of the color image is represented as a point in the three-dimensional space whose coordinate axes are the colors R, G, and B. For example, a pixel whose luminance values of the R, G, and B colors are 64, 128, and 192, respectively, is a point (64, 128, 192) in a three-dimensional space whose coordinate axes are the R, G, and B colors. Corresponds to.

As shown in FIG. 5B, when the target task is an image recognition task for a color image, the input conversion unit 110 uses a plurality of variables for specifying the characteristics of each pixel of the color image as coordinate axes. The conversion procedure described above may be executed in the space where the above is performed. As a result, the image after conversion loses the balance of hue, saturation, and brightness as compared with the image before conversion, so that the input conversion unit 110 can reduce the readability of the input data.

<Processing flow of learning method executed by information processing device 1>
FIG. 6 is a flowchart for explaining the flow of the learning process executed by the information processing apparatus 1 according to the embodiment. The process in this flowchart starts, for example, when the information processing device 1 is activated.

The input conversion unit 110 acquires input data used for learning to generate weights for each layer of a neural network including a plurality of layers for detecting a target task (S2). The input conversion unit 110 converts the acquired input data by a predetermined procedure (S4). The predetermined conversion procedure executed by the input conversion unit 110 is a conversion procedure in which the distance between arbitrary two points in the space before conversion and the distance between two points in the space after conversion have a constant multiple relationship. Is.

The weight update unit 111 updates the weights of each layer of the neural network using the conversion input data generated by the input conversion unit 110 as input data (S6). The input conversion unit 110 updates the weight until the error calculated by using the loss function based on the output of the output layer of the neural network and the correct label satisfies the condition of the update end (No in S8). continue.

When the error calculated by using the loss function based on the output of the output layer of the neural network and the correct answer label satisfies the condition of the end of update (Yes in S8), the weight update unit 111 starts from the weight of each layer of the neural network. The configured model parameters and the input API and the output API for accessing the neural network via the communication network N are provided to the user of the model parameters (S10).

When the input conversion unit 110 provides the model parameters and various APIs to the user, the process in this flowchart ends.

<Effects of the information processing device 1 according to the embodiment>
As described above, according to the information processing device 1 according to the embodiment, it is possible to suppress information leakage when the learning model is used via the communication network N.

Although the present invention has been described above using the embodiments, the technical scope of the present invention is not limited to the scope described in the above embodiments, and various modifications and changes can be made within the scope of the gist thereof. be. For example, the specific embodiment of the distribution / integration of the device is not limited to the above embodiment, and all or a part thereof may be functionally or physically distributed / integrated in any unit. Can be done. Also included in the embodiments of the present invention are new embodiments resulting from any combination of the plurality of embodiments. The effect of the new embodiment produced by the combination has the effect of the original embodiment together.

<Modification example>
In the above, the case where the input conversion unit 110 converts the input data input to the neural network by a predetermined conversion procedure has been mainly described. However, if the input data is not information to be kept secret and the output result is information to be kept secret, the input conversion unit 110 may not perform the input conversion but the output conversion unit 112 may only perform the output conversion. ..

Generally, the size of the output data output from the neural network is smaller than the input data input to the neural network. Therefore, by omitting the conversion of the input data, the calculation cost can be suppressed and the learning speed can be expected. effective.

1 ... Information processing device 10 ... Storage unit 11 ... Control unit 110 ... Input conversion unit 111 ... Weight update unit 112 ... Output conversion unit 2 ... User terminal N ... Communication network S: Learning model execution system

Claims (7)

An information processing apparatus for generating a weight of each layer of the neural network comprising a layer of multiple,
An input conversion unit that converts the input data input to the neural network by a predetermined conversion procedure to generate the conversion input data, and
A weight update unit that updates the weights of each layer of the neural network using the conversion input data generated by the input conversion unit as input data is provided.
Predetermined conversion procedure by the input conversion unit executes, in a space defining a vector obtained by arranging the elements constituting the input data, the distance between any two points in the space before the conversion, the space after the conversion Is a conversion procedure in which the distance between the two points in the above is a constant multiple relationship.
Information processing device. The output vector output by the output layer of the neural network has a constant multiple relationship between the distance between any two points in the space where the output vector exists and the distance between the two points in the space after conversion. A conversion procedure further includes an output conversion unit that converts the output vector to generate a conversion output vector.
The information processing device according to claim 1. The neural network is a task image recognition targeting a color image,
The input conversion unit executes the predetermined conversion procedure in a space whose coordinate axes are a plurality of variables for specifying the characteristics of each pixel of the color image.
The information processing device according to claim 1 or 2. The predetermined conversion procedure is any one conversion of expansion conversion, reduction conversion, equal length conversion, and parallel movement conversion, or a composite conversion of two or more conversions.
The information processing device according to any one of claims 1 or 3. An information processing apparatus for generating a weight of each layer of the neural network comprising a layer of multiple,
A weight updating section for updating the weight of each layer of the neural network the academic習用data as input data,
The output vector output by the output layer of the neural network has a constant multiple relationship between the distance between any two points in the space that defines the output vector and the distance between the two points in the space after conversion. In the conversion procedure, the output conversion unit that converts the output vector and generates the conversion output vector,
Information processing device equipped with. A learning method processors of an information processing apparatus for generating a weight of each layer of the neural network comprising a layer of multiple runs, the processor,
A step of converting the input data to be input to the neural network by a predetermined conversion procedure to generate the converted input data, and
The step of updating the weight of each layer of the neural network using the generated conversion input data as input data is executed.
Predetermined conversion procedure in the step of generating, in a space defining a vector obtained by arranging the elements constituting the input data, the distance between any two points in the space before the conversion, the in the space after the conversion The conversion procedure in which the distance between two points has a constant multiple relationship,
Learning method. For information processing devices that generate weights for each layer of a neural network with multiple layers
And generating a converted input data is converted to a predetermined conversion procedure input data to be input before Symbol neural network,
A function of updating the weight of each layer of the neural network using the generated conversion input data as input data is realized.
Predetermined conversion procedure in function of said generated in the space to define a vector obtained by arranging the elements constituting the input data, the distance between any two points in the space before the conversion, the in the space after the conversion The conversion procedure in which the distance between two points has a constant multiple relationship,
program.

Images