Machine learning program, machine learning method, and machine learning device
The machine learning program addresses accuracy deterioration by quickly updating models through data distribution analysis and selection, ensuring timely replacement of operational models to maintain performance.
Patent Information
- Application Number
- JP2021000550
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2021-01-05
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2041-01-05
AI Technical Summary
Machine learning models experience accuracy deterioration due to concept drift, and updating operational models with new data is hindered by long update times, leading to continued use of models with degraded accuracy.
A machine learning program that updates a second model by identifying the distribution of feature amounts in input data, selects a suitable model for learning, and generates a new model based on labeled training data to quickly replace the operational model.
This approach reduces the update time of machine learning models by efficiently selecting and updating models based on data distribution, maintaining accuracy and reducing the duration of using outdated models.
Smart Images

Figure 0007593115000001 
Figure 0007593115000002 
Figure 0007593115000003
Abstract
Description
[Technical field]
[0001] The present invention relates to a machine learning program, a machine learning method, and a machine learning device. [Background technology]
[0002] When performing classification, prediction, etc. based on a machine learning model created using training data, a deterioration in accuracy may occur after operation begins due to concept drift. Here, concept drift refers to a change in data distribution between the training data used to train the machine learning model and the data input after operation begins.
[0003] Fig. 12 is a diagram for explaining concept drift. As shown in Fig. 12, concept drift includes sudden drift, gradual drift, and incremental drift. Sudden drift is a drift in which data distribution changes suddenly. Gradual drift is a drift in which data distribution changes stepwise. Gradual drift is a drift in which data distribution changes gradually. Among concept drifts, gradual drift is particularly difficult to detect because data changes little by little.
[0004] In order to prevent the accuracy of the machine learning model from deteriorating due to changes in the distribution of data over time, the machine learning model is periodically retrained. However, updating the operational model itself during operation in the system by retraining is not realistic in consideration of stable operation of the system. Therefore, an updated model is prepared in parallel with the operational model. An initial updated model is created by copying the operational model. The updated model is updated to accommodate concept drift. The updated model may be updated periodically, for example, and multiple updated models may be generated in a time series. Then, the operational model is appropriately replaced with the updated model to suppress deterioration in accuracy.
[0005] In addition, as a conventional technique, there is a technique for retraining a machine learning model for an automation system with post-drift process data when there is concept drift, for the model being trained with process data including the context of the automation process.
[0006] In addition, there is a conventional technique for preventing a decrease in the accuracy of a prediction process result using a learning model caused by a timing mismatch between the input data and the learning model when the learning model is updated due to a change in the nature of the input data in stream processing. Here, stream processing is a process that returns a processing result in a timely manner for real-time input data.
[0007] In addition, a conventional technique involves calculating a data distribution of a training data set, calculating the likelihood that new data will follow the calculated data distribution, and retraining a machine learning model if the number of times the calculated likelihood is smaller than a first threshold exceeds a second threshold. [Prior art documents] [Patent documents]
[0008] [Patent Document 1] Special Publication No. 2018-513490 [Patent Document 2] JP 2017-068710 A [Patent Document 3] US Patent Application Publication No. 2018 / 0285772 Summary of the Invention [Problem to be solved by the invention]
[0009] An updated model in which model parameters are updated based on data in operation is prepared for the operation model, and when the accuracy of the operation model deteriorates due to concept drift, the accuracy deterioration can be suppressed by switching to the updated model. In this case, in order to follow the concept drift, it is better to perform the update based on newer data. Therefore, the updated model is updated based on the latest data, and the operation model is replaced with the updated updated model. However, if the update process of the updated model takes a long time, there is a problem that the operation model with deteriorated accuracy continues to be used.
[0010] Fig. 13 is a diagram for explaining the problem of continuing to use an operational model whose accuracy has deteriorated due to the time required for the update process of the updated model. Fig. 13 shows a case where an image is judged in real time. As shown in Fig. 13, real-time image judgment is performed on image data from an input sensor based on the operational model.
[0011] When the accuracy of the operational model decreases, the update model is started, and the update model is updated during operation. Then, after several updates of the update model are performed, when it is decided to update the operational model, the update model is updated last, and the operational model is updated with the last updated update model. However, if it takes a long time to update the update model, the operational model before the update continues to be used until the update of the last update model is completed and the operational model is updated.
