Apparatus and method for training a neural network

Abstract

Provides a method for training neural networks. [Solution] In the training system, the neural network 60 outputs a classification or regression result of sensor measurement values ​​S or provides the probability that sensor measurement values ​​occur. It also outputs predictions of sensor measurement values ​​in order to create a training data set or test data set for training other machine learning systems. The training method includes the steps of: performing singular value decomposition of the weight matrix or weight tensor of the neural network; grouping the singular values ​​and corresponding singular vectors into a plurality of groups based on the magnitude of the singular values; and training the neural network by fitting only the singular values ​​and / or singular vectors from the group corresponding to the smallest singular value or fitting only the singular values ​​and / or singular vectors from a predetermined number of groups corresponding to the smallest singular value.

Description

[Technical Field]

[0001] The present invention relates to a computer-implemented method for training a neural network, a method for using a neural network according to the training method, a training system, a control system, a computer program, and a machine-readable storage medium. [Background technology]

[0002] Conventional technology "PiSSA: Principal Singular Values ​​and Singular Vectors Adaptation of Large Language Models" by Meng et al. (published April 3, 2024, available at https: / / arxiv.org / html / 2404.02948v1) discloses a method for fine-tuning neural networks using singular value decomposition. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Meng et al., “PiSSA: Principal Singular Values ​​and Singular Vectors Adaptation of Large Language Models” (published April 3, 2024, available at https: / / arxiv.org / html / 2404.02948v1) [Overview of the Initiative] [Problems that the invention aims to solve]

[0004] Advantages of the invention Neural networks are used as the backbone for a wide variety of technological applications, including robotics, driver assistance systems, and manufacturing. This invention relates to a method for training neural networks, and more specifically, to a method for efficiently fine-tuning a pre-trained neural network using singular value decomposition (SVD). Large pre-trained neural networks demonstrate remarkable capabilities. However, fitting these large models to specific downstream tasks can be computationally expensive and resource-intensive. Fine-tuning the entire set of parameters often requires considerable computational resources and can lead to overfitting, especially when training data for downstream tasks is limited.

[0005] Conventional and known methods have explored various approaches to address this challenge. Parameter-efficient fine-tuning (PEFT) methods such as LoRA, ETHER, or other techniques aim to reduce the number of trainable parameters by introducing low-rank adjustments or other compact representations. While these methods improve efficiency, they often lack the ability to target specific components within the network's weight matrix, potentially leading to suboptimal performance or instability. Another related technique, PiSSA (Principal Singular Values ​​and Singular Vectors Adaptation), leverages SVD to decompose the weight matrix and fine-tunes only the low-rank approximation based on the largest singular value. This approach prioritizes the most dominant component of the weight matrix while ignoring the potential contribution of smaller singular values ​​that may be important for a particular task or data set. Furthermore, PiSSA and similar methods typically employ a fixed predetermined rank for low-rank approximations and apply a uniform fit to all parameters within that approximation. [Means for solving the problem]

[0006] In contrast, the present invention proposes to decompose the weight matrix using SVD, group the singular values and corresponding singular vectors into a plurality of groups, and optimize only the singular values and / or singular vectors from one or more groups corresponding to the smallest singular values. Advantageously, the inventors have discovered that this approach results in a reduction of catastrophic forgetting when fine-tuning a neural network, especially in contrast to PiSSA. This is because the newly learned content during fine-tuning is "stored" in directions that have stored little information in the space of the weight matrix so far. As a result, due to the reduction of catastrophic forgetting, the fine-tuned neural network can exhibit better performance for both the original task targeted when training this neural network and the new task targeted when training this neural network. Furthermore, since only a subset of the weight matrix of the neural network is adapted, the training of the neural network becomes more resource-efficient compared to standard learning procedures.

[0007] Disclosure of the Invention In a first aspect, the present invention is a computer-implemented method for training a neural network, wherein the neural network is configured to receive an input characterizing at least one sensor measurement, provide an output characterizing the classification and / or regression result of at least one sensor measurement, and / or provide the probability of occurrence of the sensor measurements in a set of sensor measurements, or the neural network is configured to provide an output characterizing the prediction of sensor measurements to create a training data set and / or a test data set for training another machine learning system, and the method for training comprises · performing a singular value decomposition of the weight matrix or weight tensor of the neural network; · grouping the singular values and corresponding singular vectors into a plurality of groups based on the magnitude of the singular values; · training the neural network by adapting singular values and / or singular vectors only from the group corresponding to the smallest singular value, or by adapting singular values and / or singular vectors only from a predetermined number of groups corresponding to the smallest singular value; relates to a method including this.

