Error-guaranteed inference computation for neural networks

The method uses an acceleration module to convolve input data with convolution kernels and control matrices to detect and correct transient errors in neural networks, enhancing reliability in hardware platforms by reducing computational effort and memory usage.

JP7581540B2Active Publication Date: 2024-11-12ROBERT BOSCH GMBH
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
JP2023573000
Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
Filing Date
2021-05-25
Publication Date
2024-11-12
Estimated Expiration
2041-05-25

AI Technical Summary

Technical Problem

Neural networks in hardware platforms for mobility applications are prone to transient errors due to high-energy photons and voltage disturbances, which are difficult to diagnose and correct, especially in densely integrated systems.

Method used

A method using an acceleration module to convolve input data with convolution kernels, generating output matrices and control matrices, and comparing elements to detect and correct transient errors efficiently, with additional checks for hardware component failures.

Benefits of technology

Reduces computational effort and memory usage for error detection, allowing precise localization of errors and enabling reliable operation of neural networks in hardware platforms, even in the presence of sporadic faults.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007581540000001
    Figure 0007581540000001
  • Figure 0007581540000002
    Figure 0007581540000002
  • Figure 0007581540000003
    Figure 0007581540000003
Patent Text Reader

Abstract

A method (100) for operating a hardware platform for inference computation of a convolutional neural network, the method (100) comprising the steps of: convolving (110) an input matrix (1) having input data of the neural network with a plurality of convolution kernels (2a-2c) using an acceleration module, thereby generating a plurality of two-dimensional output matrices (3a-3c); summing (120) the convolution kernels (2a-2c) element-wise to generate a control kernel (4); and convolving (120) the input matrix (1) with the control kernel (4) using the acceleration module, thereby generating a two-dimensional control matrix (3a-3c). The method includes a step (130) of generating a matrix (5) of a control matrix (5), a step (140) of comparing each element (5*) of the control matrix (5) with the sum of the elements (3a*-3c*) in the output matrices (3a-3c) corresponding to the element (5*), and, in response to a deviation (150) occurring for one element (5*) of the control matrix (5) as a result of this comparison (140), a step (160) of checking whether the element (3a*-3c*) of at least one output matrix (3a-3c) corresponding to this element (5*) of the control matrix (5) has been correctly calculated by at least one additional control calculation.
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present invention relates to guaranteeing the computations that occur during inference operations of a neural network against transient errors in the hardware platform used. [Background technology]

[0002] Background technology When inferring a neural network, the activations of these very large numbers of neurons are calculated by weighting and summing the inputs supplied to the neurons according to weights acquired during the training of the neural network, i.e. performing multiple multiplications and then adding up the results of these multiplications (multiply-and-accumulate (MAC)). In particular in mobile applications, such as for example in the at least partially automatic guidance of vehicles in road traffic, neural networks are implemented on hardware platforms dedicated to such calculations. These platforms are particularly efficient in terms of computing power, with regard to hardware cost and energy consumption per unit.

[0003] As the integration density of these hardware platforms increases, the probability of transient, i.e. sporadic, computational errors increases, where bits can be randomly "flipped" by high-energy photons, e.g. from background radiation, striking a memory location or a processing unit of the hardware platform. Furthermore, the hardware platforms share the on-board power supply with a number of further consumers, which may couple disturbances, e.g. voltage peaks, into the hardware platform, right in the vehicle. The tolerances in this regard become narrower as the integration density of the hardware platforms increases.

[0004] DE 10 2018202095 A1 discloses a method in which, when processing a tensor of input values ​​to generate a tensor of output values, output values ​​that are erroneously calculated by a neural network can be identified and also corrected using additional control calculations. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] DE 102018202095 A1 Summary of the Invention [Means for solving the problem]

[0006] Disclosure of the Invention In the framework of the present invention, a method for operating a hardware platform for inference calculation of a convolutional neural network has been developed. The hardware platform comprises at least one acceleration module configured to calculate the convolution of an input matrix with a convolution kernel by applying the convolution kernel to a number of positions inside the input matrix and to output the result of the convolution as a two-dimensional output matrix. By "configured" it is understood in particular that the range of tasks that the acceleration module can perform is significantly limited in favor of a much higher output in exactly such tasks compared to the CPU or GPU of a conventional computer. The input matrix and the convolution kernel may here in particular be three-dimensional, for example, which is particularly advantageous for processing image data. However, the input matrix and the convolution kernel can also be generalized to higher dimensions. Thus, for example, in the case of video data or other time-varying data, three dimensions can represent spatial coordinates and the fourth dimension can represent time.