[0012] An object of one aspect of the present invention is to reduce the update time of an update model. [Means for solving the problem]
[0013] In one embodiment, the machine learning program causes a computer to update parameters of a first machine learning model to a second machine learning model among one or more machine learning models generated by updating parameters of the first machine learning model. No. 1 data Enter The machine learning program causes the computer to execute the following process: The aboveThe second machine learning model calculated the No. 1 The machine learning program causes the computer to execute a process of identifying a first distribution of features of data. The machine learning program then causes the computer to identify the first distribution and training data used to update parameters when each of the one or more machine learning models is generated. The training data calculated by each of the one or more machine learning models The machine learning program causes the computer to execute a process of comparing a distribution of feature quantities of the training data with a distribution of feature quantities of the training data. The machine learning program then causes the computer to execute a process of selecting a specific machine learning model from the one or more machine learning models based on a result of the comparison. The machine learning program then causes the computer to execute a process of generating a new machine learning model by updating parameters of the specific machine learning model based on specific training data labeled based on the feature quantities of the data. Effect of the Invention
[0014] In one aspect, the present invention can reduce the update time of an update model. [Brief description of the drawings]
[0015] [Figure 1] FIG. 1 is a diagram illustrating a functional configuration of the machine learning device according to the first embodiment. [Diagram 2] FIG. 2 is a diagram illustrating a configuration of a classification system according to a second embodiment. [Diagram 3] FIG. 3 is a diagram for explaining the HDL technology. [Figure 4] FIG. 4 is a diagram for explaining density-based clustering. [Figure 5A] FIG. 5A is a first diagram for explaining a plurality of models managed by the model management unit. [Figure 5B] FIG. 5B is a second diagram for explaining a plurality of models managed by the model management unit. [Figure 6] FIG. 6 is a diagram illustrating a functional configuration of the machine learning device according to the second embodiment. [Figure 7] FIG. 7 is a diagram showing cluster information. [Figure 8] FIG. 8 is a diagram illustrating an example of the past model storage unit. [Figure 9] FIG. 9 is a flowchart showing a flow of processing by the deterioration detection unit. [Figure 10] FIG. 10 is a diagram showing a sequence of the update process of the drift adaptive model. [Figure 11] FIG. 11 is a diagram illustrating a hardware configuration of a computer that executes the machine learning program according to the first and second embodiments. [Figure 12] FIG. 12 is a diagram for explaining the concept drift. [Figure 13] FIG. 13 is a diagram for explaining the problem of continuing to use an operational model whose accuracy has deteriorated due to the time required for updating the HDL model. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0016] Hereinafter, examples of the machine learning program, the machine learning method, and the machine learning device disclosed in the present application will be described in detail with reference to the drawings. Note that the disclosed technology is not limited to the examples. EXAMPLES
[0017] FIG. 1 is a diagram illustrating a functional configuration of a machine learning device 5a according to a first embodiment. As illustrated in FIG. 1, the machine learning device 5a includes a deterioration detection unit 10a and a model management unit 20a. The deterioration detection unit 10a detects deterioration of an operational model, and updates the operational model when it detects deterioration of the operational model. Here, the operational model is a machine learning model in operation. The deterioration detection unit 10a copies the operational model and stores it as a drift-responsive model. Then, the deterioration detection unit 10a updates the drift-responsive model based on input data. Then, when the operational model deteriorates, the deterioration detection unit 10a copies the drift-responsive model to the operational model.
[0018] The model management unit 20a manages the drift response model in association with the distribution of the feature amount of the input data (hereinafter, referred to as "input data distribution"). The model management unit 20a manages all models used as drift response models by the deterioration detection unit 10a.
[0019] The deterioration detection unit 10a includes a model storage unit 11a, a classification unit 12a, a distribution information generation unit 13a, a judgment unit 14a, a labeling unit 15a, a model access unit 16a, and a learning unit 17a. The model management unit 20a includes a past model storage unit 21a, a registration unit 22a, a comparison unit 23a, and a selection unit 24a.
[0020] The model storage unit 11a stores a drift-adaptive model. Specifically, the model storage unit 11a stores parameters of a neural network as the drift-adaptive model. The parameters include weights between neurons. The weights between neurons are updated by learning.
[0021] The classification unit 12 a classifies the input data based on the drift response model stored in the model storage unit 11 .
[0022] The distribution information generating unit 13a generates an input data distribution for a predetermined number (eg, 300) of input data.
