Apparatus and method for determining the uncertainty of sensor signals synthesized by a generative machine learning system

Abstract

This provides a method for determining the uncertainty of sensor signals synthesized by a generative machine learning system. [Solution] The method includes the steps of: acquiring a noise sample; performing Bayesian estimation on the generative machine learning system using the noise sample as input to the generative machine learning system, thereby determining the posterior predictive distribution of the sensor signal that would have been synthesized by the generative machine learning system using the noise sample as input to the generative machine learning system; and providing a measure of the variability of the posterior predictive distribution as uncertainty about the sensor signal synthesized from the noise sample, wherein the posterior predictive distribution characterizes the distribution of latent features of the sensor signal that would have been generated using the noise sample as input.

Description

[Technical Field]

[0001] Conventional technology Bayesian diffusion is disclosed in "BayesDiff: Estimating Pixel-wise Uncertainty in Diffusion via Bayesian Inference" by Kuo et al., 2024, https: / / arxiv.org / abs / 2310.11142.

[0002] Berry and Megar's "Efficient Epistemic Uncertainty Estimation in Regression Ensemble Models Using Pairwise-Distance Estimators," 2024, https: / / arxiv.org / abs / 2308.13498, discloses pairwise distance estimators for probability distributions. [Background technology]

[0003] Technical background and advantages of the invention Current technological systems that interact with physical reality typically require some form of internal model to model the state of each environment at a given time in order to determine what action to take. Such models can be implemented in the form of machine learning models that can process sensor signals of the technological system's environment to extract meaningful information, such as the presence of other objects, people, or animals.

[0004] To determine the conditions of physical reality based on sensor signals, such machine learning systems typically require vast amounts of data to make accurate predictions. Acquiring this data is an inefficient and costly task. Therefore, generative methods can be used to determine the sensor signals that would have occurred in physical reality. Conditional generative models further allow for the provision of additional information to the generative process to define the desired characteristics that should be present in the generated sensor signals.

[0005] While generative models, such as StableDiffusion, can achieve highly realistic sensor signal generation, they are machine learning models and are therefore susceptible to inaccurate modeling of the probability distribution of sensor signals obtained from physical reality. Consequently, they may provide outputs that do not accurately represent sensor signals from physical reality (for example, if sensor signals from physical reality are recorded, the output may have extremely low-probability or impossible values).

[0006] Therefore, it is desirable to identify sensor signals generated (also called "synthesized") by generative machine learning systems that are unlikely to occur in physical reality. In known methods such as Bayesian diffusion, Bayesian estimation is used to determine an uncertainty score for the generated sensor signal to determine whether the likelihood of the generated sensor signal is an accurate representation of the possible conditions in physical reality. High uncertainty can be understood as a low-quality (e.g., artifact-containing) synthesized sensor signal, while low uncertainty can be understood as a high-quality (e.g., clearly "realistic") synthesized sensor signal. [Prior art documents] [Non-patent literature]

[0007] [Non-Patent Document 1] Kuo et al., “BayesDiff: Estimating Pixel-wise Uncertainty in Diffusion via Bayesian Inference,” 2024, https: / / arxiv.org / abs / 2310.11142 [Non-Patent Document 2] Berry and Meger, “Efficient Epistemic Uncertainty Estimation in Regression Ensemble Models Using Pairwise-Distance Estimators,” 2024, https: / / arxiv.org / abs / 2308.13498 [Overview of the project] [Problems that the invention aims to solve]

[0008] However, the inventors have found that common methods evaluate the uncertainty of the generation process based on the generated values ​​of sensor signals. For example, when generating images, uncertainty is evaluated based on pixel-level differences. The inventors have empirically found that such uncertainty evaluations based on the "raw" values ​​of the synthesized sensor signals lead to uninformative estimations of generative uncertainty. For example, evaluating pixel-level differences in images generated by machine learning systems can result in high uncertainty scores when only a small number of pixels differ significantly from a numerical standpoint. [Means for solving the problem]

[0009] Advantageously, the above problem is solved by a method having the features of claim 1. In particular, the method assesses uncertainty of a synthesized sensor signal by determining uncertainty of latent factors (also known as latent features) of the generated sensor signal. By relying on latent factors instead of values of the sensor signal, the uncertainty assessment is performed on the "semantic" level of the generated sensor signal, in other words, the content of the sensor signal is assessed in terms of uncertainty rather than its raw values. The inventors have empirically found that such an approach provides a more accurate prediction of the validity of a synthesized sensor signal relative to physical reality, thereby achieving a more accurate machine learning system trained using the generated data, or achieving a more reliable estimation of the performance of a machine learning system even when testing is performed using low-uncertainty samples.