[0007] Thus, quite generally, the neural network may be configured as a classifier for assigning surveillance data, such as, for example, camera images, thermal images, radar data, LIDAR data or ultrasound data, to one or more classes of a given classification. These classes may, for example, represent objects or conditions in the monitored area, which are important to detect. The surveillance data may, for example, come from one or more sensors mounted on the vehicle. From the class assignments provided by the neural network, actions may then be derived, for example of a driver assistance system or of a system for at least partially automatic guidance of the vehicle, adapted to the concrete traffic situation. The neural network may, for example, be a convolutional neural network (CNN) divided into layers.

[0008] In this method, an input matrix with input data of a neural network is convolved with multiple convolution kernels using an acceleration module. This means that for each position where a convolution kernel is applied inside the input matrix, the elements of the input matrix covered by the convolution kernel are weighted and summed, where the weights are given by the elements of the convolution kernel. The input matrix is ​​"scanned" in two dimensions by the convolution kernel, resulting in multiple such weighted sums, from which an output matrix corresponding to the convolution kernel is formed. Thus, multiple such output matrices are generated for multiple convolution kernels.

[0009] The convolution kernels are summed element-wise to generate the control kernel. Using the acceleration module, the input matrix is ​​convolved with the control kernel, which results in a two-dimensional control matrix, similar to the application of a convolution kernel.

[0010] The convolution kernels may in particular be of equal size, for example, although this is not necessarily required. If the convolution kernels are of different sizes, the convolution kernels may be, for example, virtually zero-filled at the edges with the size of the largest convolution kernel, and then all convolution kernels may be summed element-wise to generate the control kernel.

[0011] Each element of the control matrix is ​​compared with the sum of the elements in the output matrix that correspond to each element of the control matrix. For example, if the convolution kernel and the control kernel "scan" the input matrix in the dimensions x and y, respectively, and have the same depth as the input matrix in the third dimension z, then the output matrix and the control matrix corresponding to the convolution kernel also extend along the dimensions x and y and are "stacked" in the third dimension z. Thus, for each coordinate pair (x,y), the sum of the elements of all the output matrices with these coordinates (x,y), i.e. the sum formed along the "height column" in the z direction, should be equal to the element of the control matrix with the same coordinates (x,y). This can be deduced from the associative laws of mathematics and understood by analogy that when counting coins, an equal amount should be obtained regardless of whether the individual values ​​of the coins are added directly or whether the coins are first grouped into rolls according to denomination and then the roll values ​​are added.

[0012] In response to this comparison resulting in a deviation for an element of the control matrix, at least one additional control calculation is performed to check whether an element of at least one output matrix corresponding to this element of the control matrix has been correctly calculated.

[0013] It has been realized that by organizing the error check in this way, the additional effort in the form of calculation time and memory is significantly reduced in relation to the specific hardware platform mentioned above. For the calculation of the control matrix, an acceleration module similar to that for the calculation of the output matrix is ​​used, so that this calculation only requires a very short additional time. Since the aim is precisely to find transient, i.e. sporadically occurring errors, it can be expected that no deviations will occur for the majority of comparisons (more than 99%) in normal operating circumstances. If these cases are handled as efficiently as possible, time is invested in additional control calculations for deviation cases, which allows the error to be located more accurately. Here, the specific type of this additional control calculation and the measures taken to remove the more accurately located error are fundamentally not limited. Rather, it is advantageous that the choice of control calculations or other measures is determined in particular according to how much effort the calculations or other measures are and how frequently transient errors can be expected in a concrete application.

[0014] If deviations are identified, this can essentially be caused by miscalculating one or more of the elements in the output matrix that correspond to the elements of the control matrix and / or by miscalculating the elements of the control matrix themselves. However, it is precisely in the context of the present invention that it is important to recognize that transient errors: The probability that two transient errors occur at times that involve elements in two output matrices that are separated from each other in the z direction but have the same coordinates (x,y), or The probability that two transient errors occur at a time when at least one element of the output matrix with coordinates (x,y) and an element of the control matrix with the same coordinates (x,y) are related to the error is extremely low.

[0015] In such a case, even if the full inference calculation must be repeated because the error cannot be further limited, this means that the power loss may not be noticeable in the concrete application due to the low probability. Therefore, for the purpose of further limiting and correcting the transient error, The assumption that an element of the output matrix having exactly the same coordinates (x,y) as the element of the control matrix currently being examined has been calculated incorrectly; Or the assumption that the elements of the control matrix themselves are miscalculated. can be assumed.

