Information processor, information processing method and program

Abstract

To provide a technique for detecting an error in processing utilizing a machine learning model.SOLUTION: An information processor includes: an acquisition part for acquiring first data; a model processing part for acquiring a first calculation result for the first data by utilizing a first machine learning model, and acquiring a second calculation result for second data obtained by sign inversion of the first data by utilizing a second machine learning model; and an error detection part for detecting an error, on the basis of a comparison result between the first calculation result and the second calculation result.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present disclosure relates to an information processing device, an information processing method, and a program. [Background technology]

[0002] In recent years, machine learning models for image processing and language processing have come to be widely used. Such machine learning models are not only stored on a server and executed over a network, but also stored on an edge terminal and executed. Edge terminals equipped with such machine learning models are deployed and used in various environments. For example, an edge terminal equipped with a machine learning model for detecting defective products may be deployed on a factory production line. In addition, a mobile robot for pick-and-place equipped with a machine learning model for moving goods is deployed and used in a logistics warehouse. In addition, an edge terminal equipped with a machine learning model for face recognition is deployed at the entrance of an office.

[0003] Such edge terminals may be installed in environments that are not under the supervision of humans. For example, if an edge terminal is placed in an environment that is not suitable for the operation of a processor or the like, noise generated by such an environment may cause errors in the estimation process using a machine learning model by the edge terminal. Alternatively, disturbances (noise) may be injected with the intention of intentionally causing erroneous estimation. For example, malicious attacks using electromagnetic waves, etc. may be performed on edge terminals. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Special Publication No. 2018-524745 Summary of the Invention [Problem to be solved by the invention]

[0005] In a typical example of error detection in a conventional edge terminal, the same estimation process is repeatedly executed, and the occurrence of an error is detected based on whether the calculation results of these estimation processes match. However, if the same noise occurs repeatedly, the same malfunction will be repeated. In this case, even if the estimation process is repeatedly executed, there is a possibility that appropriate error detection will not be possible.

[0006] In view of the above problems, an object of the present disclosure is to provide a technology for detecting errors in processing using a machine learning model. [Means for solving the problem]

[0007] One aspect of the present disclosure relates to an information processing device having an acquisition unit that acquires first data, a model processing unit that utilizes a first machine learning model to acquire a first calculation result for the first data and a second machine learning model to acquire a second calculation result for second data obtained by inverting the sign of the first data, and an error detection unit that detects an error based on a comparison result between the first calculation result and the second calculation result. Effect of the Invention

[0008] According to the present disclosure, it is possible to provide a technique for detecting errors in processing using a machine learning model. [Brief description of the drawings]

[0009] [Figure 1] FIG. 1 is a schematic diagram illustrating an information processing system according to an embodiment of the present disclosure. [Diagram 2] FIG. 2 is a schematic diagram illustrating an error detection process according to one embodiment of the present disclosure. [Diagram 3] FIG. 3 is a schematic diagram illustrating an error detection process according to one embodiment of the present disclosure. [Figure 4] FIG. 4 is a block diagram illustrating a hardware configuration of an edge terminal according to an embodiment of the present disclosure. [Diagram 5]FIG. 5 is a block diagram illustrating a functional configuration of an edge terminal according to an embodiment of the present disclosure. [Figure 6] FIG. 6 is a diagram illustrating a machine learning model according to one embodiment of the present disclosure. [Figure 7] 7A and 7B are diagrams illustrating activation functions for positive and negative estimation according to one embodiment of the present disclosure. [Figure 8] 8A and 8B are diagrams illustrating activation functions for positive and negative estimation according to one embodiment of the present disclosure. [Figure 9] FIG. 9 is a flowchart illustrating an error detection process according to an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings.

[0011] In the following embodiment, an information processing device that detects an error in an estimation process using a machine learning model is disclosed.

[0012] [Summary of the Disclosure] First, an information processing system according to an embodiment of the present disclosure will be described. Fig. 1 is a schematic diagram showing an information processing system according to an embodiment of the present disclosure.

[0013] As shown in FIG. 1, an information processing system 10 includes an administrator terminal 20 and an edge terminal 100, and the administrator terminal 20 and the edge terminal 100 can be communicatively connected via a network such as the Internet.