[0008] The neural network may be provided, in particular, in the form of a pre-trained neural network. In other words, the neural network may be pre-trained. Pre-training is understood, in particular, as the neural network being adapted for a particular task before being used in a training method. By analyzing the weights of the neural network, it is possible to distinguish between a pre-trained neural network and a non-pre-trained neural network. If the weights of the neural network match a general probability distribution, particularly the standard normal distribution, the neural network can be considered non-pre-trained; otherwise, it can be considered pre-trained.

[0009] The neural network may be provided, in particular, in the form of a feed-forward neural network, in particular a convolutional neural network, a multi-layer perceptron, or a transformer.

[0010] The expression "input characterizing sensor measurements" may be understood as an input that includes or consists of values ​​that are the result of sensor measurements. In other words, the input provided to the neural network includes or consists of numerical values ​​obtained from physical reality by the sensor. The sensor may be an optical sensor, in particular a camera, lidar, radar, ultrasonic sensor, or thermal camera. However, it may be any other type of sensor. Sensor measurements obtained from all these sensors may be understood as images, that is, the sensor measurements may be images in particular. When the sensor measurements are images, the neural network may be understood to operate in particular on the pixel values ​​of the image, for example, by accepting the pixel values ​​of the image as input for performing operations on the neural network.

[0011] A neural network may be used as a virtual sensor by taking inputs that include or consist of sensor measurements from various types of sensors and predicting other physical entities that can be derived from the sensor measurements. Sensors that can be considered for such virtual sensor applications include accelerometers, gyroscopes, GPS, inertial measurement units, proximity sensors, rotary encoders, temperature sensors, humidity sensors, pressure sensors, light sensors, air quality sensors, water quality sensors, and / or weather sensors.

[0012] The expression “output characterizing classification and / or regression results” may similarly be understood as the output of a neural network containing or consisting of classification results (e.g., class index, probabilities for multiple classes, etc.) or regression results, i.e., integer or real values. When the input characterizes an optical sensor, the output can, in particular, describe the physical parameters of the scene recorded by the optical sensor, for example, the classification of the presence or absence of objects or other elements in the scene, the performance of object detection (in which case the output characterizes both classification and regression, i.e., the classification of the presence and class of objects, and regression results showing, for example, the position and / or extent of objects relative to bounding boxes), or a dense evaluation of the scene, for example, the performance of semantic segmentation and / or instance segmentation. These outputs can be generated for each of the optical sensors described above.

[0013] The expression "the probability of a sensor measurement occurring within a set of sensor measurements" can be understood as meaning that a neural network is capable of modeling the distribution of sensor measurements. Therefore, in these cases, the neural network can determine a value that indicates the probability of a sensor measurement occurring in the real world. This probability value may be, for example, likelihood, log-likelihood, density value, or simply probability. Typical neural network architectures for providing such values ​​are variational autoencoders, normalized flow models, or diffusion models. Therefore, the neural network may be configured as an encoder for a normalized flow model, a diffusion model, or a variational autoencoder.

[0014] The statement "The neural network is configured to provide an output that characterizes predictions of sensor measurements" may be understood as the neural network being a generative model, such as the generator in a generative adversarial network, or a diffusion-based model, such as StableDiffusion. Predictions regarding sensor measurements may be understood, in particular, as the neural network providing images and including those images in a data set to train and / or test other machine learning systems.

[0015] Generally, the term "sensor measurement value" is understood to include both raw sensor measurement values ​​and sensor measurement values ​​that have already been preprocessed by some form of preprocessing method, such as smoothing or extracting statistical features.

[0016] A neural network includes a weight matrix, preferably multiple weight matrices. The weight matrices are preferably assigned to the layers of the neural network; that is, the weight matrices may belong to the layers of the neural network.

[0017] In the training method, the weight matrix is ​​decomposed using singular value decomposition. Singular value decomposition decomposes a real-valued weight matrix W of shape (m × n) into the following equation: W=UΣV T It is best understood as decomposing according to the following, where U is an m × m orthogonal matrix with the left singular vectors of W as columns, and Σ is a descending order on the main diagonal (σ1 ≥ σ2 ≥ ... ≥ σ r Non-negative singular values ​​(σ) arranged in ) i It is an m × n rectangular diagonal matrix having r, where r is the rank of W and V T This is the transpose of an n×n orthogonal matrix V whose columns are the right singular vectors of W.