[0016] The occurrence of such individual, transient errors is expected to mean significant inhibition of concrete applications in many cases, e.g., in the case of conventional hardware platforms used for at least partially automated driving, compared to the further limitations described below, where complete discarding and repeating of inference calculations and possibly error correction here.

[0017] The above and all following considerations are appropriate regardless of whether the input matrix contains the complete input data of the neural network or only a part of it. In many applications, the complete input data of the neural network and the complete output matrix generated from this input data do not fit into the internal buffer ("on-chip memory") of the hardware platform, so the hardware platform processes the data one by one (in so-called tiles). The results obtained for each tile are then collected in a larger external memory outside the accelerated hardware platform.

[0018] During convolution with at least one convolution kernel, a bias value corresponding to the convolution kernel may further be added to the elements of the output matrix generated by the convolution kernel, and the sum of these bias values ​​may then be added to all elements of the control matrix.

[0019] In a particularly advantageous embodiment, an additional control calculation is used to check whether the row or column of at least one output matrix, which contains the element to be checked, has been calculated correctly. For such a check, an acceleration module can also be used, although it is not provided preferentially for this task. If information is thus obtained that an element of a certain output matrix (i.e. an element with a certain z-coordinate) has not been calculated correctly, two direct inferences can be drawn from this. On the one hand, it turns out that there is in fact an error in the output matrix, and not, for example, only an element of the control matrix is ​​incorrectly calculated. On the other hand, the specific output matrix in which the error is present, i.e. the z-coordinate of the error, is also known. Thus, the error is located on the specific element in relation to the coordinates (x,y) already determined by the first comparison.

[0020] In order to so to speak "repurpose" the acceleration module for such an output, in a particularly advantageous configuration, the input matrix is ​​expanded by elements to be examined, each of which may in particular be, for example, a simple sum of elements from a certain region of the input matrix, which are then convolved by the acceleration module with a convolution kernel corresponding to the at least one output matrix currently being examined, so that a control value is obtained.

[0021] The sum of the elements in the examined row or column is compared to a control value, and in response to a deviation from this comparison, it is determined that the row or column was incorrectly calculated, thereby also determining that the originally examined element of the output matrix was incorrectly calculated.

[0022] If it is established that an element of the output matrix was in fact calculated incorrectly, this element can be corrected by the deviation determined during the comparison. As mentioned above, it should be assumed that there is exactly one error. Therefore, both the initial comparison with an element of the control matrix and the comparison with the control value will yield the same results.

[0023] Because the probability only needs to be calculated for a single error, the search for further errors can be stopped as soon as the first error is identified.

[0024] However, cases may also arise where all elements in all output matrices that correspond to elements of the control matrix (i.e. elements with the same coordinates (x,y)) are recognized as correct by the control calculation. In this case, it can be established that an element of the control matrix was calculated incorrectly, i.e. the initial calculation of the output matrix was correct, but a single, transient error that should have been anticipated only occurred during the subsequent calculation of the control matrix. In this case, further calculations can usually be carried out with the output matrix according to the intended application. Otherwise, errors in the calculation of the control matrix can be ignored.

[0025] The previous discussion assumed that only transient errors always exist. However, the occurrence of repeated errors may no longer be a completely random transient error, but may be a sign that a hardware component or memory location is beginning to fail. For example, in the case of semiconductors, when interdiffusion occurs due to overheating or aging at the pn junction between the hole-doped layer and the electron-doped layer, the amount of energy required to flip a bit in the memory may be reduced compared to normal conditions, and for example, gamma quanta or charged particles may add this amount of energy with a higher probability from background radiation. In this case, errors still occur at random times, but the errors will gradually increase in hardware components or memory cells with damaged pn junctions.

[0026] In another particularly advantageous configuration, the error counter is incremented in response to a deviation occurring in one of the comparisons for at least one hardware component or at least one memory area considered as a cause of the deviation. In this case, for example, in the framework of general maintenance, the error counters for comparable components can be compared with each other. In this case, for example, if a hardware component is found among multiple hardware components having a comparable nominal structure, with a significantly increased error counter, it is possible that a fault is starting to occur in this hardware component.