[0014] The administrator terminal 20 may be realized by, for example, an information processing terminal such as a personal computer, a smartphone, or a tablet of an administrator who manages the edge terminal 100. For example, the administrator may be a user of the edge terminal 100, an operator of a machine learning model installed in the edge terminal 100, an administrator who manages other edge terminals 100, or the like, or may not be such an administrator.

[0015] The edge terminal 100 may be realized by an information processing device that performs error detection in the estimation process of the machine learning model as described in detail below and is equipped with the machine learning model. For example, if the machine learning model is related to image processing, the edge terminal 100 may be equipped with a camera. Also, if the machine learning model is related to voice processing, the edge terminal 100 may be equipped with a microphone. When an error is detected in the estimation process of the machine learning model, the edge terminal 100 may notify the administrator terminal 20 of the detection of the error. Note that, although the following embodiment focuses on the edge terminal 100, the error detection process according to the present disclosure is not necessarily limited to the edge terminal 100 and may be applied to any information processing device.

[0016] 2 is a schematic diagram illustrating an error detection process according to an embodiment of the present disclosure. As shown in FIG. 2, when the edge terminal 100 receives input data, it inputs the received input data to a first neural network model and executes a positive estimation process to execute a calculation process in the first neural network model. In addition to the positive estimation process, the edge terminal 100 inverts the sign of the received data, inputs the sign-inverted data to a second neural network model, and executes a negative estimation process to execute a calculation process in the second neural network model. For example, when the input data is a two-dimensional pixel value [a ij ], the sign-inverted data is expressed in matrix form as [-a ij ] matrix format data.

[0017] That is, in the positive estimation process, a first neural network model is executed on the input data, while in the negative estimation process, a second neural network model is executed on the sign-inverted data of the input data. As shown in the figure, the edge terminal 100 obtains a positive estimation calculation result from the positive estimation process, and obtains a negative estimation calculation result from the negative estimation process. For example, the positive estimation calculation result can be the calculation result of each node of the first neural network model in the positive estimation process, and the negative estimation calculation result can be the calculation result of each node of the second neural network model in the negative estimation process.

[0018] Then, the edge terminal 100 compares the positive estimation calculation result with the negative estimation calculation result, and determines the occurrence of an error in the calculation process of the machine learning model based on the comparison result. For example, when the calculation process is normally performed in the first neural network model and the second neural network model in which the activation function is point-symmetric with respect to the origin, the positive estimation calculation result and the negative estimation calculation result at each node can basically be numerical values ​​with inverted signs. Therefore, the sum of the positive estimation calculation result and the negative estimation calculation result can be zero. On the other hand, when the calculation process of the first neural network model and / or the second neural network model is not normally performed due to noise or the like, the positive estimation calculation result and the negative estimation calculation result at each node do not become numerical values ​​with inverted signs, and the sum of the positive estimation calculation result and the negative estimation calculation result can be non-zero. In this way, the edge terminal 100 compares the positive estimation calculation result and the negative estimation calculation result at the nodes of the first neural network model and the second neural network model, and can detect the occurrence of an error according to the comparison result.

[0019] The comparison between the positive estimation calculation result and the negative estimation calculation result does not need to be performed after the execution of the positive estimation process and the negative estimation process for all nodes of the first neural network model and the second neural network model as shown in Fig. 2, but may be performed after the execution of the positive estimation process and the negative estimation process for each node as shown in Fig. 3. For example, if the sum of the positive estimation calculation result and the negative estimation calculation result for a predetermined number of nodes is non-zero, it may be determined that an error has occurred. This makes it unnecessary to determine that an error has occurred and to execute the positive estimation process and the negative estimation process for further nodes when it is detected that the sum of the positive estimation calculation result and the negative estimation calculation result for a predetermined number of nodes is non-zero without executing the positive estimation process and the negative estimation process for all nodes.

[0020] Here, the edge terminal 100 may have, for example, a hardware configuration as shown in Fig. 4. That is, the edge terminal 100 has a drive device 101, a storage device 102, a memory device 103, a processor 104, an interface device 105, and a communication device 106, which are interconnected via a bus B.

[0021] Programs or instructions for implementing various functions and processes in the edge terminal 100 may be stored in a removable storage medium such as a CD-ROM (Compact Disk-Read Only Memory) or a flash memory. When the storage medium is set in the drive device 101, the programs or instructions are installed from the storage medium via the drive device 101 into the storage device 102 or the memory device 103. However, the programs or instructions do not necessarily have to be installed from the storage medium, and may be downloaded from any external device via a network or the like.