[0018] The SVD of a weight tensor can be achieved by stacking the tensor's submatrices into a single matrix, performing SVD on this single matrix, and then reshaping the resulting decomposition into a tensor representation.

[0019] For grouping purposes, take the diagonal of Σ to find the singular value σ i The singular vectors can then be extracted. The singular vectors can then be grouped based on their magnitude, i.e., their absolute values. For example, the absolute values ​​of the singular values ​​can be obtained, and then the list of absolute values ​​can be sorted. The list can then be divided into groups. The division into groups may be carried out according to various specific approaches. For example, the range of absolute values ​​can be divided into a predetermined number of ranges, and the singular values ​​corresponding to the absolute values ​​can be sorted into groups according to the specific range division to which each absolute value belongs. Similarly, the singular values ​​can be grouped such that each group contains the same number of singular values. Alternatively, the grouping of singular values ​​may be based on the exponential intervals of the ranges, or on the percentiles of the absolute values.

[0020] Next, for each group k of singular values, a corresponding set of columns from U and V can be selected based on the singular value assigned to that group k. For example, if the singular values ​​{σ1, σ5, σ7} are assigned to group k, then the corresponding columns u1, u5, u7 from U and V can be selected. T From line

number

number

[0021] Technically, the columns and rows corresponding to the indices of the singular values in the original Σ matrix are selected. This is because the singular values themselves are just scalars.

[0022] Then, for each group k, the matrix U k and Σ k and

Number

Number

[0023] Then, the output of the forward operation of the layer containing the weight matrix can be achieved by multiplying the sum of the input to the layer containing the weight matrix by matrix A k .

[0024] Training a neural network by adapting singular values and / or singular vectors can be understood, for example, as optimizing specific components of the decomposed weight matrix based on the loss value derived from the loss function. More specifically, the training focuses on optimizing the components of the weight matrix corresponding to the smallest singular values. This targeted optimization can be achieved in particular by two methods:

[0025] Method of directly adapting singular values and singular vectors: In this embodiment, the adaptation is performed on the singular values (σ i ) and / or the corresponding singular vectors (columns of U and V TIt acts directly on the row(s). This may include modifying the singular values ​​themselves according to backpropagation. This direct fitting provides fine-grained control over the individual components of the weight matrix, allowing for very detailed tuning of what is actually being optimized. By fitting only the singular values ​​and singular vectors associated with the smallest singular values, while leaving the others fixed, the network can refine less dominant features without significantly altering the core representation encoded by the larger singular values.

[0026] Reconstructed group matrix (A k A method for fitting the group matrix A: This alternative embodiment uses the singular values ​​and the vectors associated with the smallest singular values ​​corresponding to these singular values ​​to form one or more group matrices A k The matrix A is then reconstructed. Next, instead of directly fitting the singular values ​​and singular vectors, these matrices A k Fitting is performed on the matrix A. This may involve applying standard optimization algorithms (such as SGD or Adam) with a reduced learning rate, or utilizing parameter-efficient fine-tuning (PEFT) techniques, such as LoRA or AdaLoRA. This optimization is controlled by a loss function calculated based on the neural network's output and the true desired output for each training input. k The advantage of this approach is that it allows us to leverage existing optimization and fine-tuning methods while still benefiting from selective fitting based on singular value grouping. By restricting the fit to the group corresponding to the smallest singular value—that is, to either a single group or a predetermined number of the smallest singular value groups—it becomes possible to selectively refine specific aspects of the model's representation without affecting the most dominant features.

[0027] Both of these approaches target fitting the component associated with the smallest singular value, i.e., modifications during the training process. Advantageously, this strategy may be particularly beneficial in preventing catastrophic forgetting because it allows the network to learn new information while retaining knowledge encoded by the more dominant and larger singular value components, primarily by tuning the less important directions in the weight matrix space. This can be done by directly fitting the singular values / singular vectors or by using matrix A k The specific choice of which groups to fit can depend on factors such as the desired level of control, computational efficiency, and compatibility with existing training frameworks. In either case, it is possible to select either the group corresponding to the smallest absolute singularity, or a predetermined number of groups corresponding to the smallest singularity. In the former case, the group corresponding to the smallest singularity is selected. In the latter case, a predetermined number of groups is selected, which may be dynamically adjusted during training or be user-defined and fixed. This choice can depend on the underlying data set to which the neural network is fitted. For example, a simpler data set may require fitting fewer groups of singularities because fitting more groups may require redundant parameters during training.

