JP7325270B2 - Control program, control method and control device - Google Patents

Description

The present invention relates to a control program, control method and control device .

Conventionally, there is a speech recognition device that performs speech recognition using a plurality of models generated by machine learning. In such a speech recognition apparatus, out of output data obtained from a plurality of models, data with the highest matching degree is regarded as a speech recognition result (see, for example, Patent Document 1).

JP-A-6-67698

However, in the prior art, since it is necessary to store a plurality of models in memory, a storage medium with a large memory capacity is required.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a control device and a control method that can use a plurality of models while suppressing the memory capacity.

In order to solve the above-described problems and achieve the object, a control device according to an embodiment includes a storage section and a switching section. The storage unit stores a model configured by a neural network. The switching unit pseudo-switches the number of nodes of the neural network by changing parameters of the model stored in the storage unit.

According to the present invention, a plurality of models can be used while suppressing memory capacity.

FIG. 1A is a diagram showing an example of mounting a control device. FIG. 1B is a diagram showing an overview of the control method. FIG. 2 is a block diagram of the control device. FIG. 3 is a diagram illustrating an example of a parameter DB; 4A is a diagram (part 1) illustrating an example of parameters switched by a switching unit; FIG. FIG. 4B is a diagram (part 2) illustrating an example of parameters switched by the switching unit; FIG. 5 is a diagram (part 3) illustrating an example of parameters switched by the switching unit; FIG. 6 is a flow chart showing a processing procedure executed by the control device.

Hereinafter, a control device and a control method according to embodiments will be described with reference to the accompanying drawings. In addition, this invention is not limited by embodiment shown below.

First, an outline of a control device and a control method according to an embodiment will be described with reference to FIGS. 1A and 1B. FIG. 1A is a diagram showing an example of mounting a control device. FIG. 1B is a diagram showing an overview of the control method. This control method is executed by the control device 1 shown in FIG. 1A.

As shown in FIG. 1A, a control device 1 according to the embodiment is mounted on a vehicle 100. As shown in FIG. The control device 1 generates the output data Do by inputting the input data Di based on the sensor information input from the vehicle-mounted sensors of the vehicle 100 to the model M. FIG. Here, the model M has layers called an input layer, an intermediate layer (hidden layer), and an output layer, and transmits input data Di input from the input layer to the output layer while propagating each connection path. . Each layer is composed of nodes N, respectively.

At this time, the output data Do corresponding to the input data Di is generated by executing arithmetic processing based on the parameters set for each connection path.

By the way, it is conceivable that a plurality of models are stored and used by switching between them according to input data or the like. In this case, the control device needs to store a plurality of models in advance, which causes pressure on the memory capacity. Moreover, as described above, when a plurality of models are mounted on a vehicle, it is necessary to conduct inspections such as quality assurance for each model, which may lead to an increase in cost.

Therefore, in the control method according to the embodiment, one model M is used as a plurality of models M by switching the neural network structure of the model M.

Specifically, as shown in FIG. 1B, in the control method according to the embodiment, it is possible to switch between the large-scale model M1 and the small-scale model M2 having fewer nodes than the large-scale model M1.

More specifically, in the control method according to the embodiment, by changing the parameters of the model M, it is possible to switch between the large-scale model M1 and the small-scale model M2. Here, the parameters include weight parameters and bias parameters.

In the example shown in FIG. 1B, the small-scale model M2 has fewer intermediate layers L than the large-scale model M1. In this case, by setting the weight parameter of the output path of each node of the intermediate layer L to "1" and the bias parameter of each node to "0", the model M with the intermediate layer L substantially reduced can be obtained. can.

Also, when switching from the small-scale model M2 to the large-scale model M1, the weight parameter and the bias parameter may be set for the large-scale model M1.

Thus, in the control method according to the embodiment, it is possible to switch one model M between the large-scale model M1 and the small-scale model M2. Therefore, in the control method according to the embodiment, since it is sufficient to store one model M, it is possible to use a plurality of models M while suppressing the memory capacity.

In the above example, the model M is switched between the large-scale model M1 and one small-scale model M2. may be Also, in the above-described embodiment, the case where the intermediate layer L is increased or decreased has been described, but the present invention is not limited to this. That is, arbitrary nodes N can be increased or decreased.

Next, a configuration example of the control device 1 according to the embodiment will be described with reference to FIG. FIG. 2 is a block diagram of the control device 1. As shown in FIG. 2 also shows on-vehicle sensors 101. As shown in FIG. The in-vehicle sensors 101 are, for example, sensors that detect the state of the internal combustion engine (engine) of the vehicle 100 . By inputting sensor information input from the in-vehicle sensors 101 to the model M, the control device 1 can determine a state such as an engine misfire.

As shown in FIG. 2 , the control device 1 includes a storage section 2 and a control section 3 . The storage unit 2 is realized by, for example, a semiconductor memory device such as a RAM or flash memory, or a storage device such as a hard disk or an optical disk.