[0010] Disclosure of the Invention In a first aspect, the present invention relates to a computer-implemented method for determining uncertainty of a sensor signal synthesized by a generative machine learning system regarding the probability with which the sensor signal is observed in physical reality, the method comprising: obtaining a noise sample; performing Bayesian inference for the generative machine learning system using the noise sample as an input to the generative machine learning system, thereby determining a posterior predictive distribution of a sensor signal that would have been synthesized by the generative machine learning system using the noise sample as the input to the generative machine learning system; providing a variability measure of the posterior predictive distribution as the uncertainty for the sensor signal synthesized from the noise sample; the method wherein the posterior predictive distribution characterizes a distribution of latent features of a sensor signal that would have been generated using the noise sample as the input, the method being characterized in that.

[0011] It can also be understood that the above method, while generating a synthetic sensor signal, evaluates the uncertainty related to said sensor signal, that is, the probability that the synthesized sensor signal actually conforms to physical reality. Physical reality can also be understood as something that should be modeled by a generative machine learning system. That is to say, in order to generate synthetic samples of sensor signals, it can be understood that the generative machine learning system requires training based on a sample set of sensor signals recorded from physical reality. In this case, the generative machine learning system can "model" physical reality through the empirical distribution of sensor signals used to train the generative machine learning system.

[0012] It can be understood that the generative machine learning system is configured to receive samples from a noise distribution as input and provide an output characterizing a synthesized sensor signal. It should be noted that the output characterizing the synthesized sensor signal can be understood as an output comprising or consisting of the synthesized sensor signal. In order to construct a generative machine learning system that achieves this, the generative machine learning system is in particular trainable to map samples from a noise distribution to samples from a sensor signal distribution. In training, the distribution of sensor signals can in particular be represented by empirical samples of sensor signals recorded from physical reality.

[0013] In general, a generative machine learning system can provide its output unconditionally (e.g., using samples from a noise distribution as input) or conditionally (e.g., using additional information as input regarding what is to be represented in the synthesized sensor signal). The method itself is agnostic to the actual input of the generative machine learning model. For example, the generative machine learning model comprises or consists of, for example, flow matching, conditional flow matching, variational flow matching, conditional variational flow matching, normalized flows obtained by diffusion models or conditional diffusion models, continuous normalized flows, conditional normalized flows, and neural ODEs.

[0014] This method uses Bayesian estimation to determine the uncertainty regarding its output. In particular, Bayesian estimation can be based on any generative machine learning system that models its parameters through a distribution of those parameters. Alternatively, when using a neural network as or within a generative machine learning system, the final layer Laplacian approximation can be used to transform the otherwise "point estimates" of the generative machine learning system's parameters into a distribution of those parameters. Thus, advantageously, this method can be used "after the fact" for any generative machine learning system; that is, it does not require the generative machine learning system to use a distribution of its parameters.

[0015] A post-hoc prediction of the distribution of latent features in a sensor signal can be understood as the distribution of latent features becoming the post-hoc predictive distribution. Alternatively, the characterization may include a method for processing the distribution of latent features, such as scaling the latent distribution, or processing the latent features before determining the distribution of latent features (e.g., normalization or scaling).

[0016] Acquiring noise samples can be understood as deriving noise samples from a noise distribution. Alternatively, noise samples can also be provided to this method as input, for example, to be used later to synthesize sensor signals; however, here we must first evaluate whether each noise sample produces a realistic sensor signal.

[0017] Generally, common generative machine learning systems sample their output (e.g., sensor signals) by using noise samples as input. Therefore, a method for determining uncertainty can be considered a testing mechanism that uses a specific noise sample as input to a generative machine learning system to determine whether the sampled sensor signal is uncertain or not, in other words, whether the sensor signal is probable given the training data for the generative machine learning system.

