DE102017128082A1 - Meta-architecture design for a CNN network - Google Patents

Abstract

The present invention relates to a method for designing and training a convolutional neural network (CNN) meta-architecture (50) comprising the steps of constructing a base network with hyperparameters determined by training data, constructing a meta-architecture, formalized extensions (40, 40), 42, 44, 46), performing network marginalization and other meta-architecture optimization techniques, marginalizing for other meta-architecture variables including a brute-force search for variable size configurations, refining a network topology, and baseline Hyperparameters by performing a batch update on the training data over multiple iterations. The present invention also relates to a use of the above method. The present invention further relates to a drive assisting system for a motor vehicle, wherein the drive assisting system is configured to execute the above method. The present invention further relates to a non-transitory computer readable medium having instructions stored thereon that, when executed by a processor, cause a driving support system to perform the above method. The present invention also relates to a motor vehicle having a driving support system with a data processing device, wherein the commands stored on the non-transitory computer-readable medium are executed by the data processing device of the driving support system.

Description

The present invention relates to a method for designing and training a Convolutional Neural Network (CNN) meta-architecture.

Also, the present invention relates to a use of the above method in a driving support system of a motor vehicle.

Furthermore, the present invention relates to a driving support system for a motor vehicle, wherein the driving support system is configured to execute the above method.

The present invention also relates to a non-transitory computer-readable medium having instructions stored thereon that, when executed by a processor, cause a driving support system to perform the above method.

Furthermore, the present invention relates to a motor vehicle having a drive assisting system with a data processing device, wherein the commands stored on the above non-transitory computer readable medium are executed by the data processing device of the drive assisting system.

Driving assistance systems, such as driver assistance systems, are one of the fastest growing segments in automotive electronics, and there is a need for improved methods and systems for e.g. autonomous driving in different driving situations, e.g. in automated parking, driving in traffic congestion situations, driving on a highway, and many others that are executed using various driver assistance systems or generally by driving assistance systems.

The driving assistance systems, including the driver assistance systems, are systems that have been developed to automate, adapt and improve vehicle systems in terms of safety and enhanced driving. Security features are designed to prevent collisions and accidents by providing technologies that warn the driver of potential problems or avoid collisions by implementing security measures and taking control of the vehicle. In autonomous vehicles, the driving assistance systems provide inputs for performing control of the vehicle. Adaptive features can automate the lighting, provide adaptive cruise control, automate the braking process, integrate traffic alerts, communicate with smartphones, e.g. Warn the driver about other cars or various dangers, keep the vehicle in the correct lane, or indicate what is in blind spots. Driving assistance systems incorporating the aforementioned driver assistance systems are often based on inputs from multiple data sources, particularly environmental sensors such as automotive installed cameras, radar sensors, LiDAR, ultrasonic sensors, and others. The driving assistance systems often have complex algorithms, e.g. for image acquisition, image processing or feature recognition in the sensor information. Sensor information processed by neural networks has recently been integrated into the processing of such inputs of sensor data in driver assistance systems or generally in driving support systems.

Convolutional neural networks (CNNs) are a breakthrough technology known to the public under the general term "deep learning". In particular, CNNs have led to rapid technology leaps in various computer vision tasks, especially for object recognition. One standard method of using CNNs is sliding-object object detection for detecting pedestrians, cyclists, and / or vehicles.

Convolutional neural networks are already a common method for solving various computer vision tasks, such as image classification, object detection, semantic segmentation, image retrieval, image tracking, text recognition, stereo matching, and many others. However, convolutional neural networks are computationally intensive, so a major challenge in using CNNs is providing efficient network designs with well-balanced cost criteria.

There are very few guides or theories regarding the abundance of design options and hyperparameter settings of CNNs. Therefore, many researchers and users still use the trial-and-error approach to setting up CNNs. This often results in a copy of previously used CNN architectures, so that established CNN types are often persisted for a long time. Deep Learning (DL) is a fast-moving area of research. Most research is fundamentally different, with the different ideas developed in different architectures not being formalized due to a lack of theoretical background.

