Learning apparatus, learning method, and recording medium

Abstract

To provide a learning apparatus and so on which can improve learning efficiency relating to learning of a model.SOLUTION: A learning apparatus according to the present invention has state prediction means, state estimation means, and learning processing means. The state prediction means has a first model which, in response to an input of a first video including a plurality of images, can acquire a prediction result for a state indicated by a first image group having at least one image different from the plurality of images. The state estimation means has a second model which, in response to an input of a second video including the plurality of images and a second image group having at least one image different from the plurality of images, can acquire an estimation result of a state indicated by the second image group. The learning processing means subjects the first model to learning for making the prediction result closer to the estimation result.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] This disclosure relates to techniques that can be used to train models. [Background technology]

[0002] Techniques have been proposed that can be used to train models.

[0003] Specifically, for example, Patent Document 1 discloses a method for acquiring multiple new training data, individually weighting data requiring caution contained in each of the acquired multiple new training data, and training a trained model. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent Publication No. 2021-99702 Summary of the Invention [Problem to be solved by the invention]

[0005] However, the technology disclosed in Patent Document 1 has a problem in that, for example, when new training data contains elements that are common to elements contained in data already used for learning, the learning efficiency of the model may decrease.

[0006] An object of the present disclosure is to provide a learning device that can improve the efficiency of learning performed to adapt a model to a new environment. [Means for solving the problem]

[0007] In one aspect of the present disclosure, a learning device includes: a state prediction means having a first model capable of obtaining a predicted result of a state shown by a first image group having at least one image different from a first video input including a plurality of images; a state estimation means having a second model capable of obtaining an estimated result of a state shown by a second image group having at least one image different from the plurality of images in response to a second video input including the plurality of images and a second image group having at least one image different from the plurality of images; and a learning processing means that performs learning on the first model to bring the predicted result closer to the estimated result.

[0008] In another aspect of the present disclosure, a computer-implemented learning method inputs a first video including a plurality of images into a first model to obtain a predicted result of a state shown by a first group of images having at least one image different from the plurality of images, inputs a second video including the plurality of images and a second group of images having at least one image different from the plurality of images into a second model to obtain an estimated result of the state shown by the second group of images, and performs learning on the first model to bring the predicted result closer to the estimated result.

[0009] In yet another aspect of the present disclosure, a recording medium records a program that causes a computer to execute a process of inputting a first video including a plurality of images into a first model to obtain a prediction result of a state represented by a first group of images having at least one image different from the plurality of images, inputting a second video including the plurality of images and a second group of images having at least one image different from the plurality of images into a second model to obtain an estimation result of a state represented by the second group of images, and performing learning on the first model to bring the prediction result closer to the estimation result. [Effects of the Invention]

[0010] According to the present disclosure, it is possible to improve the efficiency of learning performed to adapt a model to a new environment. [Brief explanation of the drawings]

[0011] [Figure 1] FIG. 1 is a diagram showing a schematic configuration of a learning device according to the present disclosure. [Figure 2] FIG. 1 is a block diagram illustrating an example of a hardware configuration of a learning device according to the present disclosure. [Figure 3] FIG. 1 is a block diagram showing an example of the functional configuration of a learning device according to the present disclosure. [Figure 4] 10 is a flowchart showing an example of processing performed in a learning device according to the present disclosure. [Figure 5] FIG. 10 is a block diagram showing another example of the functional configuration of the learning device according to the present disclosure. [Figure 6] 10 is a flowchart showing another example of processing performed in the learning device according to the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0012] Hereinafter, preferred embodiments of the present disclosure will be described with reference to the drawings.

[0013] First Embodiment [Schematic configuration] 1 is a diagram showing a schematic configuration of a learning device according to the present disclosure. Learning device 100 performs processing related to learning of world model unit 12 using a plurality of time-series images included in video VC input from the outside.

[0014] The world model unit 12 has a model capable of predicting future changes in the state of the external environment shown by the video VC, for example. The prediction result of the world model unit 12 can be used, for example, to control the operation of a robot arm configured to grasp and move an object included in the video VC. The prediction result of the world model unit 12 can also be used, for example, to control a vehicle that performs autonomous driving in accordance with the video VC. The prediction result of the world model unit 12 can be acquired, for example, as a feature FL included in a transition feature FH described below.

[0015] For example, the learning device 100 can acquire a video VH (described below) as a first video and a video VC (described below) as a second video. Furthermore, the learning device 100 can acquire a prediction result of a state represented by a first image group having at least one image different from the first image group in response to input of a first video including a plurality of images. Furthermore, the learning device 100 can acquire an estimation result of a state represented by a second image group in response to input of a second video including the plurality of images and a second image group having at least one image different from the first image group. Furthermore, the learning device 100 can perform learning to bring the prediction result closer to the estimation result.