[0018] In a preferred embodiment, determining the posterior predictive distribution is further: The steps include: • Using Monte Carlo sampling to derive multiple synthesized sensor signals from a generative machine learning system, The steps include determining the latent characteristics of each of the synthesized sensor signals, The steps include: determining the distribution of latent features based on the determined latent features, The steps include: providing the distribution of the latent feature as a posterior predictive distribution; Includes.

[0019] In other words, sensor signals can be sampled from a machine learning model to approximate the posterior predictive distribution. For this purpose, the sampled sensor signals are processed to extract latent features that characterize each sensor signal. The distribution of these latent features can then be determined based on the sample of latent features obtained.

[0020] Advantageously, these advantageous embodiments do not generate propagated probability distributions through generative processes such as Bayesian diffusion. Samples of parameters (also called weights) of a generative machine learning model are randomly derived and then made available as fixed parameters within the model to determine a synthetic sensor signal. In this case, the determined sensor signal can be used directly to extract latent features. In this way, determining the posterior predictive distribution has substantially higher computational efficiency than known methods.

[0021] In a preferred embodiment, latent features are determined by a feature extractor, which is configured to receive a sensor signal as input and provide the latent features of the sensor signal as output.

[0022] Feature extractors for extracting features from provided sensor signals are known in the art, particularly those that can be trained unsupervised. Common methods include (variational) autoencoders, (normalized) flow, or diffusion models. However, feature extractors may be supervised-trained, for example, through contrast learning. For example, when images are used as sensor signals output by a generative machine learning system, the inventors have discovered a visual model of the CLIP model that performs as well as a feature extractor. It should be noted that CLIP-like approaches can also be used for other modalities of sensor signals. For example, any type of sensor signal can be used in a CLIP-like approach using a general model equivalent to the “visual model,” in which case training can be performed using a standard CLIP approach (e.g., using pairs of sensor signals and text descriptions of each sensor signal as input).

[0023] Training the feature extractor may be an optional further step in this method.

[0024] In particular, the distribution of latent features is given by the following formula

number

number

[0025] In another embodiment, the present invention relates to a computer-implemented method for creating or augmenting datasets for training and / or testing machine learning systems, The method involves the following steps, namely: • The step of deriving noise samples from the noise distribution, • By providing noise samples to the method for determining the above uncertainty, the steps for determining uncertainty are as follows: • If the uncertainty is below a predetermined threshold, The step of synthesizing sensor signals from a generative machine learning system used in a method for determining the uncertainty, wherein the sensor signals are synthesized using noise samples as input, The steps include: creating or enhancing a dataset by adding the synthesized sensor signals to the dataset; This includes methods.

[0026] Advantageously, the above method allows the dataset to include only sensor signals that are truly "realistic" in relation to sensor signals sampled from physical reality, in order to train a generative machine learning system. As a result, this allows the machine learning model to be trained and / or tested with only high-quality sensor signals, in other words, the dataset does not include sensor signals that contain artifacts that make them unrealistic to the real world. For the training dataset, this allows the machine learning system to be trained using only high-quality synthesized sensor signals, further improving the performance of the machine learning system. In testing, the test results are more reliable because they are those that could actually occur in the real world and relate only to sensor signals that do not constitute artifacts.

[0027] If the variability scale does not satisfy the if statement, a new noise sample can be derived, and the steps of this method can be repeated.

[0028] "Synthesizing sensor signals from a generative machine learning system used in a method for determining uncertainty" can be understood as using the same generative machine learning system that is used to assess uncertainty. This can also be understood as a step in assessing uncertainty, which is part of a method for creating and / or augmenting a dataset.

[0029] If the uncertainty is below a predetermined threshold, the sensor signal is synthesized from the generative machine learning system using the noise sample as input. In the case of a Bayesian model, the mean mode of the parameters (e.g., the expected value of each parameter) can be selected as a point estimate, and then the synthesized sensor signal can be determined using the “frozen” parameter here. If the generative machine learning system is provided with point estimates of the parameters, and for example, a final layer Laplacian approximation is used for Bayesian estimation, then the sensor signal synthesized using the noise sample as input can be determined using the point estimates of the parameters in the supplied generative machine learning system, i.e., the point estimates of the original parameters.