Therefore, it is difficult to combine ideas from the best networks from the perspective of application development. In addition, active research in the area of deep learning focuses on image recognition problems. Efficiency is typically not a design criterion in academic research, so much of the relevant networks are very large, with hundreds of layers and an ensemble of multiple networks. This is not applicable in an automotive application, especially for embedded use.

In this regard, the US 2017/293837 A1 a computer-implemented method for training a deep neural network to detect traffic scenes from multimodal sensors and knowledge data. The traffic scenes may be, for example, information related to other vehicles driving in the vicinity of a vehicle or parking lot information indicating whether or not there is a parking space in the vicinity of the vehicle. The computer-implemented method includes receiving data from the multimodal sensors and the knowledge data and extracting feature maps from the multimodal sensors and the knowledge data using a road user extractor to generate a first data set, using a static object extractor to generate a second data set and use of an extractor for additional information. The computer-implemented method further comprises training the deep neural network with training data to detect the traffic scene from a vehicle's viewpoint.

It is an object of the present invention to provide a method for designing and training a convolutional neural network (CNN) meta-architecture, a use of the above method in a driving support system of a motor vehicle, a driving support system for a motor vehicle implementing the above method non-transitory computer-readable medium having instructions stored thereon for carrying out the above method, and a motor vehicle having a driving support system with data processing means, wherein the instructions stored on the above non-transitory computer-readable medium are executed by the data processing means of the driving support system, thereby providing improved means a convolutional neural network for use with various types of training data.

The solution to this problem is provided by the independent claims. Advantageous embodiments are specified in the subclaims.

In particular, the present invention provides a method for designing and training a Convolutional Neural Network (CNN) meta-architecture, comprising the steps of constructing a base network with hyperparameters determined from training data, constructing a meta-architecture, the formalized extensions Performing a network marginalization and other meta-architecture optimization techniques, marginalization for other meta-architecture variables, including a brute-force search for variable size configurations, and refining a network topology and basic hyperparameters by performing a Batch update regarding the training data over several iterations.

According to the invention, a use of the above method is also specified in a driving assistance system of a motor vehicle.

Further, according to the present invention, there is provided a drive assisting system for a motor vehicle, wherein the drive assisting system is configured to execute the above method.

Further, according to the present invention, there is provided a non-transitory computer-readable medium having instructions stored thereon that, when executed by a processor, cause a driving support system to perform the above method.

Further, according to the present invention, there is provided a motor vehicle having a driving support system with a data processing device, wherein the commands stored in the above non-transitory computer-readable medium are executed by the data processing device of the driving support system.

The basic idea of the invention is therefore to specify a meta-architecture design for a network which has a framework of various possible architectures and which is designed based on various conventional network design aspects. Therefore, design aspects of different networks can be combined, such as Fractals of Fractals, Residuals of Residuals, Cross-Layer Filters, Wider Filters, and Cross-Channel Filters. A training algorithm based on meta-learning makes it possible to learn a specific optimal topology of the convolutional neural network. Such networks are easily scalable as the deep Learning progresses quickly. This also allows an efficient bottom-up design for hardware rather than customizing designs. In addition, the best network topology learning mechanism can use automatically provided training data. Overall, a scalable multi-level design for CNN architectures is provided by splitting various aspects into levels of abstraction that can be designed independently.

A simple CNN architecture is a linear cascade of convolutional blocks, typically known as Fully Convolutional Networks (FCN). A standard design is to continually reduce feature resolution and increase the number of channels. In this architecture, the image width and height are powers of two (e.g., 640x480), so that continuous down-sampling can be performed by a factor of two. Design limitations also include eliminating alternative pixels, thereby increasing efficiency. Design limitations also include channel upsampling by a factor of two. This creates a symmetrical design. However, upsampling does not have to be done on all layers. It may be limited to several layers to increase nonlinearity. Yet another design limitation includes e.g. five levels downsampling when the altitude deteriorates to H >> 5. In addition, a 2D structure is useful for a decoder and can not be flattened. The architecture is open for adding more layers with the same scale. This is preferably determined empirically. Preferably, at least ten layers are provided in an encoder and at least two layers are provided in a decoder (Deconv layers).