[0028] The selection of the number of groups, that is, the selection of which groups should be fitted in this method, can be understood as a hyperparameter of the training method, and the precise number of groups to be fitted can be achieved using a standard hyperparameter optimization strategy such as grid search.

[0029] Compared to known methods such as PiSSA that freeze groups with small singular values, the present invention proposes to advantageously do the opposite in order to combat catastrophic forgetting.

[0030] Furthermore, by optimizing only the group corresponding to the smallest singular value, it becomes possible to significantly reduce the total number of weights that need to be fitted during neural network training. This results in improved efficiency during training and reduced computational complexity.

[0031] In a preferred embodiment, singular values ​​and / or singular vectors can be fitted from only a predetermined number of groups corresponding to the smallest singular value, a different non-zero learning rate can be used for each group, and / or a different optimization algorithm can be selected for each group.

[0032] Advantageously, the inventors discovered that by assigning different learning rates to each selected group, they could finely control how different features were fitted. Groups corresponding to more important features (e.g., groups with larger singular values, even within the set of "smallest" groups) could be assigned a smaller learning rate, thereby retaining previously learned information that may still be important, even if not the most relevant. Conversely, groups representing less important or task-specific features (e.g., groups with smaller singular values) could be assigned a larger learning rate, enabling faster fitting to new tasks.

[0033] The advantage of using multiple different optimization algorithms is that the optimization can be tailored to diverse features. Different singular value groups typically capture different types of features or representations. These groups may also have different statistical properties or exhibit different learning dynamics. By using different optimization algorithms for each group, the training process can be tailored to the specific characteristics of each group. For example, one group may benefit from an adaptive optimization algorithm like Adam, while another group may be better suited to momentum-assisted SGD. The authors found that this flexibility allows each group to be optimized with the most appropriate algorithm, resulting in faster convergence and better overall performance.

[0034] In the same or other preferred embodiments, fitting singular values ​​and / or singular vectors may include applying a parameter-efficient fine-tuning (PEFT) method to the group.

[0035] The PEFT method aims to reduce the number of trainable parameters compared to full fine-tuning. Incorporating PEFT into this SVD-based fitting scheme further improves parameter efficiency and reduces computational complexity. This is particularly beneficial when dealing with large pre-trained models where fine-tuning the entire weight matrix can be computationally enormous. Applying PEFT to selected singular value groups further reduces the number of trainable parameters, resulting in faster training, reduced memory requirements, and a lower risk of overfitting.

[0036] Possible embodiments of the PEFT method that can be applied are LoRA, AdaLoRA, DyLoRA, or BiFit.

[0037] In preferred or other embodiments, grouping singular values ​​may include dividing the singular values ​​into a predetermined number of groups based on an equal-width range of the magnitudes of the singular values.

[0038] This can be understood as collecting the largest and smallest absolute singular values ​​from Σ and dividing the range of absolute singular values ​​into a predetermined number of ranges, which can be considered as hyperparameters of the training method. The magnitude of a singular value and the absolute singular value can be considered to refer to the same thing.

[0039] Advantageously, the authors found that uniform binning ensures uniform coverage across the entire range of singularities. This avoids potential bias towards specific regions of the spectrum and allows for a balanced fit between different levels of feature importance. This uniform coverage may be particularly useful when the distribution of singularities is relatively uniform, or when there is a lack of prior knowledge about the relative importance of different singularities.

[0040] In some embodiments, grouping singular values ​​involves assigning each singular value to a group based on a percentile threshold of the singular value's magnitude.

[0041] This can be understood as defining a predetermined number of percentile thresholds, such as the 10th percentile and the 20th percentile, taking into account the distribution of the magnitudes of singular values. Then, based on these percentile thresholds, singular values ​​are assigned to groups. A singular value is assigned to a particular group if its magnitude falls between two predetermined percentile thresholds. For example, two percentile thresholds, namely the 10th and 20th percentiles, can be defined, and all singular values ​​that are greater than the singular value at the 10th percentile and less than the singular value at the 20th percentile are inserted into one group. The predefined percentile thresholds can be understood as hyperparameters of the method.

