Method and / or apparatus for architecture search

Abstract

To provide a method for an architecture search of architecture of a neural one-shot network in order to solve a multi-task problem depending on at least one piece of target hardware.SOLUTION: The method comprises: selecting, by random sampling, an architecture parameter of a neural one-shot network for each piece of target hardware depending on initialized task weights and hardware embedding of a hypernetwork having the hardware embedding to target hardware; calculating a gradient by a prescribed loss function with respect to the hypernetwork; updating a network parameter on the basis of multiple gradient descents and / or update of task weights with respect to each piece of target hardware; calculating gradients by the prescribed loss function with respect to the neural one-shot network; aggregating the calculated gradients; and updating a network weight on the basis of the aggregated gradients.SELECTED DRAWING: None

Description

[Technical Field]

[0001] The present invention relates to a method and / or apparatus for performing architecture search for neural network architectures. [Background technology]

[0002] Conventional technology Neural networks are increasingly being used to evaluate measurands in vehicle control devices and other embedded systems. Neural networks are characterized by a significant generalization power compared to other evaluation techniques. After sufficient training, neural networks can appropriately evaluate previously unobserved situations and generate drive control signals that result in appropriate vehicle or other system responses.

[0003] This performance value means that neural networks place relatively high demands on the hardware platform on which they are implemented. Typically, a graphics processor (GPU) with a large memory facility is required. Apart from the increased cost of the hardware platform, space is often at a premium in control devices for vehicles and other embedded systems. The maximum current consumption is also limited by the energy source used, e.g., the vehicle's on-board power grid or battery, and / or by the maximum allowable heat dissipation. Therefore, German Patent Application No. 10 2019 202 816 A1 discloses a method for training a neural network, in which less relevant neurons and connections between neurons are completely deactivated.

[0004] Furthermore, the applicant has disclosed a method for multi-criteria architecture search, which addresses the search for neural one-shot architectures for multi-objective optimization. Here, a technique is proposed that allows for the optimization of multi-objective optimization problems without the need to consider the scaling of the various objectives. Furthermore, the approach provides not only a single architecture for a specific weighting of the optimization objectives, but also a set of optimal architectures for different weightings of the objectives. To achieve this, the method includes the following steps: performing random sampling weightings of the objectives from a Dirichlet distribution; drawing a random sample of architectures from a distribution that depends on the objective weightings; updating the weights to minimize a loss or cost function, where the parameters of the one-shot network are adjusted; and adjusting the architecture parameters to minimize the cost function. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] German Patent Application Publication No. 102019202816 Summary of the Invention [Problem to be solved by the invention]

[0006] It is an object of the present invention to provide an improved method and / or an improved apparatus for training neural one-shot networks. [Means for solving the problem]

[0007] The above problem is solved by a method having the features of claim 1. The above problem is solved by an apparatus having the features of claim 10.

[0008] Disclosure of the Invention According to a first aspect, there is provided a method for architecture search of an architecture of a neural one-shot network f(w,α) for solving a multitasking problem depending on at least one target hardware D. The method comprises the following steps: network parameters θ, task weights or partial objective weights λ, and at least one hardware embedding e for at least one objective hardware D hw providing a hypernetwork g having providing a neural one-shot network f(w,α) with network weights w and architecture parameters α; initializing task weights λ; In particular, the hypernetwork g is used to initialize task weights λ and at least one hardware embedding e hw selecting, in particular by random sampling, an architecture parameter α for each target hardware d of at least one target hardware D depending on Loss function

number

number

number

number

number

[0009] According to a second aspect, there is provided an apparatus for performing architecture search for an architecture of a neural one-shot network f(w,α) for solving a multitask problem depending on at least one target hardware D. The apparatus includes the following steps: network parameters θ, task weights λ, and at least one hardware embedding e for at least one target hardware D, hw providing a hypernetwork g having providing a neural one-shot network f(w,α) with network weights w and architecture parameters α; initializing task weights λ; The initialized task weights λ and at least one hardware embedding e hw selecting, in particular by random sampling, an architecture parameter α for each target hardware d of at least one target hardware D depending on Loss function

number

number

number

number

number

[0010] Embodiments described for the method also apply correspondingly to the apparatus, and vice versa.

