JP7700501B2 - Inference device, inference method, and program - Google Patents

Description

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

Various tasks such as image recognition have traditionally been performed using machine learning models. In addition, when performing a certain task, it has also been traditional to select an appropriate machine learning model from among multiple machine learning models. For example, Patent Literature 1 discloses a technology for selecting a CNN model suitable for a database from among multiple CNN (Convolutional Neural Network) models.

JP 2018-92614 A

However, CNN models generally use a large amount of hardware resources such as RAM (Random Access Memory). For this reason, in Patent Document 1, no matter which CNN model is selected from among multiple CNN models, it is not expected that the amount of hardware resources used will be significantly reduced. In addition, in Patent Document 1, the CNN model is selected before inference is performed, and the CNN model cannot be dynamically changed during inference.

On the other hand, depending on the situation during inference, it may be possible to apply not only general-purpose machine learning models such as CNN models, but also lighter machine learning models specialized for specific situations. For this reason, when the situation during inference is specific, it is believed that the amount of hardware resources used can be significantly reduced by performing inference using a lightweight machine learning model specialized for that situation. Furthermore, since the situation during inference can change from moment to moment, it is considered preferable to be able to dynamically change the machine learning model used for inference.

One embodiment of the present invention has been made in consideration of the above points, and aims to dynamically select an appropriate model from among multiple models depending on the situation at the time of inference.

To achieve the above object, an inference device according to one embodiment is an inference device that repeatedly executes inference of a predetermined task for each of a plurality of data, and has an inference unit that executes inference of the task for the data using a selection model indicating a model selected from a plurality of models, and a selection unit that selects a selection model from the plurality of models for executing inference in the next iteration based on the result of the inference.

The appropriate model can be dynamically selected from multiple models depending on the situation at the time of inference.

FIG. 2 is a diagram illustrating an example of a hardware configuration of the dynamic model switching inference device according to the present embodiment. FIG. 2 is a diagram illustrating an example of a functional configuration of a dynamic model switching inference device according to the present embodiment. FIG. 11 is a diagram illustrating an example of recognition state pattern information. FIG. 11 is a diagram illustrating an example of model type information. 13 is a flowchart showing an example of the flow of a dynamic model switching inference process. 13 is a flowchart showing an example of the flow of a model selection process. FIG. 2 is a diagram showing an example of state variable values at time t-1 and time t. FIG. 2 is a diagram showing an example of changes in state variable values at time t relative to time t−1.

One embodiment of the present invention will be described below. In this embodiment, a dynamic model switching inference device 10 will be described that is capable of dynamically selecting an appropriate model from among multiple models depending on the situation at the time of inference of an image recognition task of recognizing the type of object in an image and its position, and then switching to the selected model to perform inference. Note that in this embodiment, a model refers to a machine learning model (image recognition model) or an image recognition method that can be applied to an image recognition task.

In this embodiment, each model recognizes the type of object in an image (object type) and the position of the object in the image (image position), and calculates the recognition rate. In other words, each model outputs the recognition rate, object type, and image position. The recognition rate is an index value that represents the accuracy of feature recognition by the model, and represents, for example, the accuracy of the object type and image position output by the model.

Note that, since the object type and the position in the image are features in the image recognized by the model, hereinafter, the object type and the position in the image are also referred to as "features", respectively. Furthermore, since the recognition rate, the object type, and the position in the image are variables that represent the recognition state of the model, the recognition rate, the object type, and the position in the image are also referred to as "state variables", respectively.

However, in addition to the recognition rate, object type, and position in the image, various variables that represent the recognition state of the model in the image recognition task may be used as state variables. For example, variables that represent the distance from the imaging device to the object, conditions that affect the recognition rate of image recognition (e.g., lighting conditions in the shooting environment, time of day, etc.) may be used as state variables.

<Hardware Configuration of the Dynamic Model Switching Inference Device 10>
First, a hardware configuration of a dynamic model switching inference device 10 according to the present embodiment will be described with reference to Fig. 1. Fig. 1 is a diagram showing an example of a hardware configuration of the dynamic model switching inference device 10 according to the present embodiment.