[0027] Thus, in particular, it is possible to recognize that a hardware component or memory area is faulty, for example in response to the error counter being ascertained to be above a predefined threshold. In response to this, for example, the hardware platform can be reconfigured so that a spare hardware component or a spare memory area is used for further calculations instead of the hardware component recognized as faulty or the memory area recognized as faulty. In particular, it may be worthwhile to provide such a spare measure for fully automatic guidance of a vehicle that does not allow the driver to take over control even in the case of an error. The vehicle can reach a workshop even in the case of a breakdown ("limp home mode") and does not have to be towed at a cost.

[0028] Advantageously, optical image data, thermal image data, video data, radar data, ultrasound data and / or LIDAR data are provided as input data. These are the most important types of measurement data, on the basis of which the at least partially autonomous vehicle is oriented in the traffic space. The measurement data can be obtained by physical measurement processes and / or by partial or complete simulations of such measurement processes and / or by partial or complete simulations of technical systems that can be monitored by such measurement processes. For example, a photorealistic image of the situation can be generated by means of a computational tracing of light beams ("Raytracing") or by means of a neural generator network (for example a Generative Adversarial Network (GAN)). Here, knowledge based on a simulation of the technical system, such as for example the position of certain objects, can also be introduced as a condition. The generator network can be trained to generate images that satisfy these conditions in a targeted manner (for example a Conditional GAN ​​(cGAN)).

[0029] The output matrix can be processed to generate actuation control signals, which can then actuate quality control systems and / or medical imaging systems and / or access control systems of vehicles and / or mass-produced products. In this context, the above-mentioned error check has the advantageous effect that sporadic malfunctions, which arise "out of thin air" without any concrete trigger and are therefore usually very difficult to diagnose, are advantageously avoided.

[0030] The method may in particular be fully or partly computer-implemented. The invention therefore also relates to a computer program comprising machine-readable instructions which, when said instructions are executed in one or more computers, cause the computer to carry out one of the above-mentioned methods. In this sense, control devices for vehicles and embedded systems for technical devices, which are also capable of executing machine-readable instructions, can also be considered as computers.

[0031] The invention likewise relates to a machine-readable data carrier containing a computer program and / or a download product, which is a digital product that can be transmitted over a data network, i.e. downloaded by a user of the data network, and which can be sold for immediate download, for example in an online shop.

[0032] Furthermore, the computer may comprise a computer program, a machine-readable data carrier or a downloadable product.

[0033] Further measures for improving the invention are shown in more detail below on the basis of the drawings together with a description of preferred embodiments of the invention. [Brief description of the drawings]

[0034] [Figure 1] FIG. 1 illustrates an embodiment of a method 100. [Diagram 2]FIG. 2 illustrates the rapid calculation of a control matrix 5 by a control kernel 4. [Figure 3a] FIG. 13 illustrates precise location of an error based on rows 3a#-3c# of output matrices 3a-3c. [Figure 3b] FIG. 13 illustrates precise location of an error based on columns 3a#-3c# of output matrices 3a-3c. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0035] Working Example 1 is a schematic flow chart of an embodiment of a method 100. In step 105, data types which are of most importance for the orientation of at least partially automated vehicles, in particular in road traffic, can be provided as input data in the form of an input matrix 1.

[0036] In step 110, input matrix 1, which in this example is three-dimensional, is convolved with convolution kernels 2a-2c, which in this example are also three-dimensional, to produce two-dimensional output matrices 3a-3c, respectively. In step 120, convolution kernels 2a-2c are summed element-wise to produce control kernel 4. Input matrix 1 is convolved with control kernel 4, resulting in two-dimensional control matrix 5.

[0037] In step 140, each element 5* of the control matrix 5 is compared with the sum of the elements 3a* to 3c* in the output matrices 3a to 3c that correspond to element 5*. In step 150, it is checked whether this comparison 140 indicates a deviation. If this comparison 140 indicates a deviation (truth value 1), in step 160, it is checked whether the element 3a* to 3c* of at least one output matrix 3a to 3c that corresponds to this element 5* of the control matrix 5 has been calculated correctly.

[0038] If it is determined in step 170 that the elements 3a* to 3c* of the output matrices 3a to 3c have been calculated incorrectly, then these elements 3a* to 3c* can be corrected in step 180 by the deviation determined in the comparison.

[0039] It is also possible to subsequently check in step 190 whether elements of all output matrices 3a-3c corresponding to element 5* of control matrix 5 have been correctly calculated, and to confirm in step 200 that all these elements 3a*-3c* have been correctly calculated (truth value 1). In this case, while element 5* of control matrix 5 has not been correctly calculated in step 210, it is simultaneously confirmed that all output matrices 3a-3c are correct.