[0011] It will be understood that the steps and further optional steps of the present invention do not necessarily have to be performed in the order shown, or may be performed in other orders, and additional intermediate steps may be provided, and individual steps may further include one or more sub-steps without departing from the scope of the method of the present invention.

[0012] The neural one-shot network f(w,α) is trained based on the network parameters θ and the network weights w, i.e., by updating based on the updated network weights w and the updated network parameters θ.

[0013] Particularly preferably, one or more meta-learned predictors for a hardware objective function are prepared herein. The hardware objective function is preferably optimized with respect to latency. The predictors are preferably trained before the architecture search and are then used to calculate the objective function for the hardware. The predictors preferably use hardware embedding as well as the function g so that they can be standardized for different hardware. The predictors are preferably trained in a simple procedure, in particular by sampling the hardware and architecture at each training step. The task weights are preferably not trainable but are drawn from a probability distribution at each training iteration. The method and / or apparatus address architecture search for neural network architectures for multitask and / or multicriteria and / or hardware-sensitive approaches.

[0014] In the context of neural architecture search, one-shot models preferably refer to a technique in which individual neural networks are trained to simultaneously represent and evaluate multiple candidate architectures. Instead of training and evaluating each architecture individually, one-shot models enable efficient search in a large search space by partitioning the weights and parameters for each different architecture. One-shot models consist of a super-network that represents a larger network containing all possible architectures. Such a super-network includes various architectural components, such as convolutional layers, pooling layers, and skip connections, that can be selectively activated or deactivated for each candidate architecture. By using different activation patterns, the super-network can simulate various architectures within a single model. During the training process, the one-shot models are trained toward a proxy task, such as image classification, where a combination of architecture parameters and weight distributions is used. The model learns to adjust its weights and parameters based on the performance of each different architecture in the super-network. This enables efficient exploration of the search space because the model can simultaneously evaluate multiple architectures and update its parameters accordingly. Once training is complete, the one-shot model can select the best architecture based on its learned weights and parameters. This selected architecture can then be trained anew from scratch or fine-tuned to achieve optimal performance in the target application. Overall, the one-shot model approach to neural architecture search enables efficient exploration of the search space by training a single model to simultaneously represent and evaluate multiple candidate architectures.

[0015] A hypernetwork is a type of neural network used to generate or influence the parameters of another neural network. A hypernetwork often includes a main network, often referred to as a "base network," and side networks or "hypernetworks." The base network may be, for example, a convolutional neural network (CNN) or a recurrent neural network (RNN), or, as used herein, a one-shot network trained for multitasking, such as image classification or speech processing. The hypernetwork is used to generate or manipulate weights or other parameters of the base network or one-shot model. This provides a flexible means of adjusting the structure and / or parameters of the one-shot network without having to retrain the entire network. This can be particularly beneficial when tuning or fine-tuning the network to a task or data without requiring the full training time and resources for new training.

[0016] It is understood that the at least one image data file may be a video data file, and the same applies to the video data file to be generated in this application, where a text-in-video generation algorithm is preferably used.

[0017] The method and / or the device can be applied, for example, in the general technical framework of face recognition, and / or in the technical framework of vehicle assistance systems, and / or in the technical framework of autonomous driving, and / or in the technical framework of computer vision, and / or in the technical framework of quality monitoring of manufactured parts in automatic optical inspection, and / or in the technical framework of other technical fields, such as the evaluation and / or classification of image data.

[0018] Particularly preferably, the method and / or device can be used in the analysis of data acquired by at least one (image) sensor. The at least one sensor can, for example, determine environmental measurements in the form of a sensor signal. Such a sensor signal can, for example, be present as, inter alia, a digital image and / or a video. The sensor can, for example, be a camera and / or a LiDAR sensor and / or an ultrasonic sensor. That is to say, the present invention can be used in particular to solve multitasking problems, for example for image analysis and / or video analysis and / or audio analysis following detection and for classification or segmentation of the detected image data therein. In general, this may be sensor data of all kinds (including radar, LiDAR, ultrasonic, etc.) and / or data from multiple sensors (e.g., multiple cameras) and / or data from a combination of sensors (e.g., camera, radar and LiDAR).