[0016] [Hardware configuration] 2 is a block diagram showing an example of the hardware configuration of a learning device according to the present disclosure. As shown in FIG. 2, learning device 100 includes an interface (IF) 111, a processor 112, a memory 113, a recording medium 114, and a database (DB) 115.

[0017] The IF 111 inputs and outputs data to and from an external device, and receives, for example, a video VC from the external device.

[0018] Processor 112 is a computer such as a CPU (Central Processing Unit), and executes a prepared program to control the entire learning device 100. Processor 112 performs, for example, processing related to learning by world model unit 12.

[0019] The memory 113 is configured by a ROM (Read Only Memory), a RAM (Random Access Memory), etc. The memory 113 is also used as a working memory while the processor 112 is executing various processes.

[0020] Recording medium 114 is a non-volatile, non-transitory recording medium such as a disk-shaped recording medium or semiconductor memory, and is configured to be detachable from learning device 100. Recording medium 114 records various programs executed by processor 112. When learning device 100 executes various processes, the programs recorded on recording medium 114 are loaded into memory 113 and executed by processor 112.

[0021] The DB 115 stores, for example, data received by the IF 111, data obtained by processing by the processor 112, and the like.

[0022] [Function Configuration] Fig. 3 is a block diagram showing an example of the functional configuration of a learning device according to the present disclosure. As shown in Fig. 3, learning device 100 includes an image acquisition unit 11, a world model unit 12, a similarity calculation unit 13, a weighting processing unit 14, and a learning processing unit 15. World model unit 12 includes a transition model unit 12A and an estimation model unit 12B.

[0023] The video acquisition unit 11 functions as a video acquisition means. The video acquisition unit 11 acquires a video VC including a plurality of images in time series. The video acquisition unit 11 acquires a video including a plurality of images that belong to a period before a predetermined timing TM in the video VC as a video VH. For example, the video acquisition unit 11 acquires a video VH that includes n images from timing T1 to timing Tn (n≧2) that belong to a period before the predetermined timing TM in the video VC. In this embodiment, it is assumed that at least one image is included in the portion of the video VC after the predetermined timing TM. The video acquisition unit 11 outputs the video VH to the transition model unit 12A and outputs the video VC to the estimation model unit 12B.

[0024] According to the above-described processing, the video acquisition unit 11 can acquire a video VH that includes a plurality of images. Furthermore, according to the above-described processing, the video acquisition unit 11 can acquire a video VC that includes a plurality of images that are identical to the images included in the video VH and an image group that has at least one image that is different from the plurality of images.

[0025] The transition model unit 12A has a machine learning model MM1 that has been trained to be able to predict how a state shown by an input video will transition, for example, using the input video. In other words, the transition model unit 12A has a machine learning model MM1 that, in response to input of a video VH including a plurality of images, is able to obtain a prediction result of a state shown by an image group that includes at least one image different from the plurality of images. The machine learning model MM1 of the transition model unit 12A can be referred to as, for example, a prediction model or a state transition model.

[0026] The transition model unit 12A functions as a state prediction unit. In response to an input of a video VH, the transition model unit 12A acquires a transition feature FH, which includes a feature FK corresponding to each image of the video VH and a feature FL corresponding to a predicted result of a state that may occur after the video VH, and outputs the acquired transition feature FH to the similarity calculation unit 13. That is, in response to an input of the video VH, the transition model unit 12A can acquire a transition feature FH, which includes a feature FL corresponding to a predicted result of a state that may occur after a predetermined timing TM. The feature FL is acquired, for example, as a feature corresponding to at least one image that indicates a state that may occur after the video VH. In this embodiment, unless otherwise specified, a case will be described in which feature FL is acquired corresponding to the same number of images as the number of images included in the portion of the video VC after the predetermined timing TM.

[0027] The transition model 12A inputs the learning data GD described below into the machine learning model MM1, thereby acquiring feature values ​​FU corresponding to each image included in the learning data GD, and outputs the acquired feature values ​​FU to the similarity calculation unit 13.

[0028] The estimation model unit 12B has a machine learning model MM2 that has been trained to be able to estimate, for example, the state shown by an input video using the video. In other words, the estimation model unit 12B has a machine learning model MM2 that, in response to input of a video VC that includes a plurality of images identical to each image included in the video VH and an image group having at least one image different from the plurality of images, can obtain an estimation result of the state shown by the plurality of images and an estimation result of the state shown by the image group. The machine learning model MM2 of the estimation model unit 12B can be referred to as, for example, an observation model or a state estimation model.