[0023] The determination unit 14a determines whether the drift response model needs to be updated, and when it is determined that the drift response model needs to be updated, instructs the learning unit 17a to learn. The determination unit 14a determines whether the drift response model needs to be updated, for example, based on the rate at which the classification result by the classification unit 12a matches the correct answer specified by the user.
[0024] When the determination unit 14a determines that the drift adaptive model needs to be updated, the labeling unit 15a adds a label (correct answer) to the input data to create training data. Then, the labeling unit 15a passes the created training data to the learning unit 17a.
[0025] The model access unit 16a accesses the model management unit 20a. The model access unit 16a designates the input data distribution generated by the distribution information generation unit 13a, inquires of the model management unit 20a about a model having a similar input data distribution, and passes the model information acquired as a response from the model management unit 20a to the learning unit 17a.
[0026] Furthermore, the model access unit 16a requests the model management unit 20a to register the input data distribution generated by the distribution information generation unit 13a and the information on the drift adaptive model generated by the learning unit 17a in association with each other.
[0027] The learning unit 17a performs learning using the model information received from the model access unit 16a and the training data received from the labeling unit 15 to generate a new drift-compatible model. The generated drift-compatible model is registered in the model management unit 20a by the model access unit 16a. The learning unit 17a stores the generated drift-compatible model in the model storage unit 11a. That is, the learning unit 17a updates parameters of the drift-compatible model stored in the model storage unit 11a. Note that the learning unit 17a may perform fine tuning on the drift-compatible model when updating parameters of the drift-compatible model stored in the model storage unit 11a.
[0028] The past model storage unit 21a stores information about the drift response model in association with the input data distribution.
[0029] The registration unit 22a receives information on the drift adaptive model and the input data distribution from the model access unit 16a, and registers them in the past model storage unit 21a.
[0030] The comparison unit 23a receives the input data distribution from the model access unit 16a, and compares it with the input data distribution stored in the past model storage unit 21a to calculate the similarity.
[0031] The selection unit 24a selects, from the past model storage unit 21a, a drift-adaptive model that is most similar to the input data distribution received from the model access unit 16a, based on the similarity calculated for each drift-adaptive model by the comparison unit 23a. Then, the selection unit 24a passes information on the selected drift model to the model access unit 16a.
[0032] The classification unit 12a, the distribution information generation unit 13a, the judgment unit 14a, the labeling unit 15a, the model access unit 16a, the learning unit 17a, the registration unit 22a, the comparison unit 23a, and the selection unit 24a correspond to a control unit in the claims.
[0033] As described above, in the first embodiment, the distribution information generating unit 13a generates an input data distribution. Then, the comparing unit 23a compares the input data distribution generated by the distribution information generating unit 13a with the input data distribution stored in the past model storing unit 21, and the selecting unit 24a selects the most similar model based on the comparison by the comparing unit 23a. Also, the labeling unit 15a labels the input data to create training data. Then, the learning unit 17a performs learning using the model selected by the selecting unit 24a and the training data created by the labeling unit 15a to generate a drift-compliant model. Therefore, the learning unit 17a can perform learning at high speed, and the update time of the drift-compliant model can be shortened. EXAMPLES
[0034] First, a classification system according to Example 2 will be described. Fig. 2 is a diagram showing a configuration of the classification system according to Example 2. As shown in Fig. 2, the classification system 1 according to Example 2 includes an input sensor 2, a data storage device 3, a classification device 4, a machine learning device 5, and a display device 6.
[0035] The input sensor 2 is a sensor that acquires data to be classified. For example, in the case of classifying images, the input sensor 2 is a camera. The data storage device 3 stores the input data acquired by the input sensor 2. The data storage device 3 stores, for example, image data.
[0036] The classification device 4 is a device that uses an operational model to classify input data stored in the data storage device 3. For example, the classification device 4 classifies images captured by a camera in real time.
[0037] The machine learning device 5 is a device that generates an operation model through machine learning using the input data stored in the data storage device 3. The machine learning device 5 includes a deterioration detection unit 10 and a model management unit 20.