[0019] The present invention can be used, in particular, to classify sensor data and / or to identify the presence of objects in the sensor data and / or to perform semantic segmentation of the sensor data, e.g., with respect to traffic signs, road surfaces, pedestrians, and / or vehicles. Here, outliers can also be calculated during the classification and / or segmentation of input data, e.g., sensor data. The method and / or device can also be used to determine a continuous value or multiple continuous values, i.e., to perform a regression analysis, e.g., on distance, speed, and / or acceleration. The method and / or device can also be used to track elements, e.g., objects, in the input data. This can be done, for example, based on low-level features (e.g., edges or pixel attributes in an image).

[0020] The method and / or the device can also be used to calculate control signals for controlling technical systems, for example computer-controlled machines, such as robotic systems, vehicles, domestic appliances, power tools, manufacturing machines, personal assistants or access control systems. This may be an information transmission system, such as a surveillance system or a medical (imaging) system. Here, the method and / or the device analyze input data (e.g., scalar time series), in particular from sensors, in particular from any sensor, and based on this analysis the technical system can be operated accordingly.

[0021] Herein, the method, particularly the hypernetwork, is extended by adding hardware embeddings as additional inputs to the network, so that the architecture of the one-shot network can be sampled for various hardware. Furthermore, an optional memory database and attention mechanism can be added to the neural network.

[0022] Furthermore, the training strategy is tailored to take into account multiple target hardware devices. First, the updates for θ for multiple target hardware platforms or devices are preferably computed using multiple iterations of gradient descent. Furthermore, the updates for the one-shot model weights w are preferably computed based on multiple architectures, i.e., based on various hardware embeddings considered by the hypernetwork g.

[0023] That is, meta-learned latent predictors are prepared and trained by random sampling of architectures and target devices during each training step. These predictors are preferably used additively to the hypernetwork. They are preferably trained separately, prior to the actual architecture search, i.e., prior to the specifically described training procedure. During the architecture search, the predictors remain fixed and unchanged.

[0024] This method allows for training and / or testing of architectures within one-shot networks against a variety of hardware available both during training and for unknown hardware embedding.

[0025] The training method preferably optimizes the hypernetwork to predict architectures for multiple target hardware and also to enable prediction for unknown hardware with zero spots. The previously proposed training method improves the predictor's performance and, if necessary, reduces resource consumption.

[0026] The techniques herein optimize neural network architectures with respect to multiple objectives (also called multi-objective optimization problems), which may include the neural network's performance in a specific task (e.g., object recognition from an image) as well as the efficiency of the architecture on a target hardware device. The methods can be used in any system that uses neural networks.

[0027] The method and / or apparatus may also utilize a multi-objective search that relates model performance (e.g., accuracy, etc.) and model hardware performance (e.g., latency, FLOPS, power consumption, memory utilization). Thus, for example, an additional loss term may be added to the loss function, where the additional loss term predicts hardware performance depending on the architecture parameters of the one-shot model and / or hypermodel.

[0028] Here, a one-shot model f(w,α) is used that represents the superposition of multiple neural network architectures in a single network, where w is the weight of the individual architectures of the network, and α is an architecture parameter that determines which of the architectures embedded in the one-shot model is active. In this case, α can be the output of a function or probability distribution g with parameter θ: α=g(θ), where g is preferably represented by a neural network. The objective of the one-shot architecture search is preferably optimization θ to minimize the loss,

number

number

[0029] where w is an optimal layer parameter that may possibly depend on λ. The layer parameter w can preferably be optimized by optimizing other data-dependent losses. The loss L(α, w) is, for example, the loss on the validation data,

number

number

number

number

[0030] In a preferred embodiment, the initialization involves random sampling weighting of task weights λ from a Dirichlet distribution Dir(β) based on a cardinality hyperparameter β.

[0031] The Dirichlet distribution is a family of multivariate probability distributions used in statistics to model proportional or probability distributions categorically. The distribution is characterized by a vector of positive real numbers used as concentration parameters. These parameters affect how concentrated the weights are in the direction of the corners or edges of the simplex (in a multidimensional space whose coordinates sum to 1) represented by the Dirichlet distribution. In the context of random sampling weights, initializing task weights from a Dirichlet distribution means that the initial distribution of weights among different tasks is determined and controlled by the properties of the Dirichlet distribution. A relatively high value results in a more uniform distribution of weights, implying less initialization, while a relatively low value encourages a stronger concentration on fewer tasks, which can potentially be useful when some tasks are considered more important than others or when it is desired to optimize model performance for specific tasks.