[0030] In another embodiment, the present invention relates to a computer-implemented method for training and / or testing a machine learning system, • Steps to create training and / or test datasets according to a method for creating and / or augmenting datasets, • Steps to train and / or test a machine learning system using a dataset, This includes methods.

[0031] As part of the above method, sensor signals used in the dataset can also be annotated by human annotators or automated annotation methods. In this case, the machine learning system can be supervised and trained.

[0032] In various embodiments of all the methods presented above, the sensor signal may be a digital image or an audio signal. The machine learning system is trained on this data.

[0033] In various embodiments of all the methods presented above, the variability of the variability measure of the posterior predictive distribution may be the entropy of the posterior predictive distribution, or the variability may be determined based on the posterior predictive distribution using a pairwise distance estimator.

[0034] Embodiments of the present invention will be described in more detail with reference to the following drawings. The drawings show the following: [Brief explanation of the drawing]

[0035] [Figure 1] This flowchart shows a method for determining the uncertainty of sensor signals synthesized from a generative machine learning system. [Figure 2] This is a schematic diagram illustrating the method for creating a dataset. [Figure 3] This figure shows a training system that trains a machine learning system using a dataset. [Figure 4] This figure shows a control system, including a machine learning system, that controls actuators in the environment. [Figure 5] This figure shows a control system for controlling a robot that is at least partially autonomous. [Figure 6] This is a diagram showing a control system for controlling manufacturing machinery. [Modes for carrying out the invention]

[0036] Description of the Embodiment Figure 1 shows a flowchart of a computer-implemented method (900) for determining the uncertainty of a sensor signal synthesized by a generative machine learning system regarding the probability of observing the sensor signal in physical reality.

[0037] In the method, a generative machine learning system is acquired as input to the method. The generative machine learning system is configured to receive noise samples as input and provide a synthesized sensor signal as output. Optionally, the generative machine learning system can be trained in predetermined steps (not shown) of the method. The generative machine learning system may be a diffusion model in particular, especially a stable diffusion model or a diffusion model tuned by an adapter such as ControlNet or LoRA. In particular, the generative machine learning model can be formed from, or may include, a neural network that generates the sensor signal.

[0038] In the specific embodiment shown in Figure 1, the generated sensor signal is a digital image, but other modalities of the sensor signal are also possible in this method.

[0039] In the first step (901) of this method, noise samples are acquired. This acquisition can be achieved by receiving the noise samples as input, or by sampling the noise samples from a predetermined probability distribution, particularly a normal distribution, and even from a more specific multivariate normal distribution. The noise samples can be given in the form of real-valued vectors, that is, they can be sampled from a real-valued multivariate probability or density distribution.

[0040] In the second step (902) of the method, the noise sample is used as input to a generative machine learning model to perform a Bayesian estimation of the generative machine learning system. As a result of the Bayesian estimation, the posterior predictive distribution of the sensor signal that the generative machine learning system would have synthesized using the noise sample as input is determined. The posterior predictive distribution is configured to characterize the distribution of latent features of the sensor signal that would have been generated using the noise sample as input. This can be achieved, in particular, by Monte Carlo sampling the generative machine learning system to derive multiple samples of the synthesized sensor signal from which each latent feature is extracted, thereby determining the samples of latent features. From these samples, the probability distribution of the latent features can be determined, for example, by maximum likelihood estimation using a predetermined family of probability distributions.

[0041] In the third step, the variability measure of the posterior predictive distribution is provided as uncertainty regarding the sensor signal synthesized based on the noise sample.