[0022] The storage device 102 is realized by a hard disk drive or the like, and stores installed programs or instructions as well as files, data, etc. used in executing the programs or instructions.

[0023] The memory device 103 is realized by a random access memory, a static memory, or the like, and when a program or an instruction is activated, reads and stores the program or instruction, data, or the like from the storage device 102. The storage device 102, the memory device 103, and the removable storage medium may be collectively referred to as a non-transitory storage medium.

[0024] The processor 104 may be realized by one or more CPUs (Central Processing Units), GPUs (Graphics Processing Units), processing circuitry, etc., which may be composed of one or more processor cores, and performs various functions and processing of the edge terminal 100 in accordance with programs, instructions, data such as parameters necessary to execute the programs or instructions, etc. stored in the memory device 103.

[0025] The interface device 105 provides an interface between an external device and the edge terminal 100 .

[0026] The communication device 106 is realized by various communication circuits that execute wired and / or wireless communication processing with an external device, the Internet, a LAN (Local Area Network), a cellular network, or other communication network.

[0027] However, the above-described hardware configuration is merely an example, and the edge terminal 100 according to the present disclosure may be realized by any other suitable hardware configuration.

[0028] [Edge terminal] Next, an edge terminal 100 according to an embodiment of the present disclosure will be described. Fig. 5 is a block diagram showing a functional configuration of the edge terminal 100 according to an embodiment of the present disclosure. As shown in Fig. 5, the edge terminal 100 has an acquisition unit 110, a model processing unit 120, and an error detection unit 130. Each of the functional units of the acquisition unit 110, the model processing unit 120, and the error detection unit 130 may be realized by the processor 104 executing a computer program stored in the memory device 103 of the edge terminal 100.

[0029] The acquisition unit 110 acquires input data. For example, if the machine learning model is related to image processing, the acquisition unit 110 may acquire image data as the input data. Also, if the machine learning model is related to voice processing, the acquisition unit 110 may acquire voice data as the input data. The acquisition unit 110 passes the acquired input data to the model processing unit 120.

[0030] The model processing unit 120 uses a first machine learning model to obtain a positive estimation calculation result for the input data, and uses a second machine learning model to obtain a negative estimation calculation result for sign-inverted data obtained by inverting the sign of the input data. For example, when the machine learning model is realized by a neural network, in the positive estimation process, the model processing unit 120 inputs the input data to an input layer of the first machine learning model, and obtains output data from the output layer via one or more hidden layers. On the other hand, in the negative estimation process, the model processing unit 120 converts the input data into sign-inverted data, inputs the sign-inverted data to an input layer of the second machine learning model, and obtains output data from the output layer via one or more hidden layers.

[0031] For example, when the first machine learning model has a model architecture as shown in Fig. 6, in the positive estimation process, the model processing unit 120 inputs the input data acquired from the acquisition unit 110 to three nodes x0, x1, and x2 of the input layer of the first neural network model. After that, values ​​weighted by weighting coefficients assigned to each edge between the three nodes x0, x1, and x2 of the input layer are input to two nodes y0 and y1 of the hidden layer 1. For example, when the activation function is f, the nodes y0 and y1 output the calculation results according to the following formula.

number

[0032] Similarly, the two nodes s0 and s1 in hidden layer 2 output the calculation results according to the following equations.

number

[0033] In addition, the three nodes z0, z1, and z2 in the output layer output the calculation results according to the following equation.

number

[0034] In this way, the model processing unit 120 obtains the positive estimation calculation results at each of the nodes y0, y1, s0, s1, z0, z1, and z2 in the positive estimation process, and passes the obtained positive estimation calculation results to the error detection unit 130.

[0035] On the other hand, in the negative estimation process, the model processing unit 120 inverts the sign of the input data acquired from the acquisition unit 110, and inputs the sign-inverted data to three nodes x0, x1, and x2 in the input layer of the second neural network model. After that, values ​​weighted by weighting coefficients assigned to each edge between the three nodes x0, x1, and x2 in the input layer are input to two nodes y0 and y1 in the hidden layer 1. For example, the activation function of the second neural network model is

number

number

number

[0036] Similarly, the two nodes s0 and s1 in hidden layer 2 output the calculation results according to the following equations.

number

[0037] In addition, the three nodes z0, z1, and z2 in the output layer output the calculation results according to the following equation.

number

[0038] In this way, the model processing unit 120 obtains the negative estimation calculation results at each node y0, y1, s0, s1, z0, z1, and z2 in the negative estimation process, and passes the obtained negative estimation calculation results to the error detection unit 130.