[0032] In a preferred embodiment, the method comprises determining an architecture parameter α and at least one hardware embedded parameter e hw In a preferred embodiment, the hardware predictor is trained by drawing, in each gradient update step, the architecture and target hardware from, for example, a probability distribution or a Dirichlet distribution, in particular by random sampling.

[0033] In a preferred embodiment, the concentration hyperparameter β is initialized as a vector of ones.

[0034] Initializing β as a vector of ones results in a uniform Dirichlet distribution. In a uniform Dirichlet distribution, all possible distributions of categories (or tasks) have equal probability. This approach implies no a priori assumptions about category importance or distribution and therefore represents a "non-informative" starting condition. Using a vector of ones as the initialization value is easy to implement and understand. It provides a clear, neutral starting point for modeling, which can be particularly useful when there is no specific a priori information about the data to be modeled. While initialization begins with a vector of ones, the parameter β and its associated task weights λ can be adjusted during the training process. This allows the model to learn data based on observed performance or data characteristics and optimize the weighting accordingly. In multi-task learning scenarios or other contexts where a Dirichlet distribution is used to control the initialization or adjustment of model parameters, deciding on a vector of ones as the initialization for β can influence the learning process by providing a neutral and independent basis for exploring various weights.

[0035] In a preferred embodiment, the network parameters θ are updated by applying a Frank-Wolfe routine or by averaging the gradients of multiple gradient descents.

[0036] Task weight updating is a step that optimizes the weighting of various tasks and improves the overall performance of the model. Two distinct approaches to this process are distinguished: the Frank-Wolfe algorithm and gradient averaging over multiple iterations. The Frank-Wolfe algorithm, also known as conditioned gradient algorithms, is an optimization algorithm used for convex optimization problems. Unlike other gradient descent methods that take steps in the direction of the negative gradient of the objective function, the Frank-Wolfe algorithm searches for a solution within a convex solution space by solving a linear subproblem at each iteration. This approach is particularly useful for optimization problems where the solution must remain within a specific region or space, such as in the case of task weight optimization that must satisfy certain constraints. Gradient averaging over multiple iterations is typically applied in multitask learning scenarios, where multiple tasks are learned simultaneously. This approach considers the gradients of all tasks and averages them to find a common direction in which the weights should be updated. These techniques aim to find a compromise between various tasks and determine an update direction that is advantageous overall for all tasks. The Frank-Wolfe routine is concerned with how to aggregate gradients with respect to the θ parameters. The θ parameters are then updated using any optimizer (gradient descent + momentum, Adam, ...) based on the descent direction calculated by the Frank-Wolfe routine.

[0037] In a preferred embodiment, the aggregation is performed by averaging.

number

[0038] Other approaches to aggregation may also be advantageous.

[0039] In a preferred embodiment, the selection of the architecture parameter α includes the output of a function or probability distribution and / or a hypernetwork selection. In a preferred embodiment, the architecture parameter α is provided in a differentiable manner using the function g (or hypernetwork). To use a discrete architecture parameter α, this can be done by using a differentiable approximation such as Straight Through Gumbel-Softmax.

[0040] When the selection of the architecture parameter α is based on the output of a function, this means that a specific mathematical or algorithmic function is used to determine the optimal or likely value for α. The function here can be based on heuristics, optimization methods, or model performance metrics on a validation data set. The function evaluates potential architectures or structures and selects the one that best separates according to a specific criterion. When the selection is based on a probability distribution, this means a probabilistic approach to determining α. Instead of directly determining a fixed value, the architecture parameter is treated as a random variable following a specific distribution. This approach can be applied, for example, in the framework of a Bayesian optimization strategy, where the probability distribution represents the uncertainty about the model performance over the space of possible architectures. The selection is then performed in a manner that iteratively reduces this uncertainty by selecting new data points (model architectures) that maximize the expected information gain.

[0041] In a preferred aspect, an inference method is proposed, which comprises applying a neural one-shot network f(w,α) trained according to this aspect of the method to solve a multitask problem, in particular a classification task and / or a segmentation task.

