Learning device, learning method, and program

Abstract

Provided is a learning device, wherein a model acquisition means acquires: a first prediction model that generates a first prediction feature amount that indicates a state of each object at the next time on the basis of an input object feature amount; and a second prediction model that generates a second prediction feature amount that indicates the state of each object at the next time on the basis of the input object feature amount and an image at the next time. An identification model training means trains an identification model that identifies an output from the first prediction model and an output from the second prediction model. A re-training means re-trains the first prediction model so as to reduce the accuracy of identification by the identification model.

Description

Technical Field

[0001] The present disclosure relates to a technique for predicting the future temporal transition of an object in an image.

Background Art

[0002] A world model is known as a technique for obtaining the representation of individual objects from unlabeled images such as the position and label of an object and predicting future temporal transitions. Patent Document 1 discloses a technique for performing robot learning by combining imitation learning and reinforcement learning while obtaining a world model through an expert and predicting the future.

Prior Art Documents

Non-Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] When learning a model for predicting the temporal transition of an object in a video, a learning video is used, but the performance of the model is evaluated only based on the difference from the learning video. Therefore, when the learned model includes probabilistic transitions in the future, it is impossible to determine the plausibility of the results generated by the transitions.

[0005] One object of the present disclosure is to learn a prediction model that is not overly constrained by the learning video and can predict plausible temporal transitions.

Means for Solving the Problems

[0006] In one aspect of the present disclosure, the learning device A model acquisition means that acquires a first prediction model that generates a first prediction feature that indicates the state of each object at the next time based on input object features, and a second prediction model that generates a second prediction feature that indicates the state of each object at the next time based on the input object features and the image at the next time. A discriminant model learning means for learning a discriminant model that distinguishes between the output from the first predictive model and the output from the second predictive model, A retraining means for retraining the first prediction model such that the identification accuracy of the aforementioned identification model decreases, It is equipped with.

[0007] In other respects of this disclosure, the learning methods performed by computers are: A first prediction model generates a first prediction feature that indicates the state of each object at the next time based on the input object features, and a second prediction model generates a second prediction feature that indicates the state of each object at the next time based on the input object features and the image at the next time, A discriminant model is trained to distinguish between the output from the first predictive model and the output from the second predictive model. The first prediction model is retrained so that the recognition accuracy of the aforementioned recognition model decreases.

[0008] In other aspects of this disclosure, program teeth, A first prediction model generates a first prediction feature that indicates the state of each object at the next time based on the input object features, and a second prediction model generates a second prediction feature that indicates the state of each object at the next time based on the input object features and the image at the next time, A discriminant model is trained to distinguish between the output from the first predictive model and the output from the second predictive model. The computer is made to perform a process of retraining the first prediction model so that the recognition accuracy of the aforementioned recognition model decreases. [Brief explanation of the drawing]

[0009] [Figure 1] The concept of a learning device according to the first embodiment is shown. [Figure 2] This is a block diagram showing the hardware configuration of the learning device. [Figure 3] This is a block diagram showing the functional configuration of the learning device for performing the first learning step. [Figure 4] This is a block diagram showing the functional configuration of the learning device for performing the second learning step. [Figure 5] This is a block diagram showing the functional configuration of the learning device for performing the third learning step. [Figure 6] This is the main routine for the learning process performed by the learning device. [Figure 7] This is a flowchart of the learning process for the transition model and the inference model. [Figure 8] This is a flowchart of the training process for the discriminative model. [Figure 9] This is a flowchart of the learning process for the transition model. [Figure 10] Examples of future predictions using transition models and inference models are schematically shown. [Figure 11] The training method for the discriminative model using variation 1 is shown below. [Figure 12] This is a block diagram showing the configuration of the learning device according to the second embodiment. [Figure 13] This is a flowchart of the processing performed by the learning device according to the second embodiment. [Modes for carrying out the invention]

[0010] Preferred embodiments of this disclosure will be described below with reference to the drawings. <First Embodiment> [Concept of a learning device] FIG. 1 shows the concept of the learning device according to the first embodiment. The learning device 100 learns a prediction model that predicts the future temporal transition of an object included in an input image. Such a prediction model is also called a world model. For example, in the case of an input image obtained by photographing an environment in which a robot arm moves an object, the prediction model acquires the representation of the objects included in the input image and predicts the future temporal transition of each object. In the above example, the objects to be predicted include the robot arm itself, the object moved by the robot arm, and fixed objects arranged in the photographed environment.