[0042] Figure 2 schematically illustrates how the method according to claim 1 is used in a method (1000) for creating or augmenting a dataset (T). A noise sample (z) is derived from a probability distribution (d), which may be a probability distribution used in particular to train a generative machine learning system (61). The noise sample (z) is provided to the generative machine learning system (61) as input. In this embodiment, the generative machine learning system (61) is a stable diffusion model configured to generate an image based on the noise sample (z), but other generative machine learning systems and corresponding modalities of sensor signals are also possible. In this embodiment, a final layer Laplacian approximation is used to perform Bayesian estimation. By using the weights of the generative machine learning system (61) in estimation by Monte Carlo sampling, multiple sensor signals (x1, x2, x) are obtained from the generative machine learning system. M) (i.e., an image) is sampled. The image (x1, x2, x M ) for each image (x1, x2, x M ) are provided as input to a feature extractor (62) in order to extract latent features (e1, e2, e M ). In this embodiment, the feature extractor is provided by the visual model of a CLIP model. Other feature extractor models are also possible.

[0043] Next, the plurality of extracted latent features (e1, e2, e M ) can be used in an uncertainty module (63) that determines uncertainty regarding a sensor signal (i.e., an image) generated from a noise sample (z). In particular, the uncertainty module (63) can estimate a distribution of latent features ((e1, e2, e M ), for example, by using maximum likelihood estimation. In particular, the mean and covariance matrix of a multivariate normal distribution can be estimated using the latent features (e1, e2, e M ). In particular, the estimation can be achieved according to the following formula [[Expression]] , where M is the number of samples used during Monte Carlo sampling, and e m is a latent feature determined for the m-th sample extracted during Monte Carlo sampling, [[Expression]] , z is a noise sample, D is a training data set for a generative machine learning system, and σ 2 is any offset of the covariance matrix.

[0044] In this case, the uncertainty module (63) can determine a measure of the variability of the distribution (also called the posterior predictive distribution) as uncertainty (u). Preferably, the entropy of the posterior predictive distribution can be provided as uncertainty, but other methods such as pairwise distance estimators can also be used to measure the variability of the posterior predictive distribution. The method (1000) up to this step can be considered as one embodiment of the method (900) for determining uncertainty. The method (1000) can then proceed further by determining whether the uncertainty is less than or equal to a predetermined threshold (th). If the uncertainty (u) is less than or equal to the predetermined threshold, the determined images for the noise sample (z), i.e., images generated using the noise sample (z) as input, and images generated using either point estimates of the parameters of the generative machine learning system (61) or the expected value of the posterior predictive distribution of the generative machine learning system (61) itself, are input to the dataset (T), thereby creating and / or augmenting the dataset (T).

[0045] Figure 3 shows an embodiment of a training system (140) that trains a machine learning system (60) using dataset (T) as the training dataset (T). The training dataset (T) consists of multiple input signals (x) used for training the machine learning system (60). i The training dataset (T) includes (i.e., images), and each input signal (x i For the input signal (x i ) corresponds to or input signal (x i A desired output signal (t) characterizes the classification or regression result (i.e., a continuous value) associated with ). i ) includes.

[0046] During training, the training data unit (150) accesses a computer-implemented database (St2), which provides the training dataset (T). The training data unit (150) selects, preferably randomly, at least one input signal (x i ) and the input signal (x i ) corresponds to the desired output signal (t i ) and determine the input signal (x i The input signal (x) is sent to the machine learning system (60). The machine learning system (60) processes the input signal (x). i Based on ) the output signal (y i ) will be decided.

[0047] Desired output signal (t i ) and the determined output signal (y i ) is sent to correction unit (180).

[0048] Next, the correction unit (180) outputs the desired output signal (t i ) and the determined output signal (y i Based on this, a new parameter (Φ') for the machine learning system (60) is determined. For this purpose, the modification unit (180) uses the loss function to determine the desired output signal (t i ) and the determined output signal (y i ) is compared with the determined output signal (y i ) is the desired output signal (t i Determine a first loss value that characterizes the degree of deviation from ). In the given embodiment, a negative log-likelihood function is used as the loss function. Other loss functions are possible in alternative embodiments.

[0049] Furthermore, the determined output signal (y i ) and the desired output signal (t i ) and each include, for example, multiple sub-signals in the form of a tensor, where the desired output signal (t i The sub-signal of ) is determined by the output signal (yi It is assumed that this will correspond to the sub-signals of the input signal (x). For example, a machine learning system (60) is configured for object detection, and the first sub-signal is the input signal (x i The second sub-signal is assumed to characterize the probability of object occurrence for a portion of the determined output signal (y i ) and the desired output signal (t i If the signal includes multiple corresponding sub-signals, preferably a second loss value is determined for each corresponding sub-signal, and the determined second loss values ​​are appropriately combined, for example, by a weighted sum, to form a first loss value.