The step of constructing a meta-architecture with formalized extensions can provide a framework of proposed meta-architecture with proposed design constructs. Preferably, simple modules including 5x5 or 3x3 convolution are used along with optional pooling and nonlinearity. A 3x3 convolution is a popular choice because of the reduced number of calculations. It also has an optional stride. We further suggest choosing ReLU or Cross Channel Maxout, which generalizes ReLU and Leaky ReLU. Cross-channel can also be used as an output channel reduction mechanism. The resolution of the weights plays a key role in the complexity of the calculation and the memory bandwidth. Typically, 32-bit floating point is used. However, as many empirical evidence shows, 8-bit floating point may suffice without significantly decreasing accuracy. Some layers can optionally work with a 16-bit floating-point resolution if a larger dynamic range is needed. The deployment is preferably used for the decoding section. There are three options available: Upconv, Unpooling and Deconv.

The choice of Deconv makes it possible to model a convolution block via a transposed convolution. A suggested selection of parameters will serve as starting values for the evolutionary optimization method. Such a module parameterization may have a convolution kernel size of 3x3 or 5x5, convolution type Conv or Deconv, data resolution of 8 bits or 16 bits and a step size of one or two, nonlinearity is ReLU or cross channel maxout, pooling may be for a maximum kernel size of 2x2, and / or the number of input / output channels may be 4/8.

The architecture may also include non-convolutional blocks. Such non-convolutional blocks may e.g. a regression / soft max layer, a fully connected layer, i. a 1x1 convolution, a GAP layer that provides an accurate and efficient alternative, a 1x1 cross-channel filter, RNNs (recurrent neural networks) / LSTM (long-short-term memory) networks, and a flow / Depth geometry optimizer.

The step of performing network marginalization and other optimization techniques with respect to the meta-architecture is performed because the complexity of the meta-architecture can be very extensive. This concept can be commonly referred to as meta-architecture marginalization learning.

On the basis of the above-described facts, the dimensionality has been restricted so much that summation nodes are retained. Therefore, especially for variable size networks, it becomes difficult to model and optimize the CNN. This problem can be solved with the proposed meta-architecture marginalization learning.

In the final refinement step, optimization of the entire network topology and basic hyperparameters is achieved.

According to a modified embodiment of the invention, the step of constructing a base network with hyperparameters determined by training data includes initializing various design options and modeling the design as a hyperparameters search problem to design options using Bayesian Refine optimization algorithm. Some of the design options include, for example fractals of fractals, residuals of residuals, cross-layer filters, wider filters, and cross-channel filters, as discussed above. Hyperparameters may include, for example, parameters such as the learning rate, the number of time periods, a batch size, an activation function, the number of hidden layers and units, weight initialization, a dropout regularization technique, and a raster search or a randomized search.

According to a modified embodiment of the invention, the step of constructing a meta-architecture containing formalized extensions comprises transforming critical network design aspects into a learnable format, wherein the network topology can be learned from training data. Therefore, training data can already be directly applied to learn the network structure.

According to a modified embodiment of the invention, the step of transforming critical network design aspects into a learnable format in which the network topology of training data can be learned, starting with the base network and optimizing the network topology using Bayesian optimization.

According to a modified embodiment of the invention, the step of constructing a meta-architecture containing formalized extensions comprises applying formalized extensions from a group comprising a broader filterbank, cross-channel filters, cross-layer filters, split branching and summation Joining and handling the scaling of objects.

In this regard, some networks that use a broader filter bank have shown improved performance. For example, it has been found that 16 broader layers can achieve similar performance compared to 1000 layers. This can have training benefits of concurrency for training and / or conclusions. It can also enable feature-dependent gating.