[0011] In this embodiment, the learning device 100 basically learns a prediction model through the following three learning steps. (1) First learning step The learning device 100 learns two prediction models as models that predict the future temporal transition of an object. The first prediction model is a model (hereinafter also referred to as a "transition model") that outputs a first object feature amount indicating the state of each object at the next time based on the object feature amounts of the input objects. The second prediction model is a model (hereinafter also referred to as an "inference model") that outputs a second object feature amount indicating the state of each object at the next time based on the object feature amounts of the input objects and the actual image at the next time. (2) Second learning step The learning device 100 learns a discrimination model that discriminates between the output from the transition model and the output from the inference model. (3) Third learning step The learning device 100 relearns the transition model so that the discrimination model cannot discriminate between the output from the transition model and the output from the inference model.

[0012] By executing the above learning steps, the learning device 100 can generate a transition model that can accurately predict the future temporal transition of an object even without an image of the next time. The details of each learning step will be described later.

[0013] [Hardware configuration] Figure 2 is a block diagram showing the hardware configuration of the learning device 100. As shown in the figure, the learning device 100 includes an interface (IF) 12, a processor 13, a memory 14, a recording medium 15, a database (DB) 16, a display unit 17, and an input unit 18.

[0014] IF12 acquires input images from an external source. The input images are still images or videos capturing the state of an object in a given environment. IF12 also outputs the transition model and inference model obtained through training to an external source.

[0015] The processor 13 is a computer such as a CPU (Central Processing Unit) and controls the entire learning device 100 by executing a pre-prepared program. The processor 13 can be a CPU, GPU (Graphics Processing Unit), DSP (Digital Signal Processor), MPU (Micro Processing Unit), FPU (Floating Point Number Processing Unit), PPU (Physics Processing Unit), TPU (Tensor Processing Unit), quantum processor, microcontroller, or a combination thereof. The processor 13 performs the learning process described later.

[0016] Memory 14 consists of ROM (Read Only Memory), RAM (Random Access Memory), and other components. Memory 14 stores various programs executed by the processor 13. Memory 14 is also used as working memory while the processor 13 is executing various processes.

[0017] The recording medium 15 is a non-volatile, non-temporary recording medium such as a disk-shaped recording medium or semiconductor memory, and is configured to be detachable from the learning device 100. The recording medium 15 stores various programs that the processor 13 executes. When the learning device 100 performs various processes, the programs stored in the recording medium 15 are loaded into the memory 14 and executed by the processor 13.

[0018] DB16 may, if necessary, store the input images received through IF12 during training. DB16 may also store the inference and classification models generated during the training process.

[0019] The display unit 17 is composed of, for example, a liquid crystal display. The input unit 18 includes, for example, a keyboard or mouse. The display unit 17 and the input unit 18 are used, for example, when an operator of the learning device 100 performs necessary operation inputs.

[0020] [Functional Configuration] (First learning step) Figure 3 is a block diagram showing the functional configuration of the learning device 100a for executing the first learning step. In the first learning step, the learning device 100a comprises a loss calculation unit 21a and a learning unit 22a. In the first learning step, the learning device 100a learns a transition model MA and an inference model MB. The transition model MA and the inference model MB are composed of neural networks. In Figure 3, multiple transition model MA and inference model MB are shown, but these are shown for convenience to show that different input images are being processed; in reality, there is only one transition model MA and one inference model MB.

[0021] Now, assume there is a time-series input image from time t - 1 to time t + 1. Here, t is a natural number, and the minimum unit between each frame of the video is set to 1. The frame rate of the video can be any value, but the transition can also be estimated at an interval of n frames. In that case, for example, perform transition prediction such as t ≧ t + n ≧ t + 2n. More generally, using a random integer s (0 ≦ s < n), transition prediction such as t + s ≧ t + s + n ≧ t + s + 2n can be performed during learning.

[0022] When predicting the state from time t - 1 to time t, the object feature amounts of each object at time t - 1 predicted by the inference model MB are input into the transition model MA and the inference model MB. The object feature amount is an object center representation indicating the position, shape, etc. of each object, and for example, a feature vector can be used. Further, the actual image at time t is input into the inference model MB and the loss calculation unit 21a.