[0042] Advantageously, the authors found that percentile-based grouping ensures a balanced distribution of singular values ​​across multiple distinct groups, independent of any specific range of magnitude. This can be particularly beneficial when the distribution of singular values ​​is biased or overly concentrated in a particular region. For example, if the majority of singular values ​​are very small, equal-width binning may result in multiple groups containing very few singular values, whereas percentile-based grouping ensures that each group contains a similar proportion of the total number of singular values. This balanced distribution can lead to more stable and effective training, especially when combined with group-specific fitting strategies.

[0043] In preferred or other embodiments, the neural network is trained iteratively, and the singular value decomposition step and the grouping step are performed after each iteration or after a predetermined number of iterations.

[0044] This can be understood as training a neural network over multiple iterations or multiple training steps, in which, after each iteration or after a predetermined number of iterations, a step of performing singular value decomposition, a step of grouping singular values, and a step of processing the singular values ​​or A k The step of fitting the matrix is ​​performed. After a predetermined number of iterations, or after each iteration, the updated weight matrix of the neural network is subjected to a new singular value decomposition, and new groups are defined based on the singular values ​​of this decomposition. Then, the singular values ​​and singular vectors of the updated weight matrix are grouped into new groups, meaning that singular values ​​and singular vectors that belonged to one particular group in a previous iteration may now belong to another group after the regrouping, or that there are now entirely new singular values. For example, the singular value decomposition step and the grouping step can be repeated after each training epoch.

[0045] Advantageously, the authors discovered that this iterative approach allows for the adaptation of singular value groupings to accommodate changes in the importance of multiple different components within the weight matrix during the training process. As the network learns and the weight matrix is ​​updated, the relative importance of multiple different singular values ​​may fluctuate. By periodically recalculating the SVD and regrouping the singular values, the adaptation process can dynamically focus on the most relevant parts of the weight matrix at each stage of training. This can lead to more effective and efficient fine-tuning, particularly in complex learning scenarios where the importance of multiple different features may evolve over time. Furthermore, dynamic regrouping can help move away from local optima and improve the overall performance of the fine-tuned network.

[0046] In another embodiment, the present invention relates to a computer-implemented method for determining control signals for controlling actuators and / or displays of a technical system using a neural network, wherein the method is The steps include: obtaining a neural network by any one embodiment of the training method, The steps include: providing a neural network with an input that characterizes at least one sensor measurement; The steps include: using a neural network to determine an output that characterizes the classification and / or regression result of at least one sensor measurement, and / or determining the probability that a sensor measurement occurs in a set of sensor measurements; • A step of determining a control signal based on the output, This includes methods.

[0047] Acquiring a neural network according to a training method may be understood, in particular, as each step of the respective embodiment of the training method being part of a method for using the neural network. Alternatively, acquiring a neural network according to a training method may be understood as acquiring a neural network from a source that has been trained according to the above embodiments of the training method, for example, by downloading the neural network from the Internet, or by receiving the neural network via some form of digital data transmission, for example, by a computer-readable storage device.

[0048] If the output of the neural network is a classification and / or regression result, the control signal can be automatically selected, for example, according to the classification. If the output characterizes probabilities, the output can be used, for example, for anomaly detection in the provided input. For example, if the characterized probability is low, the control signal can be selected to switch the automatic operation of the actuator to a human operator.

[0049] In another embodiment, the present invention relates to a training system configured to carry out a training method.

[0050] In another embodiment, the present invention relates to a control system configured to carry out a method for determining a control signal.

[0051] In other embodiments, the present invention relates to a computer program that, when executed by a processor, causes a computer to perform all the steps of any one of the methods presented above.

[0052] In another embodiment, the present invention relates to a machine-readable storage medium on which a computer program is stored.

[0053] Embodiments of the present invention will be described in more detail with reference to the following drawings. [Brief explanation of the drawing]

[0054] [Figure 1] This figure shows a training system for training neural networks. [Figure 2] This figure shows a control system that uses a neural network to control an actuator or display. [Figure 3] This figure shows a control system for controlling at least a semi-autonomous robot. [Figure 4] This is a diagram showing a control system for controlling manufacturing machinery. [Modes for carrying out the invention]

[0055] Description of the Embodiment Figure 1 shows an embodiment of a training system (140) for training a neural network (60) using a training data set (T). The training data set (T) consists of multiple inputs (x) used to train the neural network (60). i) includes these inputs (x i Each of these characterizes one or more sensor measurements as described above. The training data set (T) is composed of each input (x i ) for each desired output (t i ) further includes this desired output (t i ) is input (x i ) corresponds to input (x i Characterize the classification and / or regression results of ).