[0029] The estimation model unit 12B functions as a state estimation means. In addition, in response to the input of the video VC, the estimation model unit 12B acquires feature amounts corresponding to each image of the video VC as estimated feature amounts FC, and outputs the acquired estimated feature amounts FC to the similarity calculation unit 13. That is, in response to the input of the video VC, the estimation model unit 12B can acquire estimated feature amounts FC corresponding to the estimation result of a state that occurred after a predetermined timing TM.

[0030] The estimation model unit 12B inputs the learning data GD described below into the machine learning model MM1, thereby acquiring feature values ​​FW corresponding to each image included in the learning data GD, and outputs the acquired feature values ​​FW to the similarity calculation unit 13.

[0031] The similarity calculation unit 13 functions as a similarity calculation means. Furthermore, the similarity calculation unit 13 uses the transition feature FH and the estimated feature FC to calculate a similarity DF between an image corresponding to the transition feature FH and an image corresponding to the estimated feature FC, and outputs the calculated similarity DF to the weighting processing unit 14. Furthermore, the similarity calculation unit 13 uses the feature FU and the feature FW to calculate a similarity DG between an image corresponding to the feature FU and an image corresponding to the feature FW, and outputs the calculated similarity DG to the learning processing unit 15.

[0032] The weighting processing unit 14 functions as a weight setting means. Furthermore, the weighting processing unit 14 uses the similarity DF to set a weight WF as a value relating to the importance of learning for each image included in the video VC. The weight WF can be rephrased as, for example, an index relating to the importance of learning for each image included in the video VC. Furthermore, the weight WF can be rephrased as, for example, a value relating to the necessity of learning for each image included in the video VC. A specific example of a method for setting the weight WF will be described later. Furthermore, in response to the input of the video VC, the weighting processing unit 14 outputs the video VW to the learning processing unit 15, in which a weight WF has been assigned to each image of the video VC.

[0033] The learning processing unit 15 functions as a learning processing means. The learning processing unit 15 performs learning on the machine learning models MM1 and MM2 to bring the prediction results obtained by the machine learning models MM1 closer to the estimation results obtained by the machine learning models MM2. The learning processing unit 15 acquires data including at least one image after a predetermined timing TM in the video VW as learning data GD used for training the machine learning models MM1 and MM2. The learning processing unit 15 outputs the learning data GD to the transition model unit 12A and the estimation model unit 12B. The learning processing unit 15 may acquire data including at least one image after a predetermined timing TM in the video VW and at least one image before the predetermined timing TM in the video VW as the learning data GD. In the following description, unless otherwise specified, it is assumed that weights assigned to images are included in the learning data GD.

[0034] After the training data GD is output, the training processing unit 15 uses the similarity DG obtained from the similarity calculation unit 13 to calculate the loss LG when training is performed using the training data GD. The loss LG is calculated as a value that increases as the similarity DG decreases and decreases as the similarity DG increases, for example.

[0035] The learning processing unit 15 uses the loss LG to calculate parameters PT to be applied to the machine learning model MM1 of the transition model unit 12A and parameters PE to be applied to the machine learning model MM2 of the estimation model unit 12B. It is desirable that the parameters PT and PE be calculated as parameters that contribute to reducing the loss LG, for example.

[0036] The learning processing unit 15 outputs the parameter PT to the transition model unit 12A and outputs the parameter PE to the estimation model unit 12B. Through this processing, the learning processing unit 15 can update the existing parameters set in the machine learning model MM1 of the transition model unit 12A with the parameter PT, and can also update the existing parameters set in the machine learning model MM2 of the estimation model 12B with the parameter PE.

[0037] According to the processing described above, the learning processing unit 15 can perform learning to increase the similarity DG as learning to bring the prediction results obtained by the machine learning model MM1 closer to the estimation results obtained by the machine learning model MM2.

[0038] [Specific example] Next, specific examples of processing etc. performed in the present disclosure will be described.

[0039] The video acquisition unit 11 acquires a video VC1 including a plurality of images in time series. The video acquisition unit 11 also acquires a video including a plurality of images that belong to a time period before a predetermined time TM in the video VC1 as a video VH1. The video acquisition unit 11 also outputs the video VH1 to the transition model unit 12A and outputs the video VC1 to the estimation model unit 12B. In this specific example, unless otherwise specified, a case will be described in which a video VH1 including n images from a time T1 to a time Tn that belong to a time period before the predetermined time TM in the video VC1 is acquired. In this specific example, a case will be described in which a single image IC1 corresponding to an image at a time T(n+1) is included in a portion of the video VC1 after the predetermined time TM.