[0042] Multi-task problems involve scenarios in which a model must simultaneously solve multiple tasks. This can improve the efficiency of the learning process and the generative capabilities of the model, since it learns to extract features that can be used across various tasks. Classification tasks aim to classify inputs into predefined categories. Segmentation tasks aim to identify and classify specific regions within input data, often in the context of images, for example, to identify and determine the boundaries of various objects within an image. The proposed inference method utilizes trained one-shot networks to efficiently solve such multi-task problems. By applying one-shot networks, the learned weights and architecture parameters are used to address new instances of the task with minimal training data.

[0043] In a preferred aspect, a computer program is also claimed, comprising a program code for, when the program is executed on a computer, performing at least part of the method according to the invention in one of its embodiments. In other words, a computer program (product) is proposed, which comprises instructions for causing a computer to perform one or more steps of the method according to the invention in one of its embodiments, when the program is executed by the computer.

[0044] In a preferred aspect, a computer-readable data carrier is proposed, which comprises a program code of a computer program for performing at least part of the method according to the invention in one of its embodiments, when the computer program is run on a computer. In other words, the invention relates to a computer-readable (storage) medium comprising instructions which, when run by a computer, cause the computer to perform a number of methods / steps of the method according to the invention in one of its embodiments.

[0045] The described configurations and developments can be combined with one another in any desired manner.

[0046] Other conceivable configurations, developments and realizations of the invention also include not explicitly mentioned combinations of the features of the invention described above or below with respect to the embodiments.

[0047] The accompanying drawings are provided for a further understanding of embodiments of the present invention, each illustrating an embodiment and used in conjunction with the explanation of the basic principles and concepts of the present invention.

[0048] Other embodiments and many of the advantages discussed above will become apparent in light of the drawings, in which: It should be noted that elements shown in the drawings are not necessarily drawn to scale relative to each other. [Brief explanation of the drawings]

[0049] [Figure 1] 1 is a schematic flow chart illustrating the method. [Figure 2] 1 is a schematic block circuit diagram of the method according to the invention;

[0050] In the various figures of the drawings, identical or functionally equivalent elements, components or elements are designated by the same reference numerals unless otherwise stated. DETAILED DESCRIPTION OF THE INVENTION

[0051] FIG. 1 shows a schematic flow chart of a method for training a neural one-shot network f(w,α) for solving multitasking problems relying on at least one objective hardware D.

[0052] The method can be performed, in any embodiment, at least in part by apparatus 100, which for this purpose may include multiple components not shown in detail, such as one or more supply devices and / or at least one rating calculation device. It will be understood that the supply devices may be configured together with the rating calculation device or may be distinct from the rating calculation device. The apparatus may further include a memory device and / or an output device and / or a display device and / or an input device.

[0053] The computer-implemented method includes at least the following steps.

[0054] In step S1, the network parameters θ, task weights λ, and at least one hardware embedding e for at least one target hardware D are determined. hw A hypernetwork g having

[0055] In step S2, a neural one-shot network f(w,α) with network weights w and architecture parameters α is prepared.

[0056] In step S3, the task weights λ are initialized by randomly sampling the task weights λ from a Dirichlet distribution Dir(β) based on a cardinality hyperparameter β, where the cardinality hyperparameter β is initialized as a vector of 1.

[0057] In step S4, the architecture parameter α is selected, in particular by random sampling, based on the initialized task weight λ and at least one hardware embedding e for each target hardware d of at least one target hardware D. hw The selection of the architecture parameter α is dependent on the output of a function or probability distribution. The selected architecture parameter α is preferably differentiable.

[0058] In step S5, the loss function

number

number

[0059] In step S6, for each target hardware d of at least one target hardware D, multiple gradient descent g θ Here, the calculation is done by applying the Frank-Wolfe routine or by averaging the gradients of multiple gradient descents.

[0060] In step S7, the network parameters θ are updated using multiple iterations of gradient descent g θ and / or based on the task weight λ.

[0061] In step S8, the loss function

number

number

[0062] In step S9, the calculated gradient

number

number

[0063] In step S10, the network weights w are updated based on the aggregated gradients.

[0064] FIG. 2 shows a block diagram of one embodiment of the method.

[0065] The neural one-shot network 200 is initialized with the parameters of a hypernetwork 202. The hypernetwork 202 may be a meta-network. The hypernetwork 202 has network parameters θ. The hypernetwork 202 is configured with task weights λ 204 and at least one hardware embedded e hw 206 as input. The architecture parameter α is calculated by random sampling using the initialized task weights λ 204 and at least one hardware embedded parameter e hw Depending on 206, at least one destination hardware D can be selected for each destination hardware d, which is indicated in FIG. 2 by reference numeral 208.