[0056] The neural network (60) receives input (x i ) corresponds to the output (y i The neural network (60) includes a parameter (Φ) that defines how it is mapped to the neural network (60). The parameter includes the weight matrix or weight tensor of the neural network (60). In some embodiments, a singular value decomposition (SVD) can be performed on one or more weight matrices or weight tensors of the neural network (60) before training. The singular values ​​resulting from the SVD are then grouped into groups. This grouping can be based on a range of equal widths, a percentile threshold, or other appropriate criteria. The number of groups and the specific grouping method may be predefined or may be dynamically adjusted.

[0057] For training, the training data unit (150) accesses a computer-implemented database (St2), which provides the training data set (T). The training data unit (150) takes at least one input (x) from the training data set (T). i ) and this input (x i ) corresponds to the desired output (t i ) and are preferably determined randomly, and this input (x i The input (x) is sent to the neural network (60). The neural network (60) then processes the input (x). i ) and the output (y based on the current parameters of the neural network (60), i.e., the weight matrix.i ) will be decided.

[0058] Desired output (t i ) and the determined output (y i ) and are sent to the modification unit (180). The modification unit (180) then determines new parameters (Φ') for the neural network (60) based on these values. Specifically, the modification unit (180) determines the desired output (t i ) and the determined output (y i ) and can be compared using a loss function. The loss function is the determined output (y i ) is the desired output (t i Determine the loss value that characterizes how much it deviates from ). Various loss functions are possible, such as negative log-likelihood, mean squared error, or other loss functions suitable for the specific task. Determine the output (y i ) and the desired output (t i ) can, for example, contain multiple subcomponents of a tensor form, in which case the desired output (t i The subcomponents of ) are determined by the output (y i This corresponds to a subcomponent of (x). In such cases, a loss value can be determined for each corresponding subcomponent and combined, for example, using a weighted sum, to form a single overall loss value. A particular embodiment of a subcomponent is a tensor used in supervised training for object detection, for example, with input (x i This is a tensor used as a corresponding image, describing the existence or class of the objects within the image, another tensor describing the position of each object, and another tensor describing the size dimensions of each object.

[0059] The modification unit (180) can fit the parameters of the neural network (60) based on the calculated loss values ​​and the decomposed weight matrix or weight tensor. This fitting may include modifying only the smallest group of singular values, or the singular values ​​and / or singular vectors corresponding to a predetermined number of groups corresponding to the smallest singular values. The fitting may be performed directly on the singular values ​​and singular vectors, or on a reconstructed matrix (A) formed from the grouped singular values ​​and singular vectors. k ) may be performed on ). During fitting, different non-zero learning rates and / or different optimization algorithms may be used for each group. Suitable optimization algorithms include, but are not limited to, stochastic gradient descent, Adam, AdamW, or second-order methods. Parameter-efficient fine-tuning (PEFT) methods such as LoRA, adapters, or other techniques may also be applied during the fitting process.

[0060] The above training process can be repeated iteratively. In an alternative embodiment, the singular value decomposition step and the grouping step can be performed after each iteration or after a predetermined number of iterations, thereby allowing for dynamic adjustment of groups during training. Training can be continued for a predetermined number of iterations until the loss value falls below a predetermined threshold or until the performance metric for the validation set reaches a desired level.

[0061] The training system (140) may include at least one processor (145) and at least one machine-readable storage medium (146), the at least one machine-readable storage medium (146) including instructions that, when executed by the processor (145), cause the training system (140) to perform the training method described above.

[0062] Input (x iA specific modality of one or more sensor measurements characterized by ) is selected according to the final task of the neural network (60). For example, appropriate sensor measurements are taken from the real world and appropriate desired output (t i By annotating with ), a data set (T) for supervised learning can be formed.

[0063] In another embodiment of the training system (140), the neural network (60) takes the input (x) of the data set as input. i The neural network (60) may be configured to model the probability distribution of ). In these embodiments, the neural network (60) may be a normalized flow, a variational autoencoder encoder, or a diffusion model.

[0064] Figure 2 shows an embodiment of a control system (40) that controls an actuator (10) in the actuator environment (20). The actuator (10) and the environment (20) of the actuator (10) together are referred to as the actuator system. Preferably, at equal intervals, a sensor (30) senses the state 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) that encodes the sensed state (or, if the sensor (30) includes multiple sensors, the output signal (S) of each of these sensors) is transmitted to the control system (40).

[0065] As a result, the control system (40) receives a stream of sensor signals (S). The control system (40) then calculates a series of control signals (A) depending on the stream of sensor signals (S), and these control signals (A) are then transmitted to the actuator (10).