Moreover, as shown in FIG. 2, the memory|storage part 2 is provided with model DB21 and parameter DB22. The model DB 21 is a database that stores the model M. FIG. The model DB 21 stores the large-scale model M1 described above.

The parameter DB 22 is a database that stores parameters applied to the model M. FIG. FIG. 3 is a diagram showing an example of the parameter DB 22. As shown in FIG. As shown in FIG. 3, the parameter DB 22 stores "parameter ID", "application condition", "weight parameter", "bias parameter" and the like in association with each other.

A parameter ID is an identifier that identifies a parameter, and can be said to be an identifier that identifies the large-scale model M1 and the small-scale model M2. The applicable condition indicates the condition under which the corresponding parameter is applied. Also, the weight parameter and the bias parameter are examples of the parameters described above. Also, these parameters are values derived by machine learning. In the example shown in FIG. 3, the weight parameter and the bias parameter are schematically shown as "C01" and "D01", but specific information is described in "C01" and "D01". shall be

Also, the number of parameters corresponding to the number of nodes of the model M is described in "C01" and "D01". Although the details will be described later, when switching from the large-scale model M1 to the small-scale model M2, the weight parameter and bias parameter of the nodes to be reduced are "0" or "1".

Returning to the description of FIG. 2, the control unit 3 will be described. The control unit 3 is a controller, and is realized by, for example, executing various programs stored in the storage unit 2 using a RAM as a work area by a CPU, MPU, or the like. Also, the control unit 3 can be realized by an integrated circuit such as an ASIC or FPGA, for example.

As shown in FIG. 2 , the control unit 3 includes an acquisition unit 31 , a determination unit 32 , a switching unit 33 and a calculation unit 34 . The acquisition unit 31 acquires sensor information from the in-vehicle sensors 101 . The sensor information acquired by the acquisition unit 31 is notified to the determination unit 32 .

The determination unit 32 determines whether or not to switch the number of nodes of the model M based on a predetermined condition. Specifically, the determination unit 32 refers to the parameter DB 22 and determines which application condition is met.

Here, the accuracy required for the model M can be cited as an example of applicable conditions. That is, when the accuracy is high, the large-scale model M1 is preferable, and when the accuracy is low, even the small-scale model M2 can be used.

The determination unit 32 can determine which application condition is met, for example, based on the accuracy required by a host device (not shown). Then, the determination unit 32 notifies the switching unit 33 of the parameter ID as the determination result.

The above predetermined conditions are not limited to the accuracy required for the model M. For example, they may be read from sensor information such as the number of engine revolutions, or from other information such as the current running position of the vehicle 100. It may be readable. That is, the predetermined condition can be arbitrarily set.

The switching unit 33 changes the parameters of the model M stored in the storage unit 2 to pseudo-switch the number of nodes of the neural network. Here, the switching unit 33 adjusts the number of nodes of the model M by changing parameters based on the determination result of the determination unit 32 described above. As a result, it is possible to set the model M with the number of nodes according to the current requirements.

Specifically, the switching unit 33 sets the weight parameter and the bias parameter of the parameter ID notified from the determination unit 32 to the model M, thereby switching the number of nodes of the model M in a pseudo manner.

4A, 4B, and 5 are diagrams showing specific examples of parameters switched by the switching unit 33. FIG. First, with reference to FIGS. 4A and 4B, a case of pseudo reduction of the intermediate layer L2 will be described. Note that "A1", "B1", "C1", etc. shown in FIG. 4A indicate the bias parameters of the corresponding node N, and "a11", "b11", etc. indicate the weights in the connection paths between the corresponding nodes N. shall indicate the parameters. Also, FIG. 4A is described assuming that the activation function is ReLU.

As shown in FIG. 4A, when the intermediate layer L2 consisting of the nodes N21 to N23 is pseudo-reduced, the bias parameters B1 to B3 of the nodes N21 to N23 are switched, and the weight parameter b11 in the output path of the nodes N21 to N23 ~b33 is switched.

Specifically, the switching unit 33 switches all of the bias parameters B1 to B3 of the nodes N21 to N23 to "0", sets all of the weight parameters b11, b22, and b33 to "1", and sets the other weight parameters (b12, b13 , b21, b23, b31, b32) are all switched to "0".

That is, the switching unit 33 switches the bias parameter of the node N to be reduced to "0", sets the weight parameter of the output path from the node N to the node N extending horizontally in the figure to "1", and sets the weight parameter of the output path to "1". to "0".

As a result, as shown in FIG. 4B, the nodes N21 to N23 of the intermediate layer L2 are virtually eliminated, and the nodes N11 to N13 are connected to the nodes N31 to N33 by the weight parameters a11 to a33, respectively. It is possible to switch.

In other words, the data input to the intermediate layer L2 is input to the next nodes N31 to N33 without being changed in the intermediate layer L2. In other words, the intermediate layer L2 can be rendered substantially meaningless.

Thus, the switching unit 33 can easily reduce the size of the model M by switching the structure of the model M on a layer-by-layer basis.

Next, referring to FIG. 5, a case of pseudo-removing an arbitrary node N will be described. Here, the activation function is assumed to be ReLU or Tanh.