As shown in FIG. 1, the dynamic model switching inference device 10 according to this embodiment is realized by a general computer or computer system, and has an input device 101, a display device 102, an external I/F 103, a communication I/F 104, a processor 105, and a memory device 106. Each of these pieces of hardware is connected to each other so as to be able to communicate with each other via a bus 107.

The input device 101 is, for example, a keyboard, a mouse, a touch panel, etc. The display device 102 is, for example, a display, etc. Note that the dynamic model switching inference device 10 does not necessarily have to have at least one of the input device 101 and the display device 102.

The external I/F 103 is an interface with an external device such as a recording medium 103a. The dynamic model switching inference device 10 can read and write data from and to the recording medium 103a via the external I/F 103. Examples of the recording medium 103a include a CD (Compact Disc), a DVD (Digital Versatile Disk), a SD memory card (Secure Digital memory card), and a USB (Universal Serial Bus) memory card.

The communication I/F 104 is an interface for connecting the dynamic model switching inference device 10 to a communication network. The processor 105 is, for example, various types of arithmetic devices such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit). The memory device 106 is, for example, various types of storage devices such as a HDD (Hard Disk Drive), SSD (Solid State Drive), RAM, ROM (Read Only Memory), and flash memory.

The dynamic model switching inference device 10 according to this embodiment has the hardware configuration shown in FIG. 1, and is therefore capable of implementing the dynamic model switching inference process described below. Note that the hardware configuration shown in FIG. 1 is merely an example, and the dynamic model switching inference device 10 may have other hardware configurations. For example, the dynamic model switching inference device 10 may have multiple processors 105, or multiple memory devices 106.

The dynamic model switching inference device 10 according to this embodiment may be, for example, an embedded device that has limited hardware resources compared to a general computer or computer system.

<Functional configuration of the dynamic model switching inference device 10>
Next, the functional configuration of the dynamic model switching inference device 10 according to the present embodiment will be described with reference to Fig. 2. Fig. 2 is a diagram showing an example of the functional configuration of the dynamic model switching inference device 10 according to the present embodiment.

As shown in FIG. 2, the dynamic model switching inference device 10 according to this embodiment has an inference unit 201, a state variable acquisition unit 202, a recognition state pattern determination unit 203, and a model selection unit 204. Each of these units is realized, for example, by a process in which one or more programs installed in the dynamic model switching inference device 10 are executed by the processor 105.

The dynamic model switching inference device 10 according to this embodiment also has a storage unit 205. The storage unit 205 is realized, for example, by the memory device 106. Note that the storage unit 205 may also be realized, for example, by a database server connected to the dynamic model switching inference device 10 via a communication network.

The inference unit 201 receives image data as input and performs inference processing using a model selected by the model selection unit 204 (hereinafter also referred to as the selected model), recognizing the object type and position in the image represented by the image data, and calculating the recognition rate.

The state variable acquisition unit 202 acquires state variable values that represent the inference results (i.e., recognition rate, object type, position in the image) by the inference unit 201.

The recognition state pattern determination unit 203 refers to the recognition state pattern information stored in the storage unit 205 and determines the current recognition state pattern from the state variable values (and their changes) acquired by the state variable acquisition unit 202. Here, the recognition state pattern is a pattern that represents the situation at the time of inference, and is defined by a combination of, for example, high/low recognition rate, large/small change in object type, and large/small change in position in the image, as described below. Details of the recognition state pattern information will be described later.

The model selection unit 204 refers to the model type information stored in the storage unit 205 and determines the selected model from the recognition state pattern determined by the recognition state pattern determination unit 203. Here, the model type information is information that defines an appropriate model for each recognition state pattern. An appropriate model is, for example, a model that is capable of highly accurate and lightweight image recognition for the relevant recognition state pattern (i.e., a lightweight model specialized for the relevant recognition state pattern). Details of the model type information will be described later.

The storage unit 205 stores various information (e.g., recognition state pattern information, model type information, model data for each model, etc.). In addition to the recognition state pattern information, model type information, and model data, the storage unit 205 also stores various information necessary for executing the dynamic model switching inference process described below.

<Recognition status pattern information>
Next, an example of the recognition state pattern information stored in the storage unit 205 will be described with reference to Fig. 3. Fig. 3 is a diagram showing an example of the recognition state pattern information.