In addition, cross-channel filters can model functional dependencies across channels by marginalizing a 3D folding block. Inception has used it successfully and Xception has optimized it.

Cross-layer connections provide an exponential ensemble in terms of function. Resnet-Skip connections were the easiest way to do this. Then DenseNet has brought it to extremes, with all layers having connections from their previous ones. Highway Net used learnable skip connections of data to provide cross-layer filters.

With regard to split branching and summation joining, it should be noted that there is a popular trend towards heavily used summation nodes. It was introduced by Resnet for skip connections. It is much more efficient than daisy-chain connections, which increases feature dimensionality. It is easily possible to use DMA_3D in a CNN HW block. Another use is in ResNeXt, which spans more parallel paths, and PolyNet generalizes and achieves efficiency.

Handling the scaling of objects has become important as it turned out that CNNs are not scale invariant, as theoretical findings have shown. There is enough empirical evidence to illustrate the explicit handling of scaling. This is important in analog-to-digital conversion, especially for handling small objects. Dilated convolutions are popular for handling multi-scales and can be mapped as multi-scale images at various increments. The preferred design option is the multi-scale feature mapping used in SSDs incorporating multi-scale feature mappings and MS-CNN unified multi-scale.

According to a modified embodiment of the invention, the step of performing network marginalization and other meta-architecture optimization techniques comprises a first step of freezing a first set of aspects and optimizing a second, different set of aspects and a second step of freezing the second one Set of aspects and optimizing the first set of aspects.

According to a modified embodiment of the invention, the step of transforming critical network design aspects into a learnable format comprises sequentially iterating over various combinations of aspects. Therefore, for simplicity, marginalization can be performed with respect to all other meta-architecture variables. In addition, a brute force search for variable size configurations can be performed.

According to a modified embodiment of the invention, the step of refining a network topology and basic hyperparameters comprises identifying a divergence and stopping the iterations in case of a detected divergence. In case of divergence, if the iterations stopped maintaining a previous optimal network.

These and other aspects of the invention will be apparent from and elucidated with reference to the embodiments described below. Individual features disclosed in the embodiments may, alone or in combination, form an aspect of the present invention. Features of the various embodiments may be transferred from one embodiment to another embodiment.

• 1 eine perspektivische Ansicht einer Szene, die auf eine Objekterkennung unter Verwendung eines konvolutionellen neuronalen Netzwerks gemäß einer ersten Ausführungsform angewendet wird;
• 2 ein Ablaufdiagramm eines Verfahrens zum Designen und Trainieren einer Meta-Architektur eines konvolutionellen neuronalen Netzwerks (CNN) zum Bereitstellen des zur Szenenerkennung verwendeten konvolutionellen neuronalen Netzwerks gemäß der ersten Ausführungsform;
• 3 eine schematische Ansicht eines Fully Convolutional Network (FCN), das eine lineare Kaskade von mehreren Faltungsblöcken aufweist;
• 4 eine schematische Ansicht des Fully Convolutional Network (FCN) von 3, das eine lineare Kaskade von mehreren Faltungsblöcken aufweist, die die Merkmalsauflösung fortlaufend reduzieren, mit einer zusätzlichen Eingabeschicht und einer Ausgabeschicht;
• 5 eine schematische Ansicht eines Gerüsts einer vorgeschlagenen Meta-Architektur mit vorgeschlagenen Designkonstrukten, die das FCN von 3 und Module und Verbindungen zum/vom FCN enthalten, gemäß der ersten Ausführungsform;
• 6 eine schematische Ansicht eines einzelnen Moduls mit vier Eingangskanälen und acht Ausgangskanälen;
• 7 eine schematische Ansicht zweier Darstellungen einer breiteren Filterbank als eine erste formalisierte Erweiterung;
• 8 eine schematische Ansicht zweier Darstellungen eines Cross-Channel-Filters als eine zweite formalisierte Erweiterung;
• 9 eine schematische Ansicht einer Darstellung eines Cross-Layer-Filters als eine dritte formalisierte Erweiterung;
• 10 eine schematische Ansicht dreier Darstellungen eines Cross-Layer-Filters als eine dritte formalisierte Erweiterung;
• 11 eine schematische Ansicht dreier Darstellungen der Handhabung einer Skalierung von Objekten als eine vierte formalisierte Erweiterung; und
• 12 eine schematische Ansicht einer Anwendung des designten und trainierten konvolutionellen neuronalen Netzwerks.