[0050] The modification unit (180) determines a new parameter (Φ') based on the first loss value. In the given embodiment, this is done using gradient descent, preferably stochastic gradient descent, Adam, or AdamW. In other embodiments, training may also be based on evolutionary algorithms or quadratic methods for training the neural network.

[0051] In other preferred embodiments, the training described is repeated iteratively for a predetermined number of iteration steps, or until the first loss value falls below a predetermined threshold. Alternatively or additionally, training may be terminated when the mean first loss value on the test dataset or validation dataset falls below a predetermined threshold. In at least one of the iterations, a new parameter (Φ') determined in a previous iteration is used as the parameter (Φ) of the machine learning system (60).

[0052] Furthermore, the training system (140) may include at least one processor (145) and at least one machine-readable storage medium (146) containing instructions that, when executed by the processor (145), cause the training system (140) to perform a training method according to one aspect of the present invention.

[0053] Figure 4 shows an embodiment of a control system (40) that uses a machine learning system (60) to control an actuator (10) or a display (10a). The actuator (10) and its environment (20) together are referred to as the actuator system. Preferably, at equal intervals, a sensor (30) senses the status of the actuator system. The sensor (30) may include multiple sensors. Preferably, the sensor (30) is an optical sensor that captures an image of the environment (20). The output signal (S) of the sensor (30) (or, if the sensor (30) includes multiple sensors, the output signal (S) of each sensor) encoding the sensed status is transmitted to the control system (40).

[0054] As a result, the control system (40) receives a stream of sensor signals (S). The control system then calculates a series of control signals (A) depending on the stream of sensor signals (S), and then transmits these control signals to the actuator (10).

[0055] The control system (40) receives a stream of sensor signals (S) from the sensor (30) in an optional receiving unit (50). The receiving unit (50) converts the sensor signals (S) into input signals (x). Alternatively, if the receiving unit (50) is not present, each sensor signal (S) may be directly taken in as an input signal (x). The input signals (x) may be provided, for example, as part of the sensor signals (S). Alternatively, the sensor signals (S) can be processed to generate input signals (x). That is, the input signals (x) are provided according to the sensor signals (S).

[0056] The input signal (x) is then passed to the machine learning system (60).

[0057] The machine learning system (60) is stored in parameter storage (St1) and parameterized by parameters (Φ) provided from there.

[0058] The machine learning system (60) determines an output signal (y) from an input signal (x). The output signal (y) contains information to assign one or more labels to the input signal (x). The output signal (y) is transmitted to an optional conversion unit (80), which converts the output signal (y) into a control signal (A). In this case, the control signal (A) is transmitted to the actuator (10) to control the actuator (10) accordingly. Alternatively, the output signal (y) may be directly taken as the control signal (A).

[0059] The actuator (10) receives a control signal (A), is controlled accordingly, and performs an action corresponding to the control signal (A). In this case, the actuator (10) may include a control logic circuit that converts the control signal (A) used to control the actuator (10) into other control signals.

[0060] In other embodiments, the control system (40) may include a sensor (30). In yet another embodiment, the control system (40) may optionally or additionally include an actuator (10).

[0061] In another embodiment, the control system (40) can be configured to control the display (10a) in place of or in addition to the actuator (10).

[0062] Furthermore, the control system (40) may include at least one processor (45) and at least one machine-readable storage medium (46) that stores instructions for causing the control system (40) to carry out a method according to an aspect of the present invention when executed.

[0063] Figure 5 shows an embodiment in which the control system 40 is used to control a robot that is at least partially autonomous, for example, a vehicle 100 that is at least partially autonomous.

[0064] The sensor (30) may include one or more video sensors and / or one or more radar sensors and / or one or more ultrasonic sensors and / or one or more LiDAR sensors. Some or all of these sensors are, but are not required, preferably mounted on the vehicle (100).