[0040] If the output matrices 3a-3c are all correct or if possible errors have been corrected in step 180, the output matrices 3a-3c can be subjected to further evaluation. In step 270, these output matrices 3a-3c can be processed to generate, inter alia, a drive control signal 6. Then, in step 280, the vehicle 50 and / or the sorting system 60 and / or the quality control system 70 for mass-produced products and / or the medical imaging system 80 and / or the access control system 90 can be driven by this drive control signal 6.

[0041] On the other hand, if it is determined in step 220 that the output matrices 3a-3c have not been calculated correctly, then in step 230 an error counter can be incremented for at least one hardware component or at least one memory area considered as a cause of the deviation. Then, if it is determined in step 240 that the error counter is above a predefined threshold (truth value 1), then in step 250 a hardware component or memory area can be recognized as faulty. Then, in step 260 the hardware platform can be reconfigured such that instead of the hardware component recognized as faulty or the memory area recognized as faulty, a spare hardware component or a spare memory area is utilized for further calculations.

[0042] Inside box 110, a possible configuration of the convolution with the convolution kernels 2a-2c is shown. In block 111, during the convolution, a first bias value 7a is added to the values ​​of the first output matrix 3a, a second bias value 7b to the values ​​of the second output matrix 3b and a third bias value 7c to the values ​​of the third output matrix 3c. In block 112, the sum 7a+7b+7c of these bias values ​​7a, 7b, 7c is also added to all elements of the control matrix 5.

[0043] In block 161, during an additional control calculation 160, it can be checked in particular whether the row or column 3a#-3c# containing the tested element 3a*-3c* of at least one output matrix 3a-3c has been correctly calculated. This is shown in more detail in FIG.

[0044] For this test, in particular an acceleration module provided for convolution, for example of a hardware platform, can be "repurposed". For this purpose, in block 162, the input matrix 1 is expanded by the element 11 to be tested. Then, in block 163, the element 11 to be tested is convolved with the convolution kernel 2a-2c corresponding to at least one output matrix 3a-3c using the acceleration module, so that a control value 31 is obtained. In block 164, the sum of the elements of the row or column 3a#-3c# is compared with the control value 31. If in block 165 it is determined that a deviation occurs from this comparison (truth value 1), then in block 166 it is determined that the row or column 3a#-3c# was calculated incorrectly and thus also the elements 3a*-3c* to be tested of the output matrix 3a-3c were calculated incorrectly.

[0045] Fig. 2 shows how a first check for possible calculation errors resulting from the use of a control kernel 4 on a hardware platform can be performed particularly efficiently using an acceleration module. The convolution of the input matrix 1 with each of the convolution kernels 2a-2c generates the output matrices 3a-3c. The control kernel 4 is formed by the element-wise summation of the convolution kernels 2a-2c. The convolution of the input matrix 1 with the control kernel 4 results in a control matrix 5 having a size equivalent to the output matrices 3a-3c. Each element 5* of the control matrix 5 should be equivalent to the sum of the elements 3a*-3c* corresponding to each element 5* of the output matrices 3a-3c having the same coordinates (x,y) in the plane of the respective output matrices 3a-3c.

[0046] FIG. 3 shows a further control calculation, which in block 161 makes it possible to further limit possible errors.

[0047] Figure 3a assumes that element 5* in the top left corner of control matrix 5 does not match the sum of its corresponding elements 3a*-3c* in output matrices 3a-3c. Then, for each of the output matrices 3a-3c, it is checked whether each row 3a#-3c# containing the corresponding element 3a*-3c* has been calculated correctly. As mentioned above, this allows for quicker testing than checking each element 3a*-3c* individually.

[0048] In the example shown in figure 3a, it can be seen that during this control calculation, row 3b# of output matrix 3b was calculated incorrectly, therefore it is certain that element 3b* was calculated incorrectly and a corresponding correction can be made.

[0049] The process proceeds exactly the same if the columns 3a# to 3c# of the output matrices 3a to 3c, which contain the elements to be tested 3a* to 3c* respectively, are to be tested with exact calculations, as specifically shown in FIG. 3b.