[0038] The degradation detection unit 10 detects degradation of the operation model based on HDL (High Durability Learning) technology, and updates the operation model when the degradation of the operation model is detected. HDL is a technology for estimating the accuracy degradation of a machine learning model. FIG. 3 is a diagram for explaining HDL technology. As shown in FIG. 3, in HDL technology, an HDL model is used in addition to the operation model. The HDL model is a machine learning model used in HDL technology. Also, in FIG. 3, the machine learning model classifies input data into one of two values. For example, the machine learning model inputs an image of a person, determines whether or not the person is wearing a uniform, and outputs whether or not the person is wearing a uniform.
[0039] First, an HDL model is created by copying the operational model. Then, as shown in Figure 3, the HDL model inputs the data (Data #1, Data #2, ...) that the operational model inputs, and, like the operational model, determines which of the two values the input data is. The determination is made based on the output value of the neurons in the output layer. Point 9 is plotted in the DT space (Durable Topology Space) based on the output value of the neurons in the output layer and the determination result.
[0040] Each axis of the DT space corresponds to the output value of each neuron in the output layer. The DT space is the feature space of the input data. In Figure 3, since there are three neurons in the output layer, the DT space is a three-dimensional space, but for convenience of explanation, the DT space is shown in two dimensions. The judgment result is represented by the type of point 9 (circles and ●). If the HDL model is not degraded, the collection of circles and the collection of ● are separated in the DT space. Note that the input data may be judged to be one of three or more values.
[0041] In HDL technology, points 9 are clustered based on density, and cluster A including circles and cluster B including circles are created. FIG. 4 is a diagram for explaining density-based clustering. The horizontal axis of FIG. 4 indicates feature amount, and the vertical axis indicates density of points 9 (e.g., the number of points 9 per unit interval of feature amount). FIG. 4 shows the distribution of feature amounts. For convenience of explanation, the dimension of the DT space in FIG. 4 is set to 1. As shown in FIG. 4, clusters A and B are identified based on the density of feature amounts. Each cluster is identified, for example, by a center point.
[0042] Point 9, which does not belong to either cluster A or cluster B, arises due to degradation of the HDL model. In HDL technology, re-learning is performed using the input data corresponding to point 9 belonging to cluster A and the input data corresponding to point 9 belonging to cluster B as training data, and the HDL model is updated. The input data corresponding to point 9 belonging to cluster A is labeled with a circle, and the input data corresponding to point 9 belonging to cluster B is labeled with a black circle.
[0043] In this way, in HDL technology, the occurrence of drift is considered to be a movement of the density distribution in the DT space, and the HDL model is made to follow the drift by always tracking the center of the density. Then, when deterioration of the operational model is detected, the operational model is replaced with the HDL model. Therefore, HDL technology can prevent the deterioration of the accuracy of the operational model due to concept drift. Also, in HDL technology, the labeling of the training data used for re-learning is performed automatically. Therefore, HDL technology can prevent an increase in operational costs.
[0044] Returning to Fig. 2, the degradation detection unit 10 copies the operational model and stores it as a drift-compliant model (HDL model) (t1). Then, the degradation detection unit 10 updates the drift-compliant model based on the input data (t2). Then, when the operational model deteriorates, the degradation detection unit 10 copies the drift-compliant model to the operational model (t3).
[0045] The model management unit 20 manages the drift-compatible model in association with information of the DT space. The model management unit 20 manages all models used by the deterioration detection unit 10 as drift-compatible models. FIGS. 5A and 5B are diagrams for explaining a plurality of models managed by the model management unit 20. As shown in FIG. 5A, the deterioration detection unit 10 performs classification based on the drift-compatible model before updating, and clusters the data based on the classification result (t11). Then, the deterioration detection unit 10 extracts data near the center of the cluster as learning data (t12).
[0046] Here, the degradation detection unit 10 excludes data that falls outside the cluster in the analysis in the DT space, resulting in a shortage of learning data. Therefore, the degradation detection unit 10 uses the drift-compliant model before the update when updating the drift-compliant model to cover the small amount of data.
[0047] That is, the deterioration detection unit 10 creates drift compatible model #1 using the initial model (copy of the operational model) and input data group #1, and creates drift compatible model #2 using the drift compatible model #1 and input data group #2. Similarly, the deterioration detection unit 10 creates drift compatible model #3, drift compatible model #4, .... Then, the model management unit 20 manages the drift compatible model #1, drift compatible model #2, ..., in association with the distribution of the feature amounts of the input data group #1, the distribution of the feature amounts of the input data group #2, ..., respectively.