[0065] The machine learning system (60) can be configured to detect objects in the vicinity of at least partially autonomous robots based on an input image (x). The output signal (y) may include information characterizing where the objects are located in the vicinity of the at least partially autonomous robots. In this case, the control signal (A) can be determined according to this information, for example, to avoid collisions with the detected objects.

[0066] The actuator (10) is preferably mounted on a vehicle (100) and may be supplied by the vehicle's (100) brakes, propulsion system, engine, drivetrain, or steering. The control signal (A) can be determined to control the actuator (10) so as to avoid a collision between the vehicle (100) and the detected object. The detected object is also classified according to what the machine learning system (60) determines to be most likely, such as a pedestrian or a tree, and the control signal (A) can be determined depending on this classification.

[0067] Alternatively or additionally, control signal (A) can be used to control the display (10a) so that, for example, an object detected by a machine learning system (60) is displayed. It is also conceivable that control signal (A) controls the display (10a) and generates a warning signal if the vehicle (100) is about to collide with at least one of the detected objects. The warning signal may be an audible and / or tactile signal, such as vibration of the vehicle's steering wheel.

[0068] In other embodiments, the at least partially autonomous robot may be provided by another mobile robot (not shown) that can move, for example, by flying, swimming, submerging, or walking. The mobile robot may be, in particular, an at least partially autonomous lawnmower or an at least partially autonomous cleaning robot. In all of the above embodiments, the control signal (A) can be determined to control the propulsion unit and / or steering and / or brakes of the mobile robot so that the mobile robot can avoid collision with the identified object.

[0069] In other embodiments, a robot that is at least partially autonomous may be provided by a horticultural robot (not shown) using sensors (30), preferably optical sensors, to determine the state of the plants in the environment (20). An actuator (10) can control a nozzle that sprays liquid and / or a cutting device, such as a blade. Depending on the identified species and / or the state of the identified plant, a control signal (A) can be determined so that the actuator (10) sprays the appropriate amount of liquid onto the plant and / or cuts the plant.

[0070] In other embodiments, the robot, at least partially autonomous, may be provided by a household appliance (not shown), such as a washing machine, stove, oven, microwave oven, or dishwasher. A sensor (30), such as an optical sensor, can detect the state of an object being processed by the household appliance. For example, if the household appliance is a washing machine, the sensor (30) can detect the state of the laundry inside the washing machine. In this case, the control signal (A) can be determined depending on the detected material of the laundry.

[0071] Figure 6 shows an embodiment in which a control system (40) is used to control a manufacturing machine (11) (e.g., a punch cutter, cutter, gun drill, or gripper) of a manufacturing system (200) as part of a production line. The manufacturing machine may include a conveying device for moving the manufactured product (12), such as a conveyor belt or an assembly line. The control system (40) controls actuators (10) that control the manufacturing machine (11).

[0072] The sensor (30) may be provided, for example, by an optical sensor that captures the characteristics of a manufactured product (12). Thus, the machine learning system (60) can be understood as an image classifier.

[0073] The machine learning system (60) can determine the position of the manufactured product (12) relative to the conveying device. The actuator (10) can then be controlled depending on the determined position of the manufactured product (12) for subsequent manufacturing steps of the manufactured product (12). For example, the actuator (10) can be controlled to cut the manufactured product at a specific position on the manufactured product itself. Alternatively, the machine learning system (60) can also classify whether the manufactured product is damaged and / or shows defects. In this case, the actuator (10) can be controlled to remove the manufactured product from the conveying device.

[0074] The term "computer" can be understood to encompass any device that processes predetermined computational rules. These computational rules can be in the form of software, hardware, or a combination of software and hardware.

[0075] Generally, a plural can be understood as an indexed form in which each of the elements is assigned a unique index, preferably by assigning a consecutive integer to each element contained within the plural. Preferably, the plural contains N elements, where N is the number of elements represented in the plural, and each element is assigned an integer from 1 to N. It can also be understood that the plural elements are accessible by their respective indexes.