[0023] The transition model MA predicts the transition of each object based on the input object feature amounts at time t - 1, calculates the probability distribution of the state of the object at the next time t for each object (hereinafter referred to as "state distribution"), and outputs it to the loss calculation unit 21a. The state distribution of the object can be, for example, a vector including a feature vector indicating the state of the object and a vector indicating the probability distribution of the state of the object. Note that as described above, the "state distribution" represents the state of each object as a probability distribution, and the "object feature amount" is a definite state vector sampled based on this state distribution.

[0024] The transition model MB calculates the object feature amount for each object included in the input image at time t. Also, the transition model MA calculates the state distribution of the object at the next time t for each object based on the obtained object feature amount and the image at time t, and outputs it to the loss calculation unit 21a.

[0025] The loss calculation unit 21a receives the state distribution at time t generated by the transition model MA, the state distribution at time t generated by the inference model MB, and the actual image at time t as input. First, the loss calculation unit 21a calculates the distribution distance between the state distribution at time t generated by the transition model MA and the state distribution at time t generated by the inference model MB. For example, KL divergence (Kullback-Leibler divergence) can be used as the distribution distance. Next, the loss calculation unit 21a reconstructs the image at time t from the state distribution at time t generated by the inference model MB to generate a reconstructed image, and calculates the error between the reconstructed image and the actual image at time t that was input (hereinafter referred to as the "reconstruction error"). Finally, the loss calculation unit 21a outputs the calculated distribution distance and reconstruction error to the learning unit 22a.

[0026] One example of reconstruction error is to represent the image as a 3×W×H vector using RGB (3 channels), the number of horizontal pixels W, and the number of height pixels H. The Euclidean distance between the vector of the ground truth image and the vector of the reconstructed image can then be used as the reconstruction error. In another example, the negative value (a value with a negative sign) of the log-likelihood (log likelihood) between the probability distribution where the vector of the ground truth image is the mean and the variance of each element is a predetermined σ, and the vector of the reconstructed image, can be used as the reconstruction error.

[0027] The learning device 100a performs similar processing using the object features at time t predicted by the inference model MB and the image at time t+1. Specifically, the loss calculation unit 21a calculates the distribution distance between the state distribution at time t+1 generated by the transition model MA and the state distribution at time t+1 generated by the inference model MB. The loss calculation unit 21a also reconstructs the image at time t+1 from the state distribution at time t+1 generated by the inference model MB to generate a reconstructed image, and calculates the reconstruction error between this reconstructed image and the actual image at time t+1 that was input. The loss calculation unit 21a then outputs the calculated distribution distance and reconstruction error to the learning unit 22a.

[0028] The learning unit 22a then uses the sum of the inter-distribution distances and reconstruction errors input from the two loss calculation units 21a as the loss to train the transition model MA and the inference model MB. Specifically, the learning unit 22a updates the parameters of the transition model MA and the inference model MB so that the inter-distribution distances and reconstruction errors become smaller. As a result, the transition model MA and the inference model MB are trained so that the state distribution output by the transition model MA and the state distribution output by the inference model MB become closer, and the reconstructed image output by the inference model MB becomes closer to the actual image input to the loss calculation unit 21a. Thus, in the first training step, the transition model MA and the inference model MB are trained using the input image.

[0029] In addition to the methods described above, the first learning step may also involve using other known global models such as ViMON, OP3, G-SWM, and GATSBI.

[0030] (Second learning step) Figure 4 is a block diagram showing the functional configuration of the learning device 100b for executing the second learning step. In the second learning step, the learning device 100b learns a discriminative model MC that distinguishes between the output of the transition model MA and the output of the inference model MB. The discriminative model MC is composed of a neural network. In the second learning step, the discriminative model MC is trained using the transition model MA and inference model MB that were trained in the first learning step. That is, in the second learning step, the transition model MA and the inference model MB are not the targets of training.

[0031] Let's assume we have a time-series input image from time t-1 to time t+1. As shown in Figure 4, the object features of each object at time t-1 are input to the transition model MA and the inference model MB, and the image at time t is input to the inference model MB. The transition model MA generates a state distribution at time t from the object features at time t-1, reconstructs the image at time t from the state distribution at time t, and outputs it to the discrimination model MC. The inference model MB generates a state distribution at time t from the object features at time t-1 and the actual image at time t, reconstructs the image at time t from the state distribution at time t, and outputs it to the discrimination model MC. The image output from the transition model MA is the image predicted by the transition model MA. In contrast, the image output from the inference model MB is the image generated using the actual image.