The recognition state pattern information shown in Figure 3 defines five recognition state patterns according to combinations of recognition rate, change in object type, and change in position in the image. Here, a change in object type refers to the change in the value of the object type between the current time and the previous time. Similarly, a change in position in the image refers to the change in the value of the position in the image between the current time and the previous time. Note that in Figure 3, "x" indicates a low recognition rate, a large change in object type, and a large change in position in the image, while "o" indicates a high recognition rate, a small change in object type, and a large change in position in the image. Also, "-" means that it is not defined.

Recognition state pattern 1 is a recognition state pattern that represents a case where the recognition rate is "x". In this case, changes in object type and changes in position in the image are not defined. This is because when the recognition rate is low, there is a high possibility that the object type and position in the image are incorrect.

Recognition state pattern 2 is a recognition state pattern that represents a case where the recognition rate is "o", the change in object type is "x", and the change in position in the image is "x". Recognition state pattern 3 is a recognition state pattern that represents a case where the recognition rate is "o", the change in object type is "x", and the change in position in the image is "o". Recognition state pattern 4 is a recognition state pattern that represents a case where the recognition rate is "o", the change in object type is "o", and the change in position in the image is "x". Recognition state pattern 5 is a recognition state pattern that represents a case where the recognition rate is "o", the change in object type is "o", and the change in position in the image is "o".

In the example shown in FIG. 3, only recognition state pattern 1 is used when the recognition rate is "x", but this recognition state pattern 1 may be further subdivided. That is, it may be subdivided into recognition state pattern 1-1, which is recognition rate "x", change in object type "x", and change in position in image "x", recognition state pattern 1-2, which is recognition rate "x", change in object type "o", and change in position in image "x", recognition state pattern 1-3, which is recognition rate "x", change in object type "o", and change in position in image "x", and recognition state pattern 1-4, which is recognition rate "x", change in object type "o", and change in position in image "o".

In the example shown in FIG. 3, the recognition rate is a combination of high/low and large/small change in object type and position in the image, but this may be further subdivided. That is, for example, the recognition rate may be subdivided into three categories: high/medium/low. Similarly, at least one of the change in object type and the change in position in the image may be subdivided into three categories: large/medium/small. However, subdivision into three categories is just one example, and further subdivision (i.e., four or more categories) may be used.

The example shown in Figure 3 is an example of a recognition state pattern when the three state variables are the recognition rate, object type, and position in the image, and the number of recognition state patterns can increase as the number of state variables increases.

<Model type information>
Next, an example of the model type information stored in the storage unit 205 will be described with reference to Fig. 4. Fig. 4 is a diagram showing an example of the model type information.

The model type information shown in Figure 4 defines appropriate models A to E for each of recognition state patterns 1 to 5.

Recognition state pattern 1 is a recognition state pattern with a low recognition rate. Therefore, for recognition state pattern 1, a general-purpose model A capable of recognizing multiple features (object type and position in the image) with high accuracy is set as an appropriate model. Here, model A includes, for example, a model that automatically extracts features and a model that manually selects or chooses features of interest and combines multiple features. An example of the former model is a CNN model. On the other hand, an example of the latter model is a model that combines multiple features such as SIFT (Scale-Invariant Feature Transform) features and HOG (Histgram Of Gradient) features. Note that, since a general-purpose model A such as a CNN model is used in recognition state pattern 1, the effect of reducing the amount of hardware resources used cannot be expected. This is because, in recognition state pattern 1, improving recognition performance should be prioritized over reducing the amount of hardware resources used.

Recognition state patterns 2 to 4 are recognition state patterns that have high recognition rates but large changes in certain features.

For recognition state pattern 2, there is a large change in both object type and position in the image, so model B, which can easily recognize both of these features, is an appropriate model. Examples of model B include an image recognition model or method using SIFT features, and an image recognition model or method using SURF (Speed-Up Robust Features) features. However, since there is a large change in both object type and position in the image for recognition state pattern 2, a CNN model may be used as model B, or an image recognition model with recognition performance equivalent to that of a CNN model may be used as model B.