[0048] However, the immediately preceding drift-compatible model is not necessarily optimal. FIG. 5B is a diagram showing an example in which the immediately preceding drift-compatible model is not optimal. In FIG. 5B, the input data group drifts from A to B to C to D to E. Therefore, the most recent input data group E is close to D in terms of drift, but the model based on E has the highest similarity to the model based on A.
[0049] Therefore, the model management unit 20 manages information of the DT space (for example, the central coordinates of a cluster) in association with a model, and when updating the drift-compatible model, the deterioration detection unit 10 queries the model management unit 20 based on the information of the DT space to acquire an optimal past model. Then, the deterioration detection unit 10 updates the drift-compatible model based on the acquired model. Note that if there is no appropriate model among the past models, the deterioration detection unit 10 acquires the most recent model from the model management unit 20.
[0050] Returning to FIG. 2, the display device 6 displays the classification results by the classification device 4, messages regarding updates of the drift response model and the operational model by the deterioration detection unit 10, information on the models managed by the model management unit 20, and the like.
[0051] Next, the functional configuration of the machine learning device 5 will be described. Fig. 6 is a diagram showing the functional configuration of the machine learning device 5. As shown in Fig. 6, the deterioration detection unit 10 has a model storage unit 11, a classification unit 12, a cluster generation unit 13, a judgment unit 14, a labeling unit 15, a model access unit 16, and a learning unit 17. The model management unit 20 has a past model storage unit 21, a registration unit 22, a comparison unit 23, and a selection unit 24.
[0052] The model storage unit 11 stores the drift-compatible model. Specifically, the model storage unit 11 stores parameters of a neural network as the drift-compatible model. The parameters include weights between neurons. The weights between neurons are updated by learning.
[0053] The classification unit 12 classifies the input data based on the drift-compatible model stored in the model storage unit 11. The classification unit 12 notifies the cluster generation unit 13 of the classification result and the coordinates in the DT space.
[0054] The cluster generation unit 13 performs clustering of the points 9 in the DT space based on the classification results by the classification unit 12 for a predetermined number of input data (e.g., 300) and the coordinates in the DT space, to generate clusters. The cluster generation unit 13 identifies the distribution of the points 9 in the DT space. The cluster generation unit 13 performs clustering based on the density of the points 9 in the DT space. The cluster generation unit 13 generates the same number of clusters as the number of classifications by the classification unit 12. The cluster generation unit 13 identifies the information of the clusters.
[0055] Fig. 7 is a diagram showing cluster information. As shown in Fig. 7, the cluster information includes the center coordinates of the cluster, the density (dispersion) of the cluster, and the amount of data (the number of input data). The cluster information indicates the distribution of the feature amounts of the input data.
[0056] The determination unit 14 determines whether or not the drift response model needs to be updated based on the information of the clusters generated by the cluster generation unit 13, and when it determines that the update is necessary, it instructs learning to the learning unit 17. The determination unit 14 determines whether or not the drift response model needs to be updated based on, for example, the proportion of input data that does not belong to any cluster.
[0057] When the determination unit 14 determines that the drift adaptive model needs to be updated, the labeling unit 15 creates training data by adding a classification result corresponding to a cluster as a label (correct answer) to input data corresponding to a point 9 belonging to any cluster. Then, the labeling unit 15 passes the created training data to the learning unit 17.
[0058] The model access unit 16 accesses the model management unit 20. The model access unit 16 designates information about a cluster generated by the cluster generation unit 13, queries the model management unit 20 about a model with similar cluster information, and passes the model information acquired as a response from the model management unit 20 to the learning unit 17.
[0059] Furthermore, the model access unit 16 requests the model management unit 20 to register the information of the clusters generated by the cluster generation unit 13 and the information of the drift-compliant model generated by the learning unit 17 in association with each other.
[0060] The learning unit 17 performs learning using the model information received from the model access unit 16 and the training data received from the labeling unit 15, and generates a new drift-enabled model. The generated drift-enabled model is registered in the model management unit 20 by the model access unit 16. The learning unit 17 stores the generated drift-enabled model in the model storage unit 11. That is, the learning unit 17 updates parameters of the drift-enabled model stored in the model storage unit 11. Note that, when updating parameters of the drift-enabled model stored in the model storage unit 11, the learning unit 17 may perform fine tuning on the drift-enabled model.