Claims

1. In physical reality, sensor signals (x １ , x ２ , x Ｍ Regarding the probability of observing ), the sensor signal (x) synthesized by the generative machine learning system (61) １ , x ２ , x Ｍ A computer-implemented method (900) for determining the uncertainty (u) of ), - A step (901) to acquire a noise sample (z), - Step (902) of performing Bayesian estimation on the generative machine learning system (61) using the noise sample (z) as input to the generative machine learning system (61), thereby determining the posterior predicted distribution of the sensor signal that would have been synthesized by the generative machine learning system (61) using the noise sample (z) as input to the generative machine learning system (61), ・a step (903) of providing a variability measure of a posterior prediction distribution as uncertainty related to the sensor signal (x １ , x ２ , x Ｍ ) synthesized from the noise sample (z); In a method including, The aforementioned posterior predictive distribution is the sensor signal (x) that would have been generated using the noise sample (z) as input. １ , x ２ , x Ｍ ) Latent features (e １ , e ２ , e Ｍ This characterizes the distribution of ) A method characterized by the following:

2. Determining the aforementioned posterior predictive distribution is, - Using Monte Carlo sampling, multiple synthesized sensor signals (x) are generated from the generative machine learning system (61). １ , x ２ , x Ｍ The steps to derive ) and - The synthesized sensor signal (x １ , x ２ , x Ｍ ) each of the latent features (e １ , e ２ , e Ｍ ) and - The determined latent features (e １ , e ２ , e Ｍ Based on the aforementioned latent features (e １ , e ２ , e Ｍ The steps include determining the distribution of ) and - The aforementioned latent features (e １ , e ２ , e Ｍ The steps include providing the distribution of ) as a posterior predictive distribution, including, The method according to claim 1 (900).

3. The aforementioned latent features (e １ , e ２ , e Ｍ ) is determined by the feature extractor (62), The feature extractor (62) receives the sensor signal as input (x １ , x ２ , x Ｍ ) is received as the sensor signal (x １ , x ２ , x Ｍ ) Latent features (e １ , e ２ , e Ｍ It is configured to provide ) as output, The method according to claim 2 (900).

4. The aforementioned latent features (e １ , e ２ , e Ｍ The distribution of ) is given by the following formula [Math 1] Characterized by, where M is the number of samples used during Monte Carlo sampling, and e ｍ This is a latent feature determined for the m-th sample derived during Monte Carlo sampling, [Math 2] Here, z is a noise sample, D is the training dataset for the generative machine learning system, and σ ２ This is an arbitrary offset of the covariance matrix, The method according to any one of claims 1 to 3 (900).

5. A computer-implemented method (1000) for creating or augmenting a dataset (T) for training and / or testing a machine learning system (60), The above method involves the following steps, namely: - A step of deriving noise samples (z) from the noise distribution (d), - A step of providing the noise sample (z) to the method (900) according to any one of claims 1 to 4, thereby determining the uncertainty, - If the aforementioned uncertainty is below a predetermined threshold (th), - A step of synthesizing a sensor signal from a generative machine learning system (61) used in any one of claims 1 to 4, wherein the sensor signal is synthesized using the noise sample as input to the generative machine learning system (61), - A step of adding the synthesized sensor signals to the dataset (T), thereby creating or enhancing the dataset (T), A method (1000) including the following.

6. A computer-implemented method for training and / or testing a machine learning system (60), - The step of creating the training dataset and / or test dataset (T) described in claim 5, - A step of training and / or testing the machine learning system (60) using the dataset (T), Methods that include...

7. The aforementioned sensor signal (x １ , x ２ , x Ｍ ) is a digital image or audio signal. The method according to any one of claims 1 to 6.

8. The variability of the variability measure of the posterior predictive distribution is the entropy of the posterior predictive distribution, or the variability is determined based on the posterior predictive distribution using a pairwise distance estimator. The method according to any one of claims 1 to 7.

9. A training system (140) configured to carry out the training method described in claim 6.

10. A control system (40) configured to determine a control signal (A) based on the classification of the machine learning system described in claim 6, The control signal (A) is configured to control the actuator (10) and / or the display (10a) of the control system (40).

11. A computer program, wherein the computer program is configured such that, when the computer program is executed by a processor (45, 145), it causes a computer to perform all the steps of the method described in any one of claims 1 to 8.

12. A machine-readable storage medium (46, 146) storing the computer program described in claim 11.

Images