[0040] In response to input of video VH1, the transition model unit 12A acquires a transition feature FH1 including a feature FK1 corresponding to each image of the video VH1 and a feature FL1 corresponding to one image IH1 that indicates a state that may occur after the video VH1, and outputs the acquired transition feature FH1 to the similarity calculation unit 13. The feature FK1 is acquired as a feature corresponding to each of n images from timing T1 to timing Tn. Furthermore, the feature FL1 is acquired as a feature corresponding to the image at timing T(n+1).

[0041] The transition model unit 12A can acquire, as the feature FK1 and the feature FL1, values ​​indicating features for each image or values ​​indicating features for each object included in the image, for example. That is, the transition model unit 12A can acquire, as the feature FL1 included in the transition feature FH1, values ​​indicating features for each image corresponding to the prediction result or values ​​indicating features for each object included in the image corresponding to the prediction result. The transition model unit 12A can also acquire, as the feature FK1 and the feature FL1, features for each image or each object expressed in the form of a vector, distribution, or map. Specifically, the transition model unit 12A can acquire, as the feature FK1 and the feature FL1, at least one feature vector in which multiple values ​​indicating the features of the image or the features of the object are arranged. The transition model unit 12A can also acquire, as the feature FK1 and the feature FL1, a feature distribution represented by multiple values ​​indicating the features of the image and the variance of the multiple values. Furthermore, the transition model unit 12A can acquire, for example, a feature map having a plurality of values ​​that can distinguish between characteristic parts and other parts in an image as the feature amount FK1 and the feature amount FL1.

[0042] In response to input of video VC1, the estimation model unit 12B acquires features corresponding to each image of the video VC1 as estimated features FC1 and outputs the acquired estimated features FC1 to the similarity calculation unit 13. The estimated features FC1 include features corresponding to each of the n images from timing T1 to timing Tn and features corresponding to the image at timing T(n+1). In other words, the estimated features FC1 include features that can be used to reconstruct each of the n images from timing T1 to timing Tn and features that can be used to reconstruct the image at timing T(n+1).

[0043] The estimation model unit 12B can acquire, as the estimated feature quantity FC1, for example, a value indicating the feature of each image corresponding to the estimation result or a value indicating the feature of each object included in the image corresponding to the estimation result. Furthermore, the estimation model unit 12B can acquire the estimated feature quantity FC1, which expresses the feature of each image or the feature of each object in the form of a vector, a distribution, or a map. Specifically, the estimation model unit 12B can acquire, as the estimated feature quantity FC1, at least one feature vector in which multiple values ​​indicating the feature of the image or the feature of the object are arranged. Furthermore, the estimation model unit 12B can acquire, as the estimated feature quantity FC1, a feature distribution represented by multiple values ​​indicating the feature of the image and the variance of the multiple values. Furthermore, the estimation model unit 12B can acquire, as the estimated feature quantity FC1, a feature map having multiple values ​​that can distinguish between characteristic parts and other parts in the image.

[0044] The similarity calculation unit 13 uses the transition feature FH1 and the estimated feature FC1 to calculate a similarity DF1 between an image corresponding to the transition feature FH1 and an image corresponding to the estimated feature FC1, and outputs the calculated similarity DF1 to the weighting processing unit 14.

[0045] The similarity calculation unit 13 can calculate the similarity DF1 using different methods depending on the representation format of the transition feature FH1 and the estimated feature FC1. Specifically, for example, when the transition feature FH1 and the estimated feature FC1 are represented as feature vectors, the similarity calculation unit 13 can calculate the similarity DF1 according to either Euclidean distance or cosine similarity. Furthermore, for example, when the transition feature FH1 and the estimated feature FC1 are represented as feature distributions, the similarity calculation unit 13 can calculate the similarity DF1 according to KL (Kullback-Leibler) divergence. Furthermore, for example, when the transition feature FH1 and the estimated feature FC1 are represented as feature maps, the similarity calculation unit 13 can calculate the similarity DF1 according to the least square error between an image restored from the feature map of the transition feature FH1 and an image restored from the feature map of the estimated feature FC1.

[0046] According to the above-described process, the similarity calculation unit 13 can calculate a relatively large similarity DF1 when the similarity between the image corresponding to the transition feature FH1 and the image corresponding to the estimated feature FC1 is high. Therefore, the similarity calculation unit 13 can calculate a relatively large similarity DF1 as the similarity between, for example, each image of the video VH1 and each image of the video VC1 other than image IC1. Furthermore, according to the calculation result of such similarity DF1, each image of the video VH1 and each image of the video VC1 other than image IC1 can be treated as "images corresponding to the learned environment" or "images including common elements."