[0061] The past model storage unit 21 stores information about the drift-compatible model in association with information about the cluster. FIG. 8 is a diagram showing an example of the past model storage unit 21. As shown in FIG. 8, the past model storage unit 21 stores HDL information and a model identifier in association with each drift-compatible model. The HDL information includes the amount of data, density information, and cluster center coordinates. The model identifier is an identifier for identifying the drift-compatible model. The model management unit 20 separately stores information such as parameters included in the drift-compatible model in association with the identifier.
[0062] The amount of data is the number of input data. The density information is the density (distribution) of the cluster. The cluster center coordinates are the coordinates of the center of the cluster. For example, in the drift-enabled model identified by "model20200618.model", the number of input data is "300". In addition, in the drift-enabled model identified by "model20200618.model", the densities of the two clusters are "0.823" and "0.74", and the coordinates of the centers of the two clusters are (0.3, 0.5) and (0.7, 0.2).
[0063] The registration unit 22 receives the drift-compliant model information and cluster information from the model access unit 16 and registers them in the past model storage unit 21.
[0064] The comparison unit 23 receives cluster information from the model access unit 16, and calculates the similarity by comparing it with the information of each cluster stored in the past model storage unit 21. Specifically, the comparison unit 23 normalizes the HDL information on a column-by-column basis and calculates the similarity.
[0065] The selection unit 24 selects, from the past model storage unit 21, a drift-compatible model whose HDL information is most similar to the cluster information received from the model access unit 16, based on the similarity calculated for each drift-compatible model by the comparison unit 23. Then, the selection unit 24 passes information on the selected drift model to the model access unit 16.
[0066] The classification unit 12, cluster generation unit 13, judgment unit 14, labeling unit 15, model access unit 16, learning unit 17, registration unit 22, comparison unit 23 and selection unit 24 correspond to a control unit in the claims.
[0067] Next, the flow of processing by the deterioration detection unit 10 will be described. FIG. 9 is a flowchart showing the flow of processing by the deterioration detection unit 10. As shown in FIG. 9, the deterioration detection unit 10 uses input data input by the operation model to perform classification based on the drift-compatible model (step S1). Then, the deterioration detection unit 10 performs clustering based on the classification result and the feature amount of the input data, and determines whether the accuracy of the deterioration detection has deteriorated based on the cluster information (step S2). Here, deterioration of the accuracy of the deterioration detection means that the accuracy of the drift-compatible model has deteriorated. Then, if the accuracy of the deterioration detection has not deteriorated, the deterioration detection unit 10 moves to step S6.
[0068] On the other hand, when the accuracy of deterioration detection has deteriorated, the deterioration detection unit 10 acquires a learning model from the model management unit 20 (step S3). Here, the deterioration detection unit 10 acquires a past drift-compatible model having similar features of the input data as a learning model based on the cluster information. Then, the deterioration detection unit 10 performs learning using the acquired drift-compatible model and input data belonging to one of the clusters to create a drift-compatible model (step S4), and updates the drift-compatible model (step S5).
[0069] Then, the degradation detection unit 10 determines whether the accuracy of the operational model has deteriorated (step S6), and if it has not deteriorated, returns to step S1, and if it has deteriorated, updates the operational model using the drift-compatible model (step S7), and returns to step S1.
[0070] In this way, the deterioration detection unit 10 performs learning using a past drift response model that has similar characteristics to the input data, and therefore can perform learning at high speed.
[0071] Fig. 10 is a diagram showing the sequence of the update process of the drift-compliant model. As shown in Fig. 10, the deterioration detection unit 10 performs classification based on the drift-compliant model using the input data input by the operation model (step S11). Then, the deterioration detection unit 10 generates a cluster based on the classification result and the coordinates of the DT space, and identifies the cluster information (step S12). That is, the deterioration detection unit 10 identifies the distribution of the feature amount of the input data calculated by the drift-compliant model.
[0072] Then, the degradation detection unit 10 detects degradation of the drift-compliant model (step S13), and queries the model management unit 20 about a similar model by specifying cluster information (step S14). Then, the degradation detection unit 10 generates training data using input data belonging to any of the clusters (step S15).
[0073] Upon receiving a query from the deterioration detection unit 10, the model management unit 20 compares the information of the cluster specified by the deterioration detection unit 10 with the HDL information of each model stored in the past model storage unit 21 (step S16). That is, the model management unit 20 compares the distribution of the feature amounts of the input data with the distribution of the feature amounts of the training data used in generating each model stored in the past model storage unit 21. Then, the model management unit 20 selects a model based on the comparison result (step S17). Then, the model management unit 20 notifies the deterioration detection unit 10 of the information of the selected model (step S18).