[0032] The discrimination model MC distinguishes between reconstructed images input from the transition model MA (also called "transition model-derived") and reconstructed images input from the inference model MB (also called "inference model-derived"). The discrimination model MC identifies whether the input image is from the transition model or the inference model and outputs the identification result to the loss calculation unit 21b. The loss calculation unit 21b is input with a ground truth label indicating whether the input image is from the transition model or the inference model. The loss calculation unit 21b outputs the error between the identification result of the discrimination model MC and the ground truth label as the loss to the learning unit 22b.

[0033] The same processing is performed on the image at time t. Specifically, the transition model MA generates a state distribution at time t+1 from the object features at time t, reconstructs the image at time t+1 from the state distribution, and outputs it to the discrimination model MC. The inference model MB generates a state distribution at time t+1 from the object features at time t and the actual image at time t+1, reconstructs the image at time t+1 from the state distribution, and outputs it to the discrimination model MC. The discrimination model MC identifies whether the input image originates from the transition model or the inference model and outputs the discrimination result to the loss calculation unit 21b. The loss calculation unit 21b outputs the error between the discrimination result of the discrimination model MC and the correct label as the loss to the learning unit 22b.

[0034] Then, the learning unit 22b updates the parameters of the discrimination model MC so that the sum of the losses input from the two loss calculation units 21b becomes small. In this way, the discrimination model is trained. Although the above example uses two loss calculation units 21b, in reality, it is sufficient to input the image at each time point to a single loss calculation unit 21b, calculate the loss at each time point, and output these to the learning unit 22b.

[0035] In the example above, the learning device 100b learns the discrimination model MC using reconstructed images (i.e., still images) input from the transition model MA and the inference model MB. Alternatively, the learning device 100b may learn the discrimination model MC based on a series of reconstructed images (i.e., a video) output from the transition model MA and the inference model MB at multiple consecutive time points. In this case, the loss calculation unit 21b only needs to calculate the error between the video output from the transition model MA and the video output from the inference model MB as the loss.

[0036] Furthermore, the learning device 100c may learn the discrimination model MC using object features generated by the transition model MA and the inference model MB instead of images. In this case as well, the loss calculation unit 21b learns the discrimination model MC using the error between the discrimination result of the discrimination model MC and the correct label as the loss. Moreover, the learning device 100c may learn the discrimination model MC using object features output from the transition model MA and the inference model MB at multiple consecutive time points, i.e., time-series object features.

[0037] (Third learning step) Figure 5 is a block diagram showing the functional configuration of the learning device 100c for executing the third learning step. In the third learning step, the learning device 100c retrains the transition model MA learned in the first learning step using the inference model MB learned in the first learning step and the discriminant model MC learned in the second learning step. That is, in the third learning step, the inference model MB and the discriminant model MC are not the targets of learning.

[0038] Now, let's assume we have a time-series input image from time t-1 to time t+1. As shown in Figure 5, the object features of each object at time t-1 are input to the transition model MA and the inference model MB, and the image at time t is input to the inference model MB. The transition model MA generates a state distribution at time t from the object features at time t-1, reconstructs the image at time t from the state distribution at time t, and outputs it to the discrimination model MC. The inference model MB generates a state distribution at time t from the object features at time t-1 and the actual image at time t, reconstructs the image at time t from the state distribution at time t, and outputs it to the discrimination model MC. The discrimination model MC identifies whether the input image originates from the transition model or the inference model, and outputs the discrimination result to the loss calculation unit 21c.

[0039] The loss calculation unit 21c receives a ground truth label indicating whether the input image originates from the transition model or the inference model. The loss calculation unit 21c calculates the loss based on the identification result of the identification model MC and the ground truth label, and outputs it to the learning unit 22c. Here, the learning unit 22c retrains the transition model MA so that the accuracy of identification by the identification model MC decreases. For example, the loss calculation unit 21c generates a loss that is either a negative value (a value with a negative sign) of the error between the identification result of the identification model MC and the ground truth label, or the reciprocal of the error with the ground truth label, and the learning unit 22c updates the parameters of the transition model MA so that the loss becomes smaller. In this way, the transition model MA is retrained.