Show it:

• 1 a perspective view of a scene that is applied to an object recognition using a convolutional neural network according to a first embodiment;
• 2 3 is a flowchart of a method for designing and training a convolutional neural network (CNN) meta-architecture for providing the convolutional neural network used for scene detection according to the first embodiment;
• 3 a schematic view of a Fully Convolutional Network (FCN), which has a linear cascade of multiple convolution blocks;
• 4 a schematic view of the Fully Convolutional Network (FCN) of 3 comprising a linear cascade of multiple convolution blocks that progressively reduce feature resolution, with an additional input layer and an output layer;
• 5 a schematic view of a framework of a proposed meta-architecture with proposed design constructs, the FCN of 3 and modules and connections to / from the FCN according to the first embodiment;
• 6 a schematic view of a single module with four input channels and eight output channels;
• 7 a schematic view of two representations of a wider filter bank as a first formalized extension;
• 8th a schematic view of two representations of a cross-channel filter as a second formalized extension;
• 9 a schematic view of a representation of a cross-layer filter as a third formalized extension;
• 10 a schematic view of three representations of a cross-layer filter as a third formalized extension;
• 11 a schematic view of three representations of handling a scaling of objects as a fourth formalized extension; and
• 12 a schematic view of an application of the designed and trained convolutional neural network.

1 shows a scene 10 Related to Object Recognition Using a Convolutional Neural Network (CNN) 50 is applied according to a first embodiment. As in 1 can be seen, includes the scene 10 Two people 12 , a vehicle 14 , a parking lot 16 and a street 18 as recorded characteristics.

The preceding scene 10 is evaluated in a motor vehicle having a driving support system with a camera, not shown in the figures. To recognize the scene 10 The driving support system uses a convolutional neural network 50 ,

2 FIG. 3 shows a flowchart of a method for designing and training a meta-architecture of a convolutional neural network (CNN) 50 to the CNN 50 to provide for the above vehicle. The CNN 50 is exemplary in 12 shown.

The procedure begins with step S100 which relates to constructing a base network with hyperparameters determined by training data. Therefore, various design options are initialized and the construction of the base network is modeled as a hyper-parameter search problem to refine design options using the Bayesian optimization algorithm. Some of the design options include fractal fractals, residuals residuals, cross-layer filters, wider filters, and cross-channel filters. For example, hyperparameters may include parameters including a learning rate, the number of time periods, a batch size, an activation function, the number of hidden layers and units, a weight initialization, a dropout regularization technique, and a raster search or a randomized search.

A Simple CNN Architecture is a Fully Convolutional Network (FCN) 20 that eg in 3 is shown. The Fully Convolutional Network 20 indicates a linear cascade of convolution blocks 22 on. Each of the folding blocks 22 has its own set of parameters, including as a first parameter a convolutional kernel size, which in the Embodiment of 3 3x3 or 5x5, for example, a downsampling number given by "/ 2", and as a last parameter a number of channels that are in the in 3 example is in the range of 8 to 128. Such a standard design of the FCN 20 continuously redefines the feature resolution and increases the number of channels, as in the respective views of 4 is shown. In this architecture, the image width and height are powers of 2, eg 640x480, for a channel upsampling by a factor 2 , creating a symmetrical design of the FCN 20 is produced. The architecture shown is open to adding more layers with the same scale. In this example, there are ten layers of differently configured convolution blocks 22 with an additional input layer 34 and an output layer 36 provided as in 4 is shown.