[0074] The deterioration detection unit 10 generates a new drift-compliant model by updating the parameters of the notified model based on the training data (step S19), and updates the drift-compliant model in the model storage unit 11 with the generated drift-compliant model (step S20). Then, the deterioration detection unit 10 requests the model management unit 20 to register the drift-compliant model (step S21). Then, the model management unit 20 registers the requested drift-compliant model (step S22).
[0075] In this way, the model management unit 20 selects a model based on the cluster information and HDL information specified by the deterioration detection unit 10, so that the deterioration detection unit 10 can quickly perform learning using the model selected by the model management unit 20.
[0076] As described above, in the second embodiment, the classification unit 12 classifies input data based on the drift-compliant model stored in the model storage unit 11, and the cluster generation unit 13 generates clusters based on the classification result by the classification unit 12 and the coordinates of the DT space. Then, the comparison unit 23 compares the information of the clusters generated by the classification unit 12 with the HDL information stored in the past model storage unit 21, and the selection unit 24 selects the most similar model based on the comparison by the comparison unit 23. Also, the labeling unit 15 creates training data by labeling the input data classified into any of the clusters generated by the classification unit 12. Then, the learning unit 17 performs learning using the model selected by the selection unit 24 and the training data created by the labeling unit 15 to generate a drift-compliant model. Therefore, the learning unit 17 can perform learning at high speed, and the update time of the drift-compliant model can be shortened.
[0077] Furthermore, in the second embodiment, the model access unit 16 requests the model management unit 20 to register the drift-compliant model generated by the learning unit 17, and the registration unit 22 registers the drift-compliant model requested to be registered in the past model storage unit 21. Therefore, the machine learning device 5 can accumulate drift-compliant models and can perform learning using highly similar drift-compliant models.
[0078] Furthermore, in the second embodiment, the machine learning device 5 uses the center coordinates and density of the cluster as the cluster information, and therefore can accurately select a similar model from the past model storage unit 21.
[0079] In the first and second embodiments, the machine learning devices 5a and 5 are described, but a machine learning program having similar functions can be obtained by implementing the configurations of the machine learning devices 5a and 5 by software. Therefore, a computer that executes the machine learning program will be described.
[0080] Fig. 11 is a diagram showing a hardware configuration of a computer that executes the machine learning programs according to Examples 1 and 2. As shown in Fig. 11, the computer 50 includes a main memory 51, a CPU (Central Processing Unit) 52, a LAN (Local Area Network) interface 53, and an HDD (Hard Disk Drive) 54. The computer 50 also includes a super IO (Input Output) 55, a DVI (Digital Visual Interface) 56, and an ODD (Optical Disk Drive) 57.
[0081] The main memory 51 is a memory that stores programs, intermediate results of program execution, etc. The CPU 52 is a central processing unit that reads and executes programs from the main memory 51. The CPU 52 includes a chipset having a memory controller.
[0082] The LAN interface 53 is an interface for connecting the computer 50 to another computer via a LAN. The HDD 54 is a disk device for storing programs and data, and the super IO 55 is an interface for connecting input devices such as a mouse and a keyboard. The DVI 56 is an interface for connecting a liquid crystal display device, and the ODD 57 is a device for reading and writing DVDs.
[0083] The LAN interface 53 is connected to the CPU 52 via a PCI Express (PCIe), and the HDD 54 and the ODD 57 are connected to the CPU 52 via a Serial Advanced Technology Attachment (SATA). The super IO 55 is connected to the CPU 52 via a Low Pin Count (LPC).
[0084] The machine learning program executed in computer 50 is stored on a DVD, which is an example of a recording medium readable by computer 50, and is read from the DVD by ODD 57 and installed in computer 50. Alternatively, the machine learning program is stored in a database or the like of another computer system connected via LAN interface 53, and is read from this database and installed in computer 50. The installed machine learning program is then stored in HDD 54, read into main memory 51, and executed by CPU 52.