[0047] According to the above-described process, the similarity calculation unit 13 can calculate a relatively small value of the similarity DF1 when the similarity between the image corresponding to the transition feature FH1 and the image corresponding to the estimated feature FC1 is low. Therefore, the similarity calculation unit 13 can calculate a relatively small value of the similarity DF1 as the similarity between, for example, the image IH1 and the image IC1. Furthermore, according to the calculation result of such a similarity DF1, the image IH1 and the image IC1 can be treated as "images corresponding to an unlearned environment" or "images not containing common elements."

[0048] When values ​​indicating the features of each of a plurality of objects included in an image are acquired as the transition feature FH1 and the estimated feature FC1, the similarity calculation unit 13 can calculate, as the similarity DF1, a plurality of similarities according to the number of objects included in the image. Specifically, when values ​​indicating the features of each of three objects included in an image are acquired as the transition feature FH1 and the estimated feature FC1, for example, the similarity calculation unit 13 can calculate, as the similarity DF1, three similarities corresponding to the three objects.

[0049] The weighting processing unit 14 uses the similarity DF1 to set a weight WF1 as a value relating to the importance of learning for each image included in the video VC1. In response to the input of the video VC1, the weighting processing unit 14 outputs to the learning processing unit 15 a video VW1 in which the weight WF1 is assigned to each image of the video VC1.

[0050] Here, a first specific example of a method for setting the weight WF1 by the weighting processing unit 14 will be described.

[0051] The weighting processing unit 14 sets the weight WF1 so that it becomes a relatively large value when the similarity DF1 is less than a predetermined value TD, and becomes a relatively small value when the similarity DF1 is equal to or greater than the predetermined value TD.

[0052] According to the process of this specific example, the weighting processing unit 14 can set a value such as "1.5" or "2.0" as the value of the weight WF1 corresponding to the case where the similarity DF1 is less than a predetermined value TD. According to the process of this specific example, the weighting processing unit 14 can set a value such as "0.1" or "0.5" as the value of the weight WF1 corresponding to the case where the similarity DF1 is equal to or greater than the predetermined value TD. According to the process of this specific example, the weighting processing unit 14 can assign a weight WF1 having a relatively small value to each image other than image IC1 of video VC1. According to the process of this specific example, the weighting processing unit 14 can assign a weight WF1 having a relatively large value to image IC1 of video VC1.

[0053] Next, a second specific example of a method for setting the weight WF1 by the weighting processing unit 14 will be described.

[0054] The weighting processing unit 14 sets the weight WF1 so that when the similarity DF1 is less than a predetermined value TD, the weight WF1 is "1.0", and when the similarity DF1 is equal to or greater than the predetermined value TD, the weight WF1 is "0.0".

[0055] According to the process of this specific example, the weighting processing unit 14 can assign "0.0" as the weight WF1 to each image of the video VC1 other than the image IC1. Also, according to the process of this specific example, the weighting processing unit 14 can assign "1.0" as the weight WF1 to the image IC1 of the video VC1, for example.

[0056] For example, when the weighting processing unit 14 acquires multiple similarities DF1 according to the number of objects contained in an image, it can set a weight WF1 according to whether the average value of the multiple similarities DF1 is less than a predetermined value TD, and assign the set weight WF1 to the image.

[0057] For example, even if the video VC1 contains multiple images in a portion after a predetermined timing TM, the weighting processing unit 14 can set the weight WF1 in the manner described above using the similarity DF1 calculated for each of the multiple images.

[0058] According to the processing described above, the weighting processing unit 14 can assign a weight WF1 having a relatively large value to an image, among the images included in the video VC1, for which the prediction result obtained by the machine learning model MM1 is not similar to the estimation result obtained by the machine learning model MM2. Furthermore, according to the processing described above, the weighting processing unit 14 can assign a weight WF1 having a relatively small value to an image, among the images included in the video VC1, for which the prediction result obtained by the machine learning model MM1 is similar to the estimation result obtained by the machine learning model MM2.

[0059] The learning processing unit 15 acquires data including at least one image after a predetermined timing TM in the video VW1 as learning data GD1 used for learning the transition model unit 12A and the estimation model unit 12B. The learning processing unit 15 also outputs the learning data GD1 to the transition model unit 12A and the estimation model unit 12B.