[0066] The output of the hypernetwork 202 is also available as an input to a predictor 210, which in turn receives the hardware embedding 206 as an input. The predictor 210 may be a multilayer perceptron (MLP) or a graph convolutional network (GCN). The predictor 210 may be pre-trained and frozen in this pre-trained state. The output of the predictor 210 may be used as input for a loss function 212, which in turn is the output of a one-shot network 200 formed based on at least one architecture parameter α. The results of the loss function are plotted on a graph, which maps error 214 (horizontal axis) on latency 216 (vertical axis). Here, latency is just one example of a hardware metric. For example, energy consumption may be an additional hardware metric.

[0067] By way of example, the method can be again illustrated by the following mathematical formulation:

number

Claims

1. A method for performing architecture search for an architecture of a neural one-shot network f(w, α) for solving a multitask problem depending on at least one target hardware D, the method comprising the following steps: - at least one hardware embedding e for network parameters θ, task weights λ and at least one target hardware D ｈｗ a step (S1) of preparing a hypernetwork g having - preparing (S2) a neural one-shot network f(w,α) with network weights w and architecture parameters α; - a step (S3) of initializing said task weights λ; - initialized task weights λ and at least one hardware embedding e ｈｗ a step (S4) of selecting, in particular by random sampling, an architecture parameter α for each target hardware d of at least one target hardware D depending on - loss function [Equation 1] With respect to the network parameter θ, the gradient [Equation 2] A step (S5) of calculating - Multiple gradient descents g for each target hardware d of at least one target hardware D θ and calculating - the multiple gradient descents g θ and / or based on updating the task weights λ; - loss function [Equation 3] With respect to the network weights w, the gradient [Equation 4] A step (S8) of calculating - the calculated gradient [Equation 5] a step (S9) of aggregating - updating the network weights w based on the aggregated gradients (S10); A method comprising:

2. The method (100) of claim 1 , wherein the initialization comprises randomly sampling weighting task weights λ from a Dirichlet distribution Dir(β) based on a cardinality hyperparameter β.

3. The method (100) of claim 2 , wherein the cardinality hyperparameter β is initialized as a vector of ones.

4. 4. The method according to claim 1, wherein the updating of the network parameters θ is performed by applying a Frank-Wolfe routine or by averaging the gradients of multiple gradient descents.

5. The aggregation is average formation [Equation 6] The method according to any one of claims 1 to 4, wherein the method is carried out by

6. The method of claim 1 , wherein the selection of the architecture parameter α comprises the output of a function or probability distribution and / or selection by the hypernetwork.

7. 7. A method of inference comprising applying a neural one-shot network f(w,α) trained according to the method of any one of claims 1 to 6 to solve a multitask problem, in particular a classification task and / or a segmentation task.

8. A computer program comprising a program code for performing at least part of the method according to any one of claims 1 to 7, when the computer program is run on a computer.

9. A computer readable data carrier comprising program code of a computer program for performing at least part of the method according to any one of claims 1 to 7 when the computer program is run on a computer.

10. An apparatus (100) for performing architecture search for an architecture of a neural one-shot network f(w, α) for solving a multitask problem depending on at least one target hardware D, the apparatus (100) comprising the following steps: - at least one hardware embedding e for network parameters θ, task weights λ and at least one target hardware D ｈｗ providing a hypernetwork g having - preparing a neural one-shot network f(w,α) with network weights w and architecture parameters α; - initializing the task weights λ; the initialized task weights λ and the at least one hardware embedded e ｈｗ selecting, in particular by random sampling, an architecture parameter α for each target hardware d of at least one target hardware D depending on - loss function [Equation 7] With respect to the network parameter θ, the gradient [Equation 8] and calculating - performing a plurality of gradient descents g for each target hardware d of said at least one target hardware D; θ and calculating - the multiple gradient descents g θ and / or based on updating the task weights λ; - loss function [Equation 9] With respect to the network weights w, the gradient [Equation 10] and calculating - the calculated gradient [0011] aggregating the - updating the network weights w based on the aggregated gradients; An apparatus (100) comprising a rating calculation device configured to perform:

Images