step S110 refers to the construction of a meta-architecture with formalized extensions 40 . 42 . 44 . 46 , Accordingly, a scaffold 24 a proposed meta-architecture with proposed design constructs, as with reference to 5 is apparent. The scaffolding 24 the meta-architecture has an FCN 20 with additional modules 26 which in this example have a 5x5 convolution. They are learnable connections 28 provided to the additional modules 26 with the FCN 20 connect to. The learnable connections 28 are set up the additional modules 26 with different folding blocks 22 at different layers of the FCN 20 to connect as respective input and output. More learnable connections 28 connect folding blocks 22 different layers of the FCN 20 , As in 5 is shown, are the additional modules 26 through the learnable connections 28 connected with each other. 5 shows an arrangement of four additional modules 26 which are arranged in two layers. The learnable connections 28 connect two of the additional modules 26 that directly with the same folding block 22 are connected as an input and form a first layer. Further, two of the additional modules 26 forming a second layer, directly with the same folding block 22 connected to their output this folding block 22 supply. The additional modules 26 the first layer are complete with the additional modules 26 connected to the second layer.

The modules 26 can also have an optional step size with ReLU units or cross-channel maxout. The resolution of the weights can be 32-bit floating point. However, as many empirical evidence shows, in an alternative embodiment 16 Bit or 8 bit floating point selected. Unfolding is used for the decoding section, which is achieved by Upconv, Unpooling or Deconv. In this embodiment, Deconv is chosen. Further, the modules may have a one or two step size, a pooling that is maximum for a kernel size of 2x2, and / or a number of input / output channels 30 . 32 from 4/8. A single module 26 is in 6 shown.

The architecture may also include non-convolutional blocks. Such non-convolutional blocks may e.g. a regression / soft max layer, a fully connected layer, i. a 1x1 convolution, a GAP layer that provides an accurate and efficient alternative, a 1x1 cross-channel filter, RNNs (recurrent neural networks) / LSTM (long-short-term memory) networks, and a flow / Depth geometry optimizer.

There can be formalized extensions 40 . 42 . 44 . 46 be used from a group that has a wider filter bank 40 , Cross-channel filter 42 , Cross-layer filter 44 , a handling of the scaling of objects 46 and having split branching and summation joining.

Some networks have a wider filter bank 40 use have shown improved performance. For example, it has been found that sixteen broader layers can achieve similar performance compared to 1000 layers. Two representations of a wider filter bank 40 are exemplary in 7 shown.

Cross-channel filter 42 can functional dependencies on channels by marginalizing a 3D folding block 22 model. The cross-channel filter 42 have already been successfully used by Inception and optimized by Xception. Two representations of cross-channel filters 42 are exemplary in 8th shown.

Cross-layer filter 44 allow cross-layer connections that provide an exponential ensemble in terms of function. Resnet-skip connections are an easy way to implement cross-layer filters 44 , You can use learnable skip connections of data to cross-layer filters 44 as shown by Highway Net. Various representations of cross-layer filters 44 are exemplary in the 9 and 10 shown.

Handling the scaling of objects 46 has become important in dealing with features of different resolution, especially for handling small objects. Dilated convolutions are popular for handling multi-scale and can be mapped as multi-scale images at different increments. Various representations of handling scaling of objects 46 are exemplary in 11 shown. The preferred design option is Execute multi-scale feature maps as used in SSDs using multi-scale feature maps as in 11c and MS-CNN unified multi-scale.

Split Branching and Summation Joining refers to the summation of heavily used nodes.

Key aspects of network design are transformed into an adaptive format in which the network topology of training data can be learned. This involves sequential iterating over various combinations of aspects. Therefore, marginalization can be performed with respect to all other meta-architecture variables, and a brute-force search for variable size configurations can be performed.

Starting from a basic network, as eg in 4 an optimization of the network topology is performed using Bayesian optimization.

step S120 refers to performing network marginalization and other meta-architecture optimization techniques. This can be commonly referred to as meta-architecture marginalization learning. Therefore, the dimensionality is limited, with summation nodes remaining.