[0060] For example, the learning processing unit 15 can acquire data including each image of the video VW1 as the training data GD1. Alternatively, the learning processing unit 15 can extract, from each image of the video VW1, all images to which a weight WF1 equal to or greater than a predetermined value TW has been assigned, and acquire data including the extracted images as the training data GD1. Alternatively, the learning processing unit 15 can extract, from each image of the video VW1, all images to which a weight WF1 equal to or greater than a predetermined value TW has been assigned and at least one image to which a weight WF1 less than the predetermined value TW has been assigned, and acquire data including the extracted images as the training data GD1. That is, according to the processing of this specific example, the learning processing unit 15 can acquire data including at least the image IC1 as the training data GD1. Furthermore, according to the processing of this specific example, the learning processing unit 15 can acquire data including images extracted from the video VW based on the weight WF1 as the training data GD1. According to the process of this specific example, the learning processing unit 15 can prevent, for example, images of the video VW1 to which a weight of "0.0" is assigned from being used as the learning data GD1.

[0061] The learning processing unit 15 may, for example, extract images to which a weight WF1 equal to or greater than a predetermined value TW has been assigned from among the images in the video VW1, and store the extracted images in the storage medium 114 and / or DB 115. This process allows the user to check what kind of images are included in the learning data GD1.

[0062] In response to input of the learning data GD1, the transition model 12A acquires a feature FU1 corresponding to each image included in the learning data GD1, and outputs the acquired feature FU1 to the similarity calculation unit 13. It is desirable that the transition model 12A acquires, as the feature FU1, a feature having the same expression format as the transition feature FH1, for example.

[0063] In response to input of the learning data GD1, the estimation model unit 12B acquires a feature FW1 corresponding to each image included in the learning data GD1, and outputs the acquired feature FW1 to the similarity calculation unit 13. It is desirable that the estimation model unit 12B acquires, as the feature FW1, a feature having the same expression format as the estimated feature FC1, for example.

[0064] The similarity calculation unit 13 uses the feature FU1 and the feature FW1 to calculate a similarity DG1 between an image corresponding to the feature FU1 and an image corresponding to the feature FW1, and outputs the calculated similarity DG1 to the learning processing unit 15. The similarity calculation unit 13 can calculate the similarity DG1 using, for example, a method similar to the method for calculating the similarity DF1.

[0065] The learning processing unit 15 uses the similarity DG1 obtained from the similarity calculation unit 13 after the learning data GD1 is output to calculate the loss LG1 when learning is performed using the learning data GD1.

[0066] For example, when image IC1 is used as training data GD1, the learning processing unit 15 can calculate, as the loss LG1, a value obtained by multiplying the similarity DG1 calculated for image IC1 by the weight WF1 assigned to image IC1. Furthermore, when a plurality of images including at least image IC1 are used as training data GD1, the learning processing unit 15 can calculate, as the loss LG1, a sum of multiple values ​​obtained by multiplying the similarity DG1 calculated for each of the plurality of images by the weight WF1 assigned to each of the plurality of images.

[0067] The learning processing unit 15 uses the loss LG1 to calculate a parameter PT1 to be applied to the machine learning model MM1 of the transition model unit 12A and a parameter PE1 to be applied to the machine learning model MM2 of the estimation model unit 12B. Furthermore, the learning processing unit 15 outputs the parameter PT1 to the transition model unit 12A and outputs the parameter PE1 to the estimation model unit 12B. Through this processing, the learning processing unit 15 can update the existing parameters set in the machine learning model MM1 of the transition model unit 12A with the parameter PT1, and can also update the existing parameters set in the machine learning model MM2 of the estimation model 12B with the parameter PE1.

[0068] According to the processing of this specific example, the learning processing unit 15 can calculate the loss LG1 that treats the image IC1 included in the learning data GD1 as an important image in the learning of the world model unit 12, and can calculate the parameters PT1 and PE1 that contribute to reducing the loss LG1. Furthermore, according to the processing of this specific example, the learning processing unit 15 can perform learning on the transition model unit 12A and the estimation model unit 12B to increase the similarity DF1. Furthermore, according to the processing of this specific example, the learning processing unit 15 can use the weight WF1 to calculate the loss LG1 related to learning to increase the similarity DF1.

[0069] [Processing flow] Next, a description will be given of the flow of processing performed in the learning device 100. Fig. 4 is a flowchart showing an example of processing performed in the learning device according to the present disclosure.

[0070] First, the learning device 100 acquires a first video and a second video (step S11). The learning device 100 can acquire, for example, a video VH as the first video in step S11. The learning device 100 can also acquire, for example, a video VC as the second video in step S11.

[0071] The learning device 100 acquires transition features by inputting the first video acquired in step S11 to the transition model unit 12A (step S12). The transition features in step S12 include, for example, features corresponding to each image of the first video acquired in step S11 and features corresponding to prediction results of states that may occur after the first video.