[0066] 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 inputs (x). Alternatively, if the receiving unit (50) is not provided, each sensor signal (S) may be directly obtained as input (x). The input (x) may be provided, for example, as an excerpt from the sensor signals (S), or as a portion of an image. Alternatively, the input (x) may be generated by processing the sensor signals (S). In other words, the input (x) is provided according to the sensor signals (S).

[0067] Next, the input (x) is transmitted to the neural network (60).

[0068] The neural network (60) is configured to accept an input (x) and is trained according to the training method described above using appropriate inputs and outputs in a supervised manner. The neural network (60) is parameterized by a parameter (Φ), which is stored in a parameter memory (St1) and provided by the parameter memory (St1).

[0069] The neural network (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). The control signal (A) is then transmitted to the actuator (10) to control the actuator (10) accordingly. Alternatively, the output signal (y) may be directly obtained as the control signal (A).

[0070] The actuator (10) receives a control signal (A), is controlled accordingly, and performs an action corresponding to the control signal (A). The actuator (10) may include control logic that converts the control signal (A) into a further control signal, in which case the actuator (10) is controlled using this further control signal.

[0071] In further 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).

[0072] In yet another embodiment, the control system (40) may control the display (10a) in place of or in addition to the actuator (10).

[0073] Furthermore, the control system (40) may include at least one processor (45) and at least one machine-readable storage medium (46), the at least one machine-readable storage medium (46) storing instructions that, when executed, cause the control system (40) to carry out a method according to an aspect of the present invention.

[0074] Figure 3 shows an embodiment in which a control system (40) controls at least a semi-autonomous robot (100), for example, at least a semi-autonomous vehicle (100). The sensors (30) may comprise a variety of sensor modalities, including one or more video sensors, radar sensors, ultrasonic sensors, and / or LiDAR sensors. These sensors may be integrated into the vehicle (100) or located externally. The input (x) may be, for example, a combination of images, point clouds, or sensor data from multiple sensors (30). The neural network (60) trained according to the method described above can then function as an image classifier, an object detector, or a more general-purpose sensor fusion network.

[0075] The neural network (60) may be configured to detect and classify objects in the vicinity of the robot based on an input (x). The output (y) may characterize the location, type (e.g., pedestrian, vehicle, tree) and other relevant attributes of the detected object. Based on this information, control signals (A) can then be determined to control the robot's actuators (10) and ensure safe interaction of the robot with its environment. In the case of a vehicle (100), if a pedestrian is detected crossing the road, the control signal (A) may initiate braking or steering maneuvers to avoid a collision.

[0076] Alternatively or additionally, the output (y) can characterize the probability that the input (x) occurs in the real world. If this probability is below a predetermined threshold, this may be understood as an input (x) that characterizes an anomaly in what the neural network (60) has seen during training, i.e., an anomaly in the data from the data set (T). If such an anomaly input (x) is detected, a control signal (A) can be selected to hand over the automatic operation of the actuator (10) to a human operator or driver.

[0077] Actuators (10) that can be incorporated into the robot (100) may be any components that play a role in controlling the robot's movements or providing feedback to an operator or driver. Such components may include brakes, propulsion systems, engines, drivetrains, steering mechanisms, and display interfaces (10a). Control signals (A) can instruct these actuators to perform appropriate actions, such as adjusting speed, changing direction, or displaying warnings. For example, in the case of a vehicle (100), control signals (A) may activate the vehicle's braking system if a collision is imminent, or display a warning signal on the dashboard (10a) if the vehicle is deviating from its lane. The warning signal may be visual, auditory (e.g., a warning sound), or tactile (e.g., a vibrating steering wheel).

[0078] At least semi-autonomous robots are not limited to vehicles. At least semi-autonomous robots can encompass a variety of mobile robots, including aerial drones, underwater vehicles, or walking robots. For example, an autonomous lawnmower can use its system to avoid obstacles such as trees or rocks, while a cleaning robot can use its system to move around a room and avoid furniture. In these cases, the sensor (30) may be a camera, LiDAR, or proximity sensor, and the actuator (10) controls the robot's movement and actions.

[0079] Furthermore, the present invention can be applied to other robotic systems, such as horticultural robots. In this scenario, a sensor (30), optionally an optical sensor or a multispectral camera, can be used to assess the condition of the plants. In this case, the output signal (y) can provide information about the plant species, health status, or the need for watering or pruning. An actuator (10), such as a liquid spray nozzle or a cutting device, is controlled by a control signal (A) to perform appropriate horticultural tasks, such as watering a particular plant or pruning overgrown branches.