Claims

1. A method (100) of operating a hardware platform for inference computation of a convolutional neural network, the hardware platform having at least one acceleration module configured to compute a convolution of an input matrix (1) with convolution kernels (2a-2c) by applying the convolution kernels (2a-2c) to a plurality of positions inside the input matrix (1) and to output the result of the convolution as a two-dimensional output matrix (3a-3c), comprising: - convolving (110) an input matrix (1) comprising input data of said neural network with a number of convolution kernels (2a-2c) using said acceleration module, thereby generating a number of two-dimensional output matrices (3a-3c); - summing (120) said convolution kernels (2a-2c) element-wise to generate a control kernel (4); - convolving (130) said input matrix (1) with said control kernel (4) using said acceleration module, thereby resulting in a two-dimensional control matrix (5); a step (140) of comparing each element (5*) of said control matrix (5) with the sum of the elements (3a* to 3c*) in said output matrices (3a to 3c) corresponding to said element (5*); - in response to a deviation (150) resulting from said comparison (140) for an element (5*) of said control matrix (5), checking (160) by means of at least one additional control calculation whether an element (3a*-3c*) of at least one output matrix (3a-3c) corresponding to the element (5*) of said control matrix (5) has been correctly calculated; The method (100) comprising:

2. 2. The method (100) of claim 1, wherein during the convolution (110) with at least one convolution kernel (2a-2c), bias values ​​(7a-7c) corresponding to said convolution kernels (2a-2c) are added (111) to elements of an output matrix (3a-3c) generated by said convolution kernels (2a-2c), and the sum of all bias values ​​(7a-7c) is also added (112) to all elements of the control matrix (5).

3. The method (100) according to claim 1 or 2, wherein the additional control calculation (160) checks (161) whether a row or column (3a# to 3c#) of the at least one output matrix (3a to 3c) containing an element (3a* to 3c*) to be tested has been correctly calculated.

4. In the framework of said control calculation, The input matrix (1) is expanded (162) by the elements (11) to be examined, said element under test (11) is convoluted (163) by a convolution kernel (2a-2c) corresponding to said at least one output matrix (3a-3c) using said acceleration module, thereby obtaining a control value (31); the sum of the elements of said row or column (3a#-3c#) is compared (164) with said control value (31); in response to a deviation (165) resulting from said comparison (164), it is determined (166) that said row or column (3 a#-3 c#) has been calculated incorrectly and therefore that the element (3 a*-3 c*) under test of said output matrix (3 a-3 c) has also been calculated incorrectly; 4. The method (100) of claim 3.

5. 5. The method (100) according to claim 1, wherein in response to a confirmation (166, 170) that an element (3a* to 3c*) of an output matrix (3) has been calculated incorrectly, said element (3a* to 3c*) is corrected (180) by the deviation determined in the comparison.

6. 6. A method (100) according to any one of claims 1 to 5, wherein for all output matrices (3a-3c), elements (3a*-3c*) corresponding to an element (5*) of the control matrix (5) are checked (190) as to whether said element (3a*-3c*) has been correctly calculated, and in response to a confirmation (200) that all elements (3a*-3c*) have been correctly calculated, it is confirmed (210) that the element (5*) of the control matrix (5) has not been correctly calculated.

7. 7. The method (100) of claim 1, wherein an error counter is incremented (230) in response to one of the comparisons (140, 164) causing a deviation (220) for at least one hardware component or at least one memory area considered to be a cause of the deviation.

8. 8. The method (100) of claim 7, wherein in response to determining that the error counter is above a predefined threshold (240), the hardware component or memory area is recognized (250) as having failed.

9. 10. The method (100) of claim 8, wherein the hardware platform is reconfigured (260) such that a spare hardware component or a spare memory area is utilized for further computation in place of the hardware component or the memory area identified as being faulty.

10. The method (100) according to any one of claims 1 to 9, wherein optical image data, thermal image data, video data, radar data, ultrasonic data and / or LIDAR data acquired by a physical measurement process and / or by a partial or complete simulation of such a measurement process and / or by a partial or complete simulation of a technical system that can be monitored by such a measurement process are provided as input data (105).

11. The method (100) according to any one of claims 1 to 10, wherein the output matrices (3a-3c) are processed to generate (270) driving control signals (6) and a vehicle (50) and / or a quality control system (70) of a mass-produced product and / or a medical imaging system (80) and / or an access control system (90) are driven (280) by the driving control signals (6).

12. 12. A computer program comprising machine-readable instructions that, when executed on one or more computers, cause the computers to perform a method (100) according to any one of claims 1 to 11.

13. A machine-readable data carrier comprising a computer program according to claim 12.

14. A computer comprising a computer program according to claim 12 and / or a machine-readable data carrier according to claim 13.

Citation Information

Patent Citations

  • Method and device for checking neuronal function in a neural network

    DE102018202095A1

  • Method and device for verifying a neuron function in a neural network

    US20190251005A1