[0040] [Learning Process] Next, the learning process performed by the learning device 100 will be described. Figure 6 shows the main routine of the learning process performed by the learning device 100. This process is realized when the processor 13 shown in Figure 2 executes a pre-prepared program and operates as each of the elements shown in Figures 3 to 5.

[0041] The learning process is carried out in the three learning steps described above. First, as the first learning step, the learning device 100a shown in Figure 3 learns the transition model MA and the inference model MB (step S10). Next, as the second learning step, the learning device 100b shown in Figure 4 learns the discrimination model MC (step S20). Finally, as the third learning step, the learning device 100c shown in Figure 5 retrains the transition model MA (step S30).

[0042] Figure 7 is a flowchart of the learning process for the transition model and the inference model. First, the transition model MA acquires object features for each object, and the inference model MB acquires object features for each object and the image for the next time (step S11). Next, the transition model MA predicts the state distribution for the next time based on the object features, and the inference model MB predicts the state distribution for the next time based on the object features and the image for the next time (step S12). Next, the loss calculation unit 21a calculates the distribution distance between the state distributions output by the transition model MA and the inference model MB (step S13). The loss calculation unit 21a also reconstructs the image for the next time from the state distribution output by the inference model MB and calculates the reconstruction error with the actual image for the next time (step S14). Next, the learning unit 22a trains the transition model MA and the inference model MB so that the distribution distance and the reconstruction error are small (step S15). Then, the process returns to the main routine in Figure 6.

[0043] Figure 8 is a flowchart of the learning process for the classification model. First, the transition model MA acquires object features for each object, and the inference model MB acquires object features for each object and an image for the next time (step S21). Next, the transition model MA predicts the state distribution for the next time based on the object features, and the inference model MB predicts the state distribution for the next time based on the object features and the image for the next time (step S22). Next, the transition model MA and the inference model MB output the prediction results to the classification model MC. The prediction results output by the transition model MA and the inference model MB may be still images derived from each model, videos, object features (object-centered representations), or time-series object features, as described above. The classification model MC outputs the classification result to the loss calculation unit 21b. The loss calculation unit 21b calculates the classification error between the classification result by the classification model MC and the correct label (step S23). Next, the learning unit 22b trains the classification model MC so that the classification error becomes smaller (step S24). Then, the process returns to the main routine shown in Figure 6.

[0044] Figure 9 is a flowchart of the retraining process for the transition model. First, the transition model MA acquires object features for each object, and the inference model MB acquires object features for each object and the image for the next time (step S31). Next, the transition model MA predicts the state distribution for the next time based on the object features, and the inference model MB predicts the state distribution for the next time based on the object features and the image for the next time (step S32). Next, the transition model MA and the inference model MB input the prediction results to the discrimination model MC. The prediction results output by the transition model MA and the inference model MB may be still images, videos, object features (object-centered representations), or time-series object features, as described above. The discrimination model MC outputs the discrimination result to the loss calculation unit 21c. The loss calculation unit 21c calculates the loss using the discrimination result from the discrimination model MC and the correct label (step S33). Next, the learning unit 22c retrains the transition model MA so that the discrimination accuracy of the discrimination model MC decreases (step S34). Then, the process returns to the main routine shown in Figure 6, and the learning process ends.

[0045] [effect] As described above, in the first embodiment, a discrimination model MC is generated to distinguish between images derived from the transition model and images derived from the inference model, and the transition model MA is retrained so that discrimination by the discrimination model MC becomes difficult, that is, the discrimination accuracy decreases. As a result, the transition model MA can predict the most likely image at a given time as the predicted image at that time. Consequently, even if the prediction result of the transition model MA differs from the result of the inference model MB, it is possible to prevent the transition from being judged as correct even when a plausible future prediction is made considering probabilistic behavior, or to prevent the result from appearing unnatural.

[0046] [Example of future prediction] Next, we will explain an example of future prediction using the transition model MA and inference model MB learned by the learning device 100 described above. Figure 10 schematically shows an example of future prediction using the transition model MA and inference model MB.

[0047] Suppose we are given images from time t0 to t2. In this case, the prediction device uses the inference model MB to predict the state for the time periods t0 to t2 when images exist. That is, at time t0, the prediction device inputs the image at time t0 into the inference model MB. At time t1, the prediction device inputs the object features of each object generated by the inference model MB at time t, along with the actual image at time t1, into the inference model MB. Furthermore, at time t2, the prediction device inputs the object features of each object generated by the inference model MB at time t1, along with the actual image at time t2, into the inference model MB.