For recognition state pattern 3, the change in object type is large, so model C, which can easily recognize object types, is an appropriate model. Examples of model C include an image recognition model or image recognition method using SIFT features, and an image recognition model or image recognition method using HOG features.

It is known that image recognition models or methods using SIFT features are excellent at grasping fine features, while image recognition models or methods using HOG features are excellent at grasping rough features. For this reason, for example, when recognizing the type of object of the same item (for example, when recognizing the type of pattern on a cup, etc.), an image recognition model or method using SIFT features is superior. On the other hand, for example, when recognizing the type of object (for example, when recognizing whether the object is a cup or not), an image recognition model or method using HOG features is superior.

Because recognition state pattern 4 involves large changes in position within the image, model D, which can recognize the position within the image with a light weight, is an appropriate model. Here, a model capable of light weight recognition (hereinafter also referred to as a lightweight model) refers to a model that specializes in recognizing specific features and is capable of image recognition with reduced resources. Examples of model D include an image recognition model or image recognition method using SURF features, and an image recognition model or image recognition method using HOG features.

Recognition state pattern 5 is a recognition state pattern with a high recognition rate and small changes in each feature. In this case, the most lightweight model E is the appropriate model. Model E can be, for example, an image recognition model or image recognition method using SURF features. However, other lightweight image recognition models or image recognition methods may also be used as model E.

Note that it is not necessary for all of models A to E to be different models, and some of the models may be the same model. For example, models D and E may both be image recognition models or image recognition methods that use SURF features.

The specific examples of each model above are merely examples, and various image recognition models or methods can be used in addition to the image recognition models or methods using the above-mentioned SIFT features, SURF features, or HOG features. For example, image recognition methods such as Haar-like features, template matching, etc. may be used.

<Dynamic model switching inference processing>
Next, the dynamic model switching process according to this embodiment will be described with reference to Fig. 5. Fig. 5 is a flowchart showing an example of the flow of the dynamic model switching inference process. Note that, in the following, the dynamic model switching process at a certain time t will be described, where t is an index representing time.

First, the inference unit 201 acquires image data at time t (step S101). Note that the inference unit 201 may acquire image data from another device or equipment (e.g., an imaging device, etc.) connected to the dynamic model switching inference device 10, or may acquire image data from the storage unit 205.

Next, the inference unit 201 performs inference processing using the current selected model with the image data acquired in step S101 as input, and outputs state variable values representing the recognition rate, object type, and position in the image (step S102). These state variable values are stored in the storage unit 205 as the inference results at time t. Note that when t=1 (i.e., when the inference processing is performed for the first time), a predetermined model (for example, a general-purpose model such as a CNN model) may be used as the selected model.

Next, the state variable acquisition unit 202 acquires the state variable values output in step S102 above (i.e., the state variable values at time t) and the state variable values at the previous time (i.e., the state variable values at time t-1) (step S103).

Next, the dynamic model switching inference device 10 executes a model selection process (step S104). In this model selection process, a selected model to be used in the next inference process at time t+1 is determined based on the state variable values at time t, or on both the state variable values at time t and time t-1. The model selection process will be described in detail later.

Next, the inference unit 201 determines whether or not there is next image data (step S105). If there is next image data, the inference unit 201 updates the time t to t+1 (step S106) and returns to the above step S101. As a result, the above steps S101 to S104 are repeatedly executed for each time t. On the other hand, if there is no next image data, the inference unit 201 ends the dynamic model switching inference process.

<Model selection process>
Next, the model selection process in step S104 will be described with reference to Fig. 6. Fig. 6 is a flow chart showing an example of the flow of the model selection process.

First, the recognition state pattern determination unit 203 determines whether the recognition rate included in the state variable value at time t is equal to or greater than a predetermined threshold th ₁ (step S201). If the recognition rate is equal to or greater than the threshold th ₁ , the recognition rate is determined to be high, whereas if it is less than the threshold th _1, the recognition rate is determined to be low. However, for example, thresholds th ₁₁ and th ₁₂ may be prepared, and if the recognition rate is equal to or greater than the threshold th ₁₁ , the recognition rate may be determined to be high, if it is less than the threshold th ₁₂ , the recognition rate may be determined to be low, and if it is equal to or greater than the threshold th ₁₂ and less than the threshold th ₁₁ , the result may be the same as the determination result of this step at time t-1.

