DE102017212328A1 - Function monitoring for AI modules - Google Patents

Abstract

Method (100) for training an artificial intelligence module, KI module (1), which converts input variables (11a-11d) into output variables (13a-13d) through a parameterized internal processing chain (12), wherein the parameters (14) the processing chain (12) are determined (120) such that learning values (15a-15d) of the input quantities (11a-11d) are translated to the matching learning values (16a-16d) of the output quantities (13a-13d) ( 130), wherein the learning values (15a-15d) of the input quantities (11a-11d) are superimposed (110) with a predetermined test pattern (21, 21a-21d) and a response (22, 22a-22d) is determined (140) generating the test pattern (21, 21a-21d) in the values of the outputs (13a-13d) provided by the internal processing chain (12). Method (200) of testing an artificial intelligence module, KI Module (1), which inputs (11a-11d) through a parametrized internal processing chain (12) in output in the operation of the KI module (1) the real input variables (11a-11d) a predetermined test pattern (21, 21a-