[0048] On the other hand, since no actual images exist from time t3 onward, the prediction device uses the transition model MA to make predictions. Specifically, at time t3, the prediction device inputs the object features of each object generated by the inference model MB at time t2 into the transition model MA. Furthermore, at time t4, the prediction device inputs the object features of each object generated by the transition model MA at time t3 into the transition model MA. In this way, future predictions can be made using the transition model MA even from time t3 onward, when no actual images exist.

[0049] [Differentiation] Next, a modified version of the first embodiment will be described. The following modifications can be combined and applied as appropriate. (Variation 1) In the above embodiment, during the training of the discrimination model MC, the image at each time step is input to the discrimination model MC, as shown in Figure 4, the loss is calculated, and the discrimination model MC is updated. Alternatively, the discrimination model MC may be trained using images that have passed through the transition model MA and the inference model MB multiple times at consecutive time points.

[0050] Figure 11 shows the learning method for the discrimination model according to Modification 1. As shown in the figure, the object features of each object at time t-2 are input to the transition model MA and the inference model MB, the object features output by them are input to the transition model MA and the inference model MB respectively, and the reconstructed images output by them are input to the discrimination model MC. The discrimination model MC outputs a discrimination result for the reconstructed images of the transition model MA and the inference model MB at time t, and the loss calculation unit 21b calculates the loss using the discrimination result. In this way, the discrimination model MC may be learned using the results of passing the object features at a certain time through the transition model MA and the inference model MB multiple times in succession.

[0051] The reason for performing this processing is as follows: When the transition model MA finally obtained by the learning device 100 is put into practical use, as explained with reference to Figure 10, it is conceivable to predict the future temporal changes of an object over multiple time points in the absence of actual images. Therefore, the above method makes it possible to train the discrimination model MC to accurately distinguish between objects based on the results of passing the transition model MA through multiple consecutive times.

[0052] This method can also be applied when retraining the transition model MA after training the discriminative model MC. That is, when retraining the transition model MA, the learning device 100c may retrain the transition model MA in such a way that the discriminative accuracy of the discriminative model MC decreases, using the results of passing the transition model MA and the inference model MB through multiple consecutive cycles.

[0053] (Modification 2) In the learning process described above, as shown in Figure 6, the learning process is completed by performing the first to third learning steps using all the prepared input images. Alternatively, the prepared input images may be divided into predetermined units, and the first to third learning steps shown in Figure 6 may be repeated for each unit.

[0054] (Variation 3) In the first learning step, some action may be input to the transition model MA and the inference model MB. For example, when applying the method of the above embodiment to predicting an environment in which an object is moved using a robotic arm, an action specifying the movement of the robotic arm may be input to the transition model MA and the inference model MB at a certain time. This enables the transition model MA to predict future temporal transitions if the above action is performed.

[0055] (Modification 4) In the above embodiment, the transition model MA and the inference model MB are input to images, which are either still images or moving images. However, these images are not necessarily images captured by a normal camera, but may be data generated by a sensor or the like. For example, distance images measured using a depth camera or point cloud data generated using a LiDAR can also be used as input images in this embodiment.

[0056] Furthermore, supplementary data may be input to the transition model MA and inference model MB in addition to images. For example, audio data collected by a microphone, or output data from sensors such as weight sensors and pressure sensors may be input to the transition model MA and inference model MB along with images. In this case, there are two possible ways to use the supplementary data. In the first method, the supplementary data may be used as a target for transition prediction in the same way as the images described above. That is, in the first learning step, the learning device 100a should calculate the inter-distribution distance and reconstruction error of the state distribution for the supplementary data in addition to the images, and learn the transition model MA and inference model MB. On the other hand, in the second method, the supplementary data may not be used as a target for transition prediction, but only as supplementary data for image transition prediction. In this case, the supplementary data should be treated in transition prediction in the same way as the actions in the modified example 3 described above.

[0057] [Examples of application] The above embodiment can be applied to the medical and healthcare fields. For example, the above embodiment can be applied to transition prediction and control of a medical robot that assists in patient surgery. In this case, the medical robot uses the transition model MA and inference model MB (hereinafter referred to as the "world model") learned by the learning device 100 of the embodiment to grasp the environment in the operating room, the patient's physical condition, and the instruments and procedures necessary for surgery. Specifically, the medical robot can grasp the state of the patient's body in real time based on information obtained from cameras and sensors, and can select the instruments necessary for surgery and adjust the procedures.