[0039] The error detection unit 130 detects an error based on a comparison result between the positive estimation calculation result and the negative estimation calculation result. For example, the positive estimation calculation result may be a calculation result at a node of the machine learning model acquired by inputting input data to the machine learning model, and the negative estimation calculation result may be a calculation result at the node acquired by inputting sign-inverted data to the machine learning model. In the case of the machine learning model shown in FIG. 6, the error detection unit 130 may compare the positive estimation calculation result and the negative estimation calculation result at each node of y0, y1, s0, s1, z0, z1, and z2 acquired by the model processing unit 120, and determine whether an error has occurred in the calculation process of the machine learning model based on the comparison result. At this time, it is not necessary to perform the comparison after the execution of the positive estimation process and the negative estimation process for each node of the first neural network model and the second neural network model, and after the execution of the positive estimation process and the negative estimation process on a node-by-node basis, the error detection unit 130 may compare the positive estimation calculation result and the negative estimation calculation result on a node-by-node basis. For example, if the sum of the positive estimation calculation result and the negative estimation calculation result for a predetermined number of nodes is non-zero, the error detection unit 130 may determine that an error has occurred. This makes it unnecessary to perform the positive estimation process and the negative estimation process for all nodes, and to determine that an error has occurred when it is detected that the sum of the positive estimation calculation result and the negative estimation calculation result for a predetermined number of nodes is non-zero, and to perform the positive estimation process and the negative estimation process for further nodes.

[0040] For example, the activation function of the first machine learning model and the activation function of the second machine learning model may be for positive estimation and for negative estimation, and may be point-symmetric functions. In this case, the error detection unit 130 may detect an error based on whether the sum of the positive estimation calculation result and the negative estimation calculation result is zero or not. For example, a function f as shown in FIG. 7 may be used as such an activation function. In the positive estimation process, as shown in FIG. 7A, f(y) is output for an input y. On the other hand, in the negative estimation process, as shown in FIG. 7B,

number

number

[0041] Such a point-symmetric activation function is not limited to the function shown in Fig. 7, and may be, for example, a sigmoid function as shown in Fig. 8. In positive estimation, as shown in Fig. 8A, f(y) is output for an input y. On the other hand, in negative estimation, as shown in Fig. 8B,

number

number

[0042] In the above-mentioned embodiment, the positive estimation process and the negative estimation process are performed by inputting the input data and the sign-inverted data to the first neural network model and the second neural network model, which have activation functions that are point-symmetric with each other and have the same parameters, but the present disclosure is not limited to this. In one embodiment, the model processing unit 120 may obtain a negative estimation calculation result for the input data by using a sign-inverted model as a third neural network model, which has the same activation function as the first neural network and has parameters obtained by inverting the parameters of the first neural network model. That is, the error detection unit 130 may detect an error based on a comparison result between the positive estimation calculation result obtained by inputting the input data to the first neural network model and the negative estimation calculation result obtained by inputting the input data to the sign-inverted model.

[0043] Specifically, when the machine learning model shown in FIG. 6 is used for positive estimation processing, the weighting coefficient yw n ,sw m ,zw l and bias value yb n ,sb m ,zb l The error detection unit 130 may compare the positive estimation calculation result at each node acquired by inputting the input data to a machine learning model for positive estimation processing with the negative estimation calculation result at each node acquired by inputting the input data to a sign-inversion model for negative estimation processing, and detect an error based on the comparison result.

[0044] In this way, when the comparison result between the positive estimation calculation result and the negative estimation calculation result in the node is obtained, if the comparison result in at least one node indicates the occurrence of an error, the error detection unit 130 determines that an error has occurred in the calculation process of the machine learning model, and may redo the estimation calculation of the node if the number of times is less than a predetermined number, or may notify the administrator terminal 20 of the occurrence. However, the trigger for such notification of the occurrence of an error is not limited to this, and the notification may be made when the comparison result in a predetermined number or more nodes indicates the occurrence of an error. This may avoid excessive notification of the occurrence of an error.