[0072] The learning device 100 acquires estimated features by inputting the second video acquired in step S11 to the estimation model unit 12B (step S13). The estimated features in step S13 include, for example, features corresponding to each image of the second video acquired in step S11.

[0073] Next, the learning device 100 calculates the similarity between the image corresponding to the transition feature acquired in step S12 and the image corresponding to the estimated feature acquired in step S13 (step S14).

[0074] Next, the learning device 100 uses the similarity calculated in step S14 to set a weight as a value relating to the importance of learning for each image included in the video VC (step S15).

[0075] Next, learning device 100 assigns the weight set in step S15 to each image of the second video acquired in step S11 (step S16).

[0076] Next, the learning device 100 acquires data including at least one image after a predetermined timing TM in the second video to which the weighting in step S16 has been applied as learning data to be used for learning the transition model unit 12A and the estimation model 12B (step S17).

[0077] Next, the learning device 100 uses the learning data obtained in step S17 to perform learning on the machine learning model MM1 of the transition model unit 12A and the machine learning model MM2 of the estimation model 12B (step S18).

[0078] As described above, according to this embodiment, images included in a video that correspond to an unlearned environment can be treated as important images in the learning of the world model unit 12, and images included in the video that correspond to an environment that has already been learned can be treated as unimportant images in the learning of the world model unit 12. Therefore, according to this embodiment, it is possible to improve the learning efficiency related to the learning that is performed to adapt the model to a new environment.

[0079] Second Embodiment FIG. 5 is a block diagram showing another example of the functional configuration of the learning device according to the present disclosure.

[0080] The learning device 500 has the same hardware configuration as the learning device 100. The learning device 500 also has a state prediction means 511, a state estimation means 512, and a learning processing means 513. The state prediction means 511 has a model 611. The state estimation means 512 has a model 612.

[0081] The state prediction means 511 can be realized, for example, by using the function of the transition model unit 12A. The state estimation means 512 can be realized, for example, by using the function of the estimation model unit 12B. The learning processing means 513 can be realized, for example, by using the function of the learning processing unit 15.

[0082] FIG. 6 is a flowchart showing another example of processing performed in the learning device according to the present disclosure.

[0083] The state prediction means 511 inputs a first video including a plurality of images into the model 611, thereby obtaining a prediction result of the state shown by a first image group having at least one image different from the plurality of images (step S51).

[0084] The state estimation means 512 inputs a second video including a plurality of images and a second group of images having at least one image different from the plurality of images into the model 612, thereby obtaining an estimation result of the state indicated by the second group of images (step S52).

[0085] The learning processing means 513 performs learning on the model 611 to bring the prediction result closer to the estimation result (step S53).

[0086] According to this embodiment, it is possible to improve the efficiency of learning that is performed to adapt a model to a new environment.

[0087] A part or all of the above-described embodiments can be described as, but not limited to, the following supplementary notes.

[0088] (Appendix 1) a state prediction means having a first model capable of obtaining a prediction result of a state shown by a first image group having at least one image different from the first image group, in response to an input of a first video including a plurality of images; a state estimation means having a second model capable of acquiring an estimation result of a state shown by a second image group in response to input of a second video including the plurality of images and a second image group having at least one image different from the plurality of images; a learning processing means for performing learning on the first model so as to bring the prediction result closer to the estimation result; A learning device having the above configuration.

[0089] (Appendix 2) further comprising a similarity calculation means for calculating a similarity between a first feature amount obtained by inputting learning data used for the learning into the first model and a second feature amount obtained by inputting the learning data into the second model; 2. The learning device according to claim 1, wherein the learning processing means performs, as the learning, learning on the first model to increase the similarity.

[0090] (Appendix 3) 3. The learning device of claim 2, wherein the learning processing means acquires data including at least one image in the second video after a predetermined timing as the learning data.

[0091] (Appendix 4) The learning device of Appendix 2, wherein the learning processing means acquires, as the learning data, data including images extracted from each image included in the second video, the images whose prediction results are not similar to the estimation results.

[0092] (Appendix 5) further comprising a weight setting means for setting a weight as an index relating to the importance of learning for each image included in the second video; The learning device of Appendix 2, wherein the learning processing means acquires as the learning data data including all images extracted from each image included in the second video, the weights of which are set to a predetermined value or more.

[0093] (Appendix 6) further comprising a weight setting means for setting a weight as an index relating to the importance of learning for each image included in the second video; The learning device of Appendix 2, wherein the learning processing means acquires, as the learning data, data including all images extracted from each image included in the second video, for which the weight is set to a predetermined value or more, and at least one image for which the weight is set to less than the predetermined value.