[0058] The following is an example of a specific process when using a surgical robot as a medical robot. (1) Surgical robots are equipped with cameras and sensors to understand the patient's internal condition, the location of surgical instruments, and the procedure in real time. (2) The world model learned by the embodiment provides information necessary during surgery based on the information grasped by the surgical robot. (3) The surgical robot selects the instruments and adjusts the procedure required for the surgery based on information provided from a learned world model.

[0059] This allows surgical robots utilizing a world model to work collaboratively with doctors and nurses. The surgical robot can autonomously operate in cooperation with doctors and nurses by using a world model that predicts their movements. Furthermore, by using a predictive model to accurately recognize the patient's condition, the surgical robot can perform surgery with minimal invasiveness.

[0060] <Second Embodiment> Figure 12 is a block diagram showing the configuration of a learning device according to the second embodiment. The learning device 70 according to the second embodiment includes a model acquisition means 71, an identification model learning means 72, and a retraining means 73.

[0061] Figure 13 is a flowchart of the processing performed by the learning device 70 according to the second embodiment. The model acquisition means 71 acquires a first prediction model that generates a first prediction feature that indicates the state of each object at the next time based on the input object feature, and a second prediction model that generates a second prediction feature that indicates the state of each object at the next time based on the input object feature and the image at the next time (step S71). The discrimination model learning means 72 learns a discrimination model that distinguishes between the output from the first prediction model and the output from the second prediction model (step S72). The retraining means 73 retrains the first prediction model so that the discrimination accuracy of the discrimination model decreases (step S73).

[0062] According to the learning device 70 of the second embodiment, it becomes possible to learn a model capable of predicting plausible temporal transitions.

[0063] Some or all of the above embodiments may also be described as follows, but are not limited to the following:

[0064] (Note 1) A model acquisition means that acquires a first prediction model that generates a first prediction feature that indicates the state of each object at the next time based on input object features, and a second prediction model that generates a second prediction feature that indicates the state of each object at the next time based on the input object features and the image at the next time. A discriminant model learning means for learning a discriminant model that distinguishes between the output from the first predictive model and the output from the second predictive model, A retraining means for retraining the first prediction model such that the identification accuracy of the aforementioned identification model decreases, A learning device equipped with the following features.

[0065] (Note 2) The aforementioned discrimination model learning means is a learning device described in Appendix 1 that learns the discrimination model using the first predicted feature and the second predicted feature.

[0066] (Note 3) The learning device described in Appendix 1, wherein the discriminant model learning means learns the discriminant model using a first predicted image reconstructed from the first predicted features and a second predicted image reconstructed from the second predicted features.

[0067] (Note 4) The discriminant model learning means is a learning device according to Appendix 1, which learns the discriminant model using a first predicted feature output by the first predictive model after iterating through multiple predictions, and a second predicted feature output by the second predictive model after iterating through multiple predictions.

[0068] (Note 5) The learning device described in Appendix 1, wherein the retraining means retrains the first prediction model such that the discrimination accuracy of the discrimination model decreases between the first prediction feature quantity output by the first prediction model after iterating through predictions multiple times and the second prediction feature quantity output by the second prediction model after iterating through predictions multiple times.

[0069] (Note 6) The first prediction model predicts a first state distribution for each object at the next time step, and generates the first predicted features based on the first state distribution. The learning device described in Appendix 1 predicts a second state distribution for each object at the next time step and generates a second predicted feature based on the second state distribution.

[0070] (Note 7) The system comprises a predictive model learning means for learning the first predictive model and the second predictive model, The prediction model learning means learns the first prediction model and the second prediction model such that the distance between the first state distribution and the second state distribution becomes small, and the error between the image reconstructed from the second prediction feature and the actual image at the next time step becomes small. The model acquisition means is a learning device according to Appendix 6 that acquires the first predictive model and the second predictive model learned by the predictive model learning means.

[0071] (Note 8) The aforementioned object includes a patient and medical equipment, The first prediction model outputs a first prediction feature that indicates the state of the patient and medical device at the following time point. The second prediction model is a learning device as described in Appendix 1, which outputs a second predictive feature that indicates the state of the patient and medical device at the next time.