[0080] Figure 4 shows an embodiment in which a control system (40) is used to control a manufacturing machine (11) of a manufacturing system (200), such as a punch cutter, cutter, gun drill, or gripper, 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 assembly line. The control system (40) controls an actuator (10), which in turn controls the manufacturing machine (11).

[0081] The sensor (30) may be provided, for example, by an optical sensor that captures the characteristics of a manufactured product (12).

[0082] The neural network (60) can determine the position of the manufactured product (12) relative to the conveying device. Then, for subsequent manufacturing processes of the manufactured product (12), the actuator (10) can be controlled depending on the determined position of the manufactured product (12). For example, the actuator (10) can be controlled to cut the manufactured product at a specific point on the manufactured product itself. Alternatively, the neural network (60) can be assumed to classify whether the manufactured product is damaged and / or shows defects. In that case, the actuator (10) can be controlled to remove the manufactured product from the conveying device.

[0083] Generally, the term "computer" may be understood to encompass any device for processing a given set of computational rules. These computational rules may be in the form of software, hardware, or a combination of software and hardware.

[0084] Generally, plurals are often understood to be subscripted, that is, each element of a plural is assigned a unique subscript, preferably by assigning consecutive integers to the multiple elements contained in the plural. Preferably, if a plural contains N elements, and N is the number of elements in that plural, these elements are assigned integers from 1 to N. It may also be understood that each element contained in a plural is accessible through its subscript.

Claims

1. A computer-implemented method for training a neural network (60), The neural network is configured to accept an input (x) characterizing at least one sensor measurement (S), to provide an output (y) characterizing the classification and / or regression results of the at least one sensor measurement (S), and / or to provide the probability that the sensor measurement (S) occurs in a set of sensor measurements, or The neural network (60) is configured to provide an output (y) that characterizes predictions of sensor measurements in order to create a training data set and / or test data set for training other machine learning systems. The aforementioned method for training is - A step of performing singular value decomposition of the weight matrix or weight tensor of the neural network (60), - A step of grouping the singular values ​​and corresponding singular vectors into multiple groups based on the magnitude of the singular values, - A step of training the neural network (60) by fitting singular values ​​and / or singular vectors only from the group corresponding to the smallest singular value, or by fitting singular values ​​and / or singular vectors only from a predetermined number of groups corresponding to the smallest singular value, Methods that include...

2. The singular value and / or singular vector are fitted from only a predetermined number of groups corresponding to the smallest singular value. A different non-zero learning rate is used for each group, and / or a different optimization algorithm is selected for each group. The method according to claim 1.

3. Fitting the singular values ​​and / or singular vectors includes applying a parameter-efficient fine-tuning (PEFT) method to the group. The method according to claim 1 or 2.

4. Grouping the aforementioned singular values ​​includes dividing the singular values ​​into a predetermined number of groups based on an equal-width range of the magnitudes of the singular values. The method according to any one of claims 1 to 3.

5. Grouping the singular values ​​includes assigning the singular values ​​to each group based on percentile thresholds of the magnitude of the singular values. The method according to any one of claims 1 to 3.

6. The neural network (60) is trained iteratively, and the singular value decomposition step and the grouping step are performed after each iteration or after a predetermined number of iterations. The method according to any one of claims 1 to 5.

7. A computer-implemented method for determining a control signal (A) for controlling an actuator (10) and / or a display (10a) of a technical system, using a neural network (60) according to any one of claims 1 to 6, The aforementioned method, - A step of obtaining a neural network (60) according to any one of claims 1 to 6, - Providing the neural network (60) with an input (x) that characterizes at least one sensor measurement (S), - The neural network (60) determines an output (y) that characterizes the classification and / or regression result of the at least one sensor measurement (S), and / or determines the probability that the sensor measurement (S) occurs in the set of sensor measurement values. - A step of determining the control signal (A) based on the output (y), Methods that include...

8. A training system (140) configured to carry out the training method described in any one of claims 1 to 6.

9. A control system (40) configured to carry out the method described in claim 7.

10. A computer program configured, when executed by a processor (45, 145), to cause a computer to perform all steps of the method according to any one of claims 1 to 7.

11. A machine-readable storage medium (46, 146) in which the computer program according to claim 10 is stored.

Images