[0045] In addition, in the computational processing of a neural network, computations are typically performed in sequence from the input layer through the hidden layer to the output layer, but when an error is detected in a predetermined number of nodes or more, computations in subsequent layers and nodes may be stopped. This can prevent unnecessary computations from continuing.

[0046] [Error detection processing] Next, an error detection process according to an embodiment of the present disclosure will be described. Fig. 9 is a flowchart showing the error detection process according to an embodiment of the present disclosure. The error detection process may be performed, for example, by the edge terminal 100, and more specifically, may be realized by the processor 104 of the edge terminal 100 executing a computer program or instructions stored in the memory device 103.

[0047] In step S101, the edge terminal 100 accepts input data. For example, the edge terminal 100 can acquire image data, audio data, and the like as the input data.

[0048] In step S102, the edge terminal 100 uses the first machine learning model to obtain a positive estimation calculation result for the input data, and uses the second machine learning model to obtain a negative estimation calculation result for the sign-inverted data obtained by inverting the sign of the input data. Specifically, the edge terminal 100 obtains the positive estimation calculation result and the negative estimation calculation result at each node of the machine learning model.

[0049] In step S103, the edge terminal 100 detects an error based on the comparison result between the positive estimation calculation result and the negative estimation calculation result. Specifically, when the edge terminal 100 detects that an error has occurred in the calculation process in a predetermined number of nodes or more based on the comparison result between the positive estimation calculation result and the negative estimation calculation result in each node, the edge terminal 100 may notify the administrator terminal 20 of the occurrence of the error.

[0050] For example, when the first machine learning model and the second machine learning model use activation functions that are point-symmetric with respect to the origin, the edge terminal 100 may calculate the sum of the positive estimation calculation result and the negative estimation calculation result at the node, and determine whether an error has occurred in the calculation process at the node based on whether the calculated sum is zero or not. Specifically, when the calculated sum is zero, the edge terminal 100 may determine that no error has occurred in the calculation process at the node. On the other hand, when the calculated sum is non-zero, the edge terminal 100 may determine that an error has occurred in the calculation process at the node.

[0051] According to the above-described embodiment, it becomes possible to appropriately detect the occurrence of an error in calculation processing due to noise or the like in the edge terminal 100 that is placed in an environment where monitoring by an administrator is difficult.

[0052] Although the examples of the present disclosure have been described in detail above, the present disclosure is not limited to the specific embodiments described above, and various modifications and variations are possible within the scope of the gist of the present disclosure described in the claims. [Explanation of symbols]

[0053] 10. Information Processing Systems 20 Administrator terminal 100 Edge Devices 110 Acquisition Department 120 Model processing section 130 Error detection unit

Claims

1. An acquisition unit that acquires first data; a model processing unit that uses a first machine learning model to obtain a first operation result for the first data, and uses a second machine learning model to obtain a second operation result for second data obtained by inverting the sign of the first data; an error detection unit that detects an error based on a comparison result between the first calculation result and the second calculation result; An information processing device having the above configuration.

2. the first operation result is an operation result at a first node of the first machine learning model obtained by inputting the first data into the first machine learning model; The information processing device according to claim 1 , wherein the second calculation result is a calculation result at the first node obtained by inputting the second data to the second machine learning model.

3. The information processing device according to claim 1 , wherein the activation function of the first machine learning model and the activation function of the second machine learning model are point-symmetric functions.

4. The information processing device according to claim 3 , wherein the error detection unit detects the error based on whether or not a sum of the first calculation result and the second calculation result is zero.

5. the model processing unit obtains a third operation result for the first data by using a third machine learning model having parameters obtained by inverting the signs of the parameters of the first machine learning model; The information processing device according to claim 1 , wherein the error detection unit detects an error based on a comparison result between the first calculation result and the third calculation result.

6. Obtaining first data; Obtaining a first operation result for the first data by using a first machine learning model, and a second operation result for second data obtained by inverting the sign of the first data by using a second machine learning model; Detecting an error based on a comparison result between the first calculation result and the second calculation result; An information processing method implemented by a computer.

7. Obtaining first data; Obtaining a first operation result for the first data by using a first machine learning model, and a second operation result for second data obtained by inverting the sign of the first data by using a second machine learning model; Detecting an error based on a comparison result between the first calculation result and the second calculation result; A program that causes a computer to execute the following.

Images