The threshold th ₁ (or the thresholds th ₁₁ and th ₁₂ ) may be a preset fixed value, or may be a variable value according to past recognition rates. For example, after calculating the average value μ ₁ and variance σ ₁ of the recognition rates at each time from time t-ΔT (where ΔT is a predetermined time width) to time t, th ₁₁ = μ ₁ + σ ₁ , th ₁₂ = μ ₁ - σ ₁ may be used. However, in order to prevent th ₁₁ from becoming too low, for example, a preset tuning value may be set to R' (e.g., R'=70%), and th ₁₁ =max(μ ₁ + σ ₁ , R').

If the recognition rate is determined to be low in step S201 above, the recognition state pattern determination unit 203 refers to the recognition state pattern information shown in Figure 3 and determines that the current recognition state pattern is "recognition state pattern 1" (step S202).

Then, the model selection unit 204 refers to the model type information shown in FIG. 4 and selects model A corresponding to recognition state pattern 1 as the selected model (step S203).

On the other hand, if the recognition rate is determined to be high in step S201, the recognition state pattern determination unit 203 calculates the change in the state variable values (excluding the recognition rate) at time t relative to time t-1 (step S204). Note that the recognition rate is excluded because the recognition rate has already been determined to be high or low in step S201.

Here, an example of a method for calculating the change in state variable values (excluding the recognition rate) will be described. For example, as shown in the upper diagram of FIG. 7, the state variable values representing the result of inference processing performed by a certain selection model using image data at time t-1 as input are a recognition rate of "80%, object type "0: cup", and position in the image "X=100, Y=150". Also, as shown in the lower diagram of FIG. 7, the state variable values representing the result of inference processing performed by a certain selection model using image data at time t as input are a recognition rate of "85%, object type "0: cup", and position in the image "X=150, Y=100". Note that the recognition rate can take a value between 0 and 100, and the object type can take a category value representing the type of object. Also, the position in the image can take an XY coordinate value with the top left corner of the image represented by the image data as the origin, with the right direction being the positive direction of the X axis and the downward direction being the positive direction of the Y axis.

At this time, the recognition state pattern determination unit 203, for example, sets the change to "0" if the object type at time t is the same as the object type at time t-1, and sets the change to "1" if they are different. The recognition state pattern determination unit 203 also calculates the change in the position in the image, for example, by subtracting the X and Y coordinate values of the position in the image at time t-1 from the X and Y coordinate values of the position in the image at time t, respectively. As a result, the change in object type and the change in position in the image are as shown in FIG. 8. That is, the change in object type at time t relative to time t-1 is "0", and the change in position in the image at time t relative to time t-1 is "X=50, Y=0".

Next, the recognition state pattern determination unit 203 performs a threshold determination on the changes in state variable values other than the recognition rate to determine whether the changes are large or small (step S205). That is, the recognition state pattern determination unit 203 determines whether the change in object type is large or small, and whether the change in position in the image is large or small. Details of determining whether the change in object type and the change in position in the image are large or small are described below.

When determining whether the change in object type is large/small, for example, the recognition state pattern determination unit 203 calculates the change rate Rv of the object type, and determines that the change in the object type is large if this change rate Rv is equal to or greater than a predetermined threshold th ₂ , and determines that the change in the object type is small if it is less than the threshold th _2. Here, the change rate Rv of the object type is calculated, for example, as Rv = (n/N) x 100, where n is the number of changes in the object type between time t-ΔT and time t, and N is the number of time indexes included in the time span ΔT.

However, for example, thresholds _th21 and _th22 may be prepared, and if the rate of change Rv of the object type is equal to or greater than the threshold _th21 , it may be determined that the change in the object type is large, and if it is less than the threshold _th22 , it may be determined that the change in the object type is small, and if it is less than the threshold _th21 and equal to or greater than the threshold _th22 , it may be determined that the change in the object type is the same as the determination result of this step at time t-1 (or, if this step has not been executed at time t-1, the most recent determination result of this step at a time prior to that).