[0072] (Note 9) A learning method performed by a computer, A first prediction model generates a first prediction feature that indicates the state of each object at the next time based on the input object features, and a second prediction model generates a second prediction feature that indicates the state of each object at the next time based on the input object features and the image at the next time, A discriminant model is trained to distinguish between the output from the first predictive model and the output from the second predictive model. A learning method for retraining the first prediction model such that the recognition accuracy of the aforementioned recognition model decreases.

[0073] (Note 10) A first prediction model generates a first prediction feature that indicates the state of each object at the next time based on the input object features, and a second prediction model generates a second prediction feature that indicates the state of each object at the next time based on the input object features and the image at the next time, A discriminant model is trained to distinguish between the output from the first predictive model and the output from the second predictive model. A recording medium containing a program that causes a computer to perform a process of retraining the first prediction model so that the recognition accuracy of the aforementioned recognition model decreases.

[0074] Although the present disclosure has been described above with reference to embodiments and examples, the present disclosure is not limited to the above embodiments and examples. Various modifications to the structure and details of the present disclosure can be understood by those skilled in the art within the scope of the present disclosure. [Explanation of Symbols]

[0075] 13 processors 21a, 21b, 21c Loss calculation section 22a, 22b, 22c Learning Department 100a, 100b, 100c Learning Devices MA Transition Model MB Inference Model MC Identification Model

Claims

1. A model acquisition means for acquiring a first prediction model that generates a first prediction feature that indicates the state of each object at the next time based on input object features, and a second prediction model that generates a second prediction feature that indicates the state of each object at the next time based on the input object features and the image at the next time. A discriminant model learning means for learning a discriminant model that distinguishes between the output from the first predictive model and the output from the second predictive model, A retraining means for retraining the first prediction model so that the identification accuracy of the aforementioned identification model decreases, A learning device equipped with the following features.

2. The learning device according to claim 1, wherein the discrimination model learning means learns the discrimination model using the first predicted feature and the second predicted feature.

3. The learning device according to claim 1, wherein the discrimination model learning means learns the discrimination model using a first predicted image reconstructed from the first predicted features and a second predicted image reconstructed from the second predicted features.

4. The learning device according to claim 1, wherein the discriminant model learning means learns the discriminant model using a first predicted feature quantity output by the first predictive model after iterating through multiple predictions, and a second predicted feature quantity output by the second predictive model after iterating through multiple predictions.

5. The learning device according to claim 1, wherein the retraining means retrains the first prediction model such that the discrimination accuracy of the discrimination model decreases between the first prediction feature quantity output by the first prediction model after iterating through predictions multiple times and the second prediction feature quantity output by the second prediction model after iterating through predictions multiple times.

6. The first prediction model predicts a first state distribution for each object at the next time step, and generates the first predicted features based on the first state distribution. The learning device according to claim 1, wherein the second prediction model predicts a second state distribution for each object at the next time step and generates a second predicted feature based on the second state distribution.

7. The system comprises a predictive model learning means for learning the first predictive model and the second predictive model, The prediction model learning means learns the first prediction model and the second prediction model such that the distance between the first state distribution and the second state distribution becomes smaller, and the error between the image reconstructed from the second prediction feature and the actual image at the next time step becomes smaller. The learning device according to claim 6, wherein the model acquisition means acquires the first predictive model and the second predictive model learned by the predictive model learning means.

8. The aforementioned object includes a patient and medical equipment, The first prediction model outputs a first predictive feature that indicates the state of the patient and medical device at the following time point. The learning device according to claim 1, wherein the second prediction model outputs a second predictive feature that indicates the state of the patient and medical device at the next time.

9. A learning method performed by a computer, A first prediction model generates a first prediction feature that indicates the state of each object at the next time based on the input object features, and a second prediction model generates a second prediction feature that indicates the state of each object at the next time based on the input object features and the image at the next time, A discriminant model is trained to distinguish between the output from the first prediction model and the output from the second prediction model. A learning method for retraining the first prediction model such that the recognition accuracy of the aforementioned recognition model decreases.

10. A first prediction model generates a first prediction feature that indicates the state of each object at the next time based on the input object features, and a second prediction model generates a second prediction feature that indicates the state of each object at the next time based on the input object features and the image at the next time, A discriminant model is trained to distinguish between the output from the first prediction model and the output from the second prediction model. A program that causes a computer to perform a process of retraining the first prediction model so that the recognition accuracy of the aforementioned recognition model decreases.

Images