[0085] In addition, in the first and second embodiments, a case where a neural network is used as the machine learning model has been described, but the machine learning devices 5a and 5 may use other machine learning models. In the first and second embodiments, a case where input data is classified based on a machine learning model has been described, but a system instead of the classification system 1 may make a judgment or a prediction based on a machine learning model. In the first and second embodiments, a case where an output value of a neuron in an output layer is used as a feature of input data has been described, but the machine learning devices 5a and 5 may use other values as a feature of input data. In the second embodiment, a case where a central coordinate and density of a cluster are used has been described, but the machine learning device 5 may use other information as information indicating a distribution of the feature of input data. In the first and second embodiments, a case where a drift-compatible model is updated has been described, but the machine learning devices 5a and 5 may use a past machine learning model other than the drift-compatible model. [Explanation of symbols]
[0086] 1 Classification System 2 Input Sensor 3 Data storage device 4 Classifier 5,5a Machine learning device 6 Display device 10, 10a Deterioration detection section 11, 11a Model memory section 12,12a Classification section 13 Cluster Generation Section 13a Distribution information generation section 14, 14a Judgment section 15,15a Label attachment section 16,16a Model Access Section 17,17a Learning Section 20,20a Model Management Department 21,21a Past model memory section 22,22a Registration Section 23,23a Comparison section 24,24a Selection section 50 Computer 51 Main memory 52 CPU 53 LAN Interface 54 HDD 55 Super IO 56 DVI 57 ODD
Claims
1. inputting first data into a second machine learning model among one or more machine learning models generated by updating a parameter of a first machine learning model, and identifying a first distribution of features of the first data calculated by the second machine learning model; comparing the first distribution with a distribution of features of the training data calculated by each of the one or more machine learning models for training data used to update parameters when each of the one or more machine learning models is generated; selecting a particular machine learning model from the one or more machine learning models based on a result of the comparison; generating a new machine learning model by updating parameters of the specific machine learning model based on specific training data labeled based on the feature amount of the first data; A machine learning program that causes a computer to execute processing.
2. the comparing process is performed by referring to model information including the one or more machine learning models; and causing the computer to execute a process of adding the new machine learning model to the model information by associating the new machine learning model with a distribution of features of the specific training data.
2. The machine learning program according to claim 1 .
3. the first distribution includes center coordinates and density of clusters created by clustering the first data based on a determination result obtained by inputting the first data into the second machine learning model and feature amounts of the first data; The comparing process includes a process of comparing a center coordinate and a density of the first distribution with a center coordinate and a density of a cluster associated with each of the one or more machine learning models.
2. The machine learning program according to claim 1 .
4. further causing the computer to execute a process of determining whether or not an output accuracy of the second machine learning model has deteriorated based on the identified first distribution; The machine learning program according to claim 1 , 2 or 3 , characterized in that the comparing process is executed when the determining process determines that the second machine learning model has deteriorated.
5. the machine learning model includes a neural network; The feature amount is a value output by a neuron in an output layer of the neural network.
5. The machine learning program according to claim 1, wherein the machine learning program is a program for executing a machine learning process.
6. inputting first data into a second machine learning model among one or more machine learning models generated by updating a parameter of a first machine learning model, and identifying a first distribution of features of the first data calculated by the second machine learning model; comparing the first distribution with a distribution of features of the training data calculated by each of the one or more machine learning models for training data used to update parameters when each of the one or more machine learning models is generated; selecting a particular machine learning model from the one or more machine learning models based on a result of the comparison; generating a new machine learning model by updating parameters of the specific machine learning model based on specific training data labeled based on the feature amount of the first data; A machine learning method characterized in that processing is executed by a computer.
7. inputting first data into a second machine learning model among one or more machine learning models generated by updating a parameter of a first machine learning model, and identifying a first distribution of features of the first data calculated by the second machine learning model; comparing the first distribution with a distribution of features of the training data calculated by each of the one or more machine learning models for training data used to update parameters when each of the one or more machine learning models is generated; selecting a particular machine learning model from the one or more machine learning models based on a result of the comparison; A control unit that generates a new machine learning model by updating parameters of the specific machine learning model based on specific training data labeled based on the feature amount of the first data. A machine learning device comprising:
Citation Information
Patent Citations
Distributed processing system, learning model creating method, data processing method, learning model creating program, and data processing program
JP2017068710A
Method and apparatus for operating an automation system
JP2018513490A
Image processing apparatus, image processing method, and program
JP2020146433A
Dynamic updating of machine learning models
US20180285772A1
Predictive model updating system, predictive model updating method, and predictive model updating program
WO2016151618A1