The threshold th ₂ (or the thresholds th ₂₁ and th ₂₂ ) may be a preset fixed value, or may be a variable value according to the past rate of change Rv. For example, after calculating the average value μ ₂ and variance σ ₂ of the rate of change Rv of the object type at each time from time t-ΔT to time t, th ₂₁ = μ ₂ + σ ₂ and th ₂₂ = μ ₂ - σ ₂ may be set. However, in order to prevent th ₂₁ from becoming too low, for example, a predetermined tuning value may be set to Rv' (for example, Rv' = 30%), and th ₂₁ = max (μ ₂ + σ ₂ , Rv').

- When determining whether the change in the position in the image is large/small For example, the recognition state pattern determination unit 203 calculates the change rate Rd of the position in the image, and determines that the change in the position in the image is large if the change rate Rd is equal to or greater than a predetermined threshold th ₃ , and determines that the change in the position in the image is small if the change rate Rd is less than the threshold th _3. Here, the change rate Rd of the position in the image is, for example, the higher of the change rate Rdx in the X-axis direction from time t-ΔT to time t and the change rate Rdy in the Y-axis direction from time t-ΔT to time t (that is, Rd=max(Rdx, Rdy)). In addition, Rdx is calculated, for example, by Rdx=(dx/Px)×100, where dx is the amount of movement in the X-axis direction from time t-ΔT to time t and Px is the total number of pixels in the X-axis direction of the image represented by the image data. Similarly, Rdy is calculated, for example, as Pdy = (dy/Py) x 100, where dy is the amount of movement in the Y-axis direction between time t-ΔT and time t, and Py is the total number of pixels in the Y-axis direction of the image represented by the image data.

However, for example, thresholds _th31 and _th32 may be prepared, and if the rate of change Rd of position in the image is equal to or greater than the threshold _th31 , it may be determined that the change in position in the image is large, and if it is less than the threshold _th32 , it may be determined that the change in position in the image is small, and if it is less than the threshold _th31 and equal to or greater than the threshold _th32 , it may be determined that the result is the same as the determination result of this step at time t-1 (or, if this step has not been executed at time t-1, the most recent determination result of this step at a time prior to that).

The threshold th ₃ (or the thresholds th ₃₁ and th ₃₂ ) may be a preset fixed value, or may be a variable value according to the past rate of change Rd. For example, after calculating the average value μ ₃ and variance σ ₃ of the rate of change Rd of the position in the image at each time from time t-ΔT to time t, th ₃₁ = μ ₃ + σ ₃ , th ₃₂ = μ ₃ - σ ₃ may be set. However, in order to prevent th ₃₁ from becoming too low, for example, a predetermined tuning value Rd' (for example, Rd' = 200%) may be set, and th ₃₁ = max (μ ₃ + σ ₃ , Rd').

Next, the recognition state pattern determination unit 203 refers to the recognition state pattern information shown in FIG. 3 and determines whether the current recognition state pattern is recognition state pattern 1 to recognition state pattern 5 (step S206). That is, the recognition state pattern determination unit 203 determines "recognition state pattern 2" if the change in object type is large and the change in position in the image is also large, "recognition state pattern 3" if the change in object type is large and the change in position in the image is small, "recognition state pattern 4" if the change in object type is small and the change in position in the image is large, and "recognition state pattern 5" if the change in object type is small and the change in position in the image is also small.

Next, the model selection unit 204 refers to the model type information shown in FIG. 4 and selects the model corresponding to the recognition state pattern determined in step S206 above as the selected model (step S207). That is, the model selection unit 204 selects model B as the selected model if the recognition state pattern is determined to be "recognition state pattern 2" in step S206 above, model C if the recognition state pattern is determined to be "recognition state pattern 3", model D if the recognition state pattern is determined to be "recognition state pattern 4", and model E if the recognition state pattern is determined to be "recognition state pattern 5".

Then, following step S203 or step S207, the model selection unit 204 reads the model data of the selected model from the storage unit 205 (step S208). As a result, the selected model is used in the inference process at the next time t+1.