[0094] (Appendix 7) further comprising a weight setting means for setting a weight as an index relating to the importance of learning for each image included in the second video; 3. The learning device according to claim 2, wherein the learning processing means calculates a loss related to the learning using the weights.

[0095] (Appendix 8) 3. The learning device of claim 2, wherein the first feature and the second feature are acquired as features expressed in the form of a vector, a distribution, or a map.

[0096] (Appendix 9) 1. A computer-implemented learning method comprising: inputting a first video including a plurality of images into a first model to obtain a prediction result of a state represented by a first image group having at least one image different from the plurality of images; inputting a second video including the plurality of images and a second group of images having at least one image different from the plurality of images into a second model, thereby obtaining an estimation result of the state represented by the second group of images; A learning method in which the first model is trained to bring the prediction result closer to the estimation result.

[0097] (Appendix 10) inputting a first video including a plurality of images into a first model to obtain a prediction result of a state represented by a first image group having at least one image different from the plurality of images; inputting a second video including the plurality of images and a second group of images having at least one image different from the plurality of images into a second model, thereby obtaining an estimation result of the state represented by the second group of images; A recording medium having a program recorded thereon that causes a computer to execute a process of performing learning on the first model to bring the prediction result closer to the estimation result.

[0098] Although the present disclosure has been described above with reference to the embodiments, the present disclosure is not limited to the above-described embodiments and examples. Various modifications that can be understood by those skilled in the art can be made to the configuration and details of the present disclosure within the scope of the present disclosure. Furthermore, each embodiment can be combined with other embodiments as appropriate. [Explanation of symbols]

[0099] 11 Video acquisition unit 12 World Model Department 12A Transition Model Section 12B Estimation Model Section 13 Similarity calculation unit 14 Weighting processing section 15 Learning processing unit 100 Learning Device

Claims

1. a state prediction means having a first model capable of obtaining a prediction result of a state shown by a first image group having at least one image different from the first image group, in response to input of a first video including a plurality of images; a state estimation means having a second model capable of acquiring an estimation result of a state shown by a second image group in response to input of a second video including the plurality of images and a second image group having at least one image different from the plurality of images; a learning processing means for performing learning on the first model so as to bring the prediction result closer to the estimation result; A learning device having the above configuration.

2. the system further comprises a similarity calculation means for calculating a similarity between a first feature amount obtained by inputting learning data used for the learning into the first model and a second feature amount obtained by inputting the learning data into the second model, 2. The learning device according to claim 1, wherein the learning processing means performs, as the learning, learning on the first model to increase the degree of similarity.

3. The learning device according to claim 2 , wherein the learning processing means acquires, as the learning data, data including at least one image in the second video after a predetermined timing.

4. The learning device according to claim 2, wherein the learning processing means acquires as the learning data data data including images extracted from each image included in the second video, the prediction result of which is not similar to the estimation result.

5. further comprising a weight setting means for setting a weight as an index relating to the importance of learning for each image included in the second video; The learning device according to claim 2, wherein the learning processing means acquires as the learning data data including all images extracted from each image included in the second video and for which the weight is set to a predetermined value or more.

6. further comprising a weight setting means for setting a weight as an index relating to the importance of learning for each image included in the second video; The learning device described in claim 2, wherein the learning processing means acquires as the learning data data including all images extracted from each image included in the second video, the images having the weight set to a predetermined value or more, and at least one image having the weight set to less than the predetermined value.

7. further comprising a weight setting means for setting a weight as an index relating to the importance of learning for each image included in the second video; The learning device according to claim 2 , wherein the learning processing means calculates a loss related to the learning using the weight.

8. The learning device according to claim 2 , wherein the first feature amount and the second feature amount are acquired as feature amounts expressed in the form of a vector, a distribution, or a map.

9. 1. A computer-implemented learning method comprising: inputting a first video including a plurality of images into a first model to obtain a prediction result of a state represented by a first group of images having at least one image different from the plurality of images; inputting a second video including the plurality of images and a second group of images having at least one image different from the plurality of images into a second model, thereby obtaining an estimation result of the state shown by the second group of images; A learning method in which the first model is trained to bring the prediction result closer to the estimation result.

10. inputting a first video including a plurality of images into a first model to obtain a prediction result of a state represented by a first group of images having at least one image different from the plurality of images; inputting a second video including the plurality of images and a second group of images having at least one image different from the plurality of images into a second model, thereby obtaining an estimation result of the state shown by the second group of images; A recording medium having a program recorded thereon that causes a computer to execute a process of performing learning on the first model to bring the prediction result closer to the estimation result.

Images