When the lower node column Nc shown in FIG. 5 is pseudo-reduced, the switching unit 33 sets the bias parameter and the weight parameter for the node column Nc to "0".

Specifically, the switching unit 33 sets the bias parameters A3, B3, and C3 of the node N13, the node N23, and the node N33, which constitute the node string Nc, to all "0", and Switch all weight parameters to '0'.

Here, the weight parameters for the node N13, the node N23, and the node N33 are the weight parameters (a13, a23, a33, b13, b23, b33) of the input paths for the node N13, the node N23, and the node N33, and the weight parameters for the node N13, the node N13, and the node N33. Both N23 and node N33 output path weight parameters (a31, a32, a33, b31, b32, b33) are shown.

That is, data "0" is input to the nodes N13, N23 and N33, and "0" is output from the nodes N13, N23 and N33 to the following nodes N. Become. It should be noted that the data output from the nodes N13, N23, and N33 need only be "0", so only the weight parameter of the output path is switched while the bias parameter and the weight parameter of the input path are held. good too.

Although the case of artificially reducing the node string Nc has been described here, it is also possible to reduce arbitrary nodes N in units of one node. It should be noted that here, when restoring the pseudo-reduced node N, the weight parameter and the bias parameter are respectively restored.

Returning to the description of FIG. 2, the calculation unit 34 will be described. The calculation unit 34 inputs the input data Di based on the sensor information to the model M to generate the output data Do. Here, as described above, the number of nodes of the model M used by the calculation unit 34 for calculation processing is pseudo-changed by the switching unit 33 .

That is, the calculation unit 34 performs calculation processing using either the large-scale model M1 or the small-scale model M2.

Next, a processing procedure executed by the control device 1 according to the embodiment will be described with reference to FIG. FIG. 6 is a flowchart showing a processing procedure executed by the control device 1. As shown in FIG. Note that the processing procedure described below is executed by the control unit 3 each time sensor information is acquired.

As shown in FIG. 6, when the control device 1 acquires sensor information (step S101), it determines whether high accuracy is required (step S102). If the controller 1 determines that high accuracy is required (step S102, Yes), it sets parameters for the large-scale model M1 in the model M (step S103).

After that, the control device 1 performs arithmetic processing using the large-scale model M1 (step S104), and terminates the processing. If the control device 1 determines in the determination process of step S102 that high accuracy is not required (step S102, No), it sets parameters for the small-scale model M2 in the model M (step S105).

After that, the control device 1 performs arithmetic processing using the small-scale model M2 (step S104), and terminates the processing.

As described above, the control device 1 according to the embodiment includes the storage section 2 and the switching section 33 . The storage unit 2 stores a model M configured by a neural network. The switching unit 33 changes the parameters of the model M stored in the storage unit 2 to pseudo-switch the number of nodes of the neural network. Therefore, according to the control device 1 according to the embodiment, it is possible to use a plurality of models while suppressing the memory capacity.

By the way, in the above-described embodiment, the case where the large-scale model M1 and the small-scale model M2 are models M with different accuracies has been described, but the present invention is not limited to this. That is, the large-scale model M1 and the small-scale model M2 may be models M having completely different purposes. That is, for example, knocking may be determined using the large-scale model M1, and engine failure may be determined using the small-scale model M2.

Also, in the above-described embodiment, the case where the control device 1 is mounted on the vehicle 100 has been described, but the present invention is not limited to this. That is, the present invention can be applied to various devices including personal computers.

Further effects and modifications can be easily derived by those skilled in the art. Therefore, the broader aspects of the invention are not limited to the specific details and representative embodiments so shown and described. Accordingly, various changes may be made without departing from the spirit or scope of the general inventive concept defined by the appended claims and equivalents thereof.

Reference Signs List 1 control device 31 acquisition unit 32 determination unit 33 switching unit 34 calculation unit M model

Claims (6)

A control program to be executed by a computer,
Among the parameters of the model configured by the neural network stored in the storage unit , the weight parameter of the node to be increased or decreased and the bias parameter are changed to change the number of nodes of the neural network in a pseudo manner.
control program. 2. The control program according to claim 1 , further comprising determining whether or not switching of the number of nodes is necessary, and switching the number of nodes according to a determination result. 3. The control program according to claim 2, further comprising determining whether the number of nodes needs to be switched according to the accuracy required for the model. When removing at least one layer that constitutes the neural network, setting the weight parameter of each output path of the layer to 1 and setting the bias parameter of each node of the intermediate layer to 0.
A control program according to any one of claims 1 to 3. A control method for controlling a control device,
Among the parameters of the model configured by the neural network stored in the storage unit , the weight parameter of the node to be increased or decreased and the bias parameter are changed to change the number of nodes of the neural network in a pseudo manner.
control method. The control unit of the control device is
Among the parameters of the model configured by the neural network stored in the storage unit, the weight parameter of the node to be increased or decreased and the bias parameter are changed to change the number of nodes of the neural network in a pseudo manner.
Control device.

Images