According to step S130 Marginalization is performed on other meta-architecture variables, including brute-force search for variable-sized configurations. This includes a first step to freezing a first set of aspects and optimizing a second, different set of aspects. Subsequently, a second step of freezing the second set of aspects and optimizing the first set of aspects is performed.

step S140 refers to the refinement of the network topology and the basic hyperparameters by performing a batch update of the training data over several iterations to optimize the entire network topology and basic hyperparameters. If a divergence is detected, the iterations are stopped. In this case, if the iterations are stopped, a previous optimal network can be maintained.

12 shows a structure of the CNN 50 as a network of encoder / decoder type. The CNN 50 receives an input 52 that the CNN 50 passes. The CNN 50 has a coding section 54 on, the input first 52 processed. The CNN 50 also has a decoding section 56 which then performs a reconstruction to an edition 58 to create. 12 also provides a definition of channels and filters in a right part of 12 you can see.

LIST OF REFERENCE NUMBERS

10

scene

12

person

14

vehicle

16

parking spot

18

Street

20

Fully Convolutional Network

22

folding block

24

framework

26

module

28

connection

30

input channel

32

output channel

34

input layer

36

output layer

40

wider filter bank

42

Cross-channel filter

44

Cross-layer filter

46

Handling the scaling of objects

50

convolutional neural network, CNN

52

input

54

coding

56

Decoding section

58

output

QUOTES INCLUDE IN THE DESCRIPTION

This list of the documents listed by the applicant has been generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Cited patent literature

• US 2017293837 A1 [0012]

Claims (12)

A method of designing and training a meta-architecture of a convolutional neural network (CNN) (50), comprising the steps of: Constructing a base network with hyperparameters of training data; Constructing a meta-architecture having formalized extensions (40, 42, 44, 46); Performing network marginalization and other optimization techniques related to the meta-architecture; Marginalization for other meta-architecture variables, including a brute-force search for variable-size configurations; and Refining a network topology and basic hyperparameters by performing a batch update on the training data over multiple iterations. Method according to Claim 1 characterized in that the step of constructing a basic network with hyperparameters determined by training data comprises initializing various design options and modeling the construction as a hyperparameters search problem to refine design options using the Bayesian optimization algorithm. Method according to Claim 1 or 2 characterized in that the step of constructing a meta-architecture having formalized extensions comprises transforming critical network design aspects into a learnable format, wherein the network topology can be learned from training data. Method according to Claim 3 characterized in that the step of transforming critical network design aspects into a learnable format wherein the network topology can be learned from training data comprises starting with the base network and optimizing the network topology using Bayesian optimization. A method according to any one of the preceding claims, characterized in that the step of constructing a meta-architecture having formalized extensions (40, 42, 44, 46) comprises applying formalized extensions (40, 42, 44, 46) from a group comprising a broader filterbank (40), cross-channel filters (42), cross-layer filters (44), split-branching and summation-joining, and handling scaling of objects (46). A method according to any one of the preceding claims, characterized in that the step of performing network marginalization and other optimization techniques related to the meta-architecture comprises a first step of freezing a first set of aspects and optimizing a second, another set of aspects and a second step for freezing the second set of aspects and optimizing the first set of aspects. Method according to Claim 6 characterized in that the step of transforming critical network design aspects into a learnable format comprises sequentially iterating over various combinations of aspects. The method of any one of the preceding claims, characterized in that the step of refining a network topology and basic hyperparameters comprises identifying a divergence and stopping the iterations in case of a detected divergence. Use of the method according to one of Claims 1 to 8th in a driving support system of a motor vehicle. A driving assistance system for a motor vehicle, wherein the driving support system is configured to perform the method of any one of Claims 1 to 8th perform. A non-transitory computer-readable medium having instructions stored therein that, when executed by a processor, cause a driving support system to perform the method of any one of Claims 1 to 8th perform. A motor vehicle comprising: a driving support system having a data processing device, wherein the non-transitory computer-readable medium according to Claim 11 stored commands are executed by the data processing device of the driving support system.

Images