<Summary>
As described above, the dynamic model switching inference device 10 according to the present embodiment can dynamically switch the selected model to a model corresponding to the pattern according to the result of inference processing performed by the current selected model for an image recognition task and the recognition state pattern representing the pattern of change. As a result, for example, when the recognition rate is low, a general-purpose model with a higher recognition rate can be selected as the selected model, and when a certain feature changes significantly, a lightweight model specialized for recognizing that feature can be selected as the selected model. Therefore, in particular, when a certain recognition rate or higher is obtained, inference processing can be performed by a lightweight model specialized for recognizing a specific feature, and it is possible to reduce the amount of hardware resources used compared to when inference processing is always performed by a general-purpose model.

Note that while this embodiment focuses on image recognition tasks, this is merely one example, and the invention can be applied to a variety of tasks in which data is input and inference processing is repeatedly performed using a model. For example, the invention can be applied to a variety of tasks, such as data classification, regression, prediction, and anomaly detection.

The present invention is not limited to the specifically disclosed embodiments above, and various modifications, changes, and combinations with known technologies are possible without departing from the scope of the claims.

10 Dynamic model switching inference device 101 Input device 102 Display device 103 External I/F
103a Recording medium 104 Communication I/F
105 Processor 106 Memory device 107 Bus 201 Inference unit 202 State variable acquisition unit 203 Recognition state pattern determination unit 204 Model selection unit 205 Storage unit

Claims (7)

An inference device that repeatedly executes inference of a predetermined task for each of a plurality of data,
an inference unit that executes inference of the task on the data using a selection model that indicates a model selected from a plurality of models;
a selection unit that selects a selection model for performing inference in a next iteration from among the plurality of models based on a result of the inference;
having
the result of the inference includes a recognition result for each of one or more features of the data, and a recognition rate representing the likelihood of the recognition result;
The selection unit is
An inference device that, depending on a pattern of the recognition result and the recognition rate included in a result of inference in a current iteration, selects a model corresponding to the pattern as a selected model for performing inference in the next iteration . The selection unit is
The inference device according to claim 1 , wherein if the recognition rate is less than a predetermined threshold, a general-purpose model capable of recognizing the one or more features is selected as the selected model as a model corresponding to the pattern . The selection unit is
The inference device according to claim 1 , wherein if the recognition rate is equal to or greater than the threshold, a lightweight model capable of recognizing a specific feature is selected as the selected model as a model corresponding to the pattern . An inference device as claimed in any one of claims 1 to 3, wherein the pattern is represented by a combination of whether or not the recognition rate is greater than or equal to the threshold value and the presence or absence of a change between the recognition result included in the inference result in the current iteration and the recognition result included in the inference result in the previous iteration, or the magnitude of the change . the task is an image recognition task,
the one or more features include at least one of a type of object in an image represented by the data and a position of the object in the image ;
5. The inference device of claim 4, wherein the pattern is represented by a combination of whether or not the recognition rate is equal to or greater than the threshold, whether or not there is a change between the type of object included in the inference result in the current iteration and the type of object included in the inference result in the previous iteration, and the magnitude of change between the position of the object included in the inference result in the current iteration and the position of the object included in the inference result in the previous iteration . An inference method for repeatedly executing inference of a predetermined task for each of a plurality of data, comprising:
an inference procedure for performing inference of the task on the data using a selection model indicating a model selected from a plurality of models;
a selection step of selecting, based on a result of the inference, from among the plurality of models, a selection model for performing inference in a next iteration;
The computer executes
the result of the inference includes a recognition result for each of one or more features of the data, and a recognition rate representing the likelihood of the recognition result;
The selection step comprises:
An inference method comprising: selecting, depending on a pattern of the recognition result and the recognition rate contained in a result of inference in a current iteration, a model corresponding to the pattern as a selected model for performing inference in the next iteration . A program for repeatedly executing inference of a predetermined task for each of a plurality of data,
an inference procedure for performing inference of the task on the data using a selection model indicating a model selected from a plurality of models;
a selection step of selecting, based on a result of the inference, from among the plurality of models, a selection model for performing inference in a next iteration;
on the computer ,
the result of the inference includes a recognition result for each of one or more features of the data, and a recognition rate representing the likelihood of the recognition result;
The selection step comprises:
a program for selecting, depending on a pattern of the recognition result and the recognition rate included in a result of inference in a current iteration, a model corresponding to the pattern as a selection model for performing inference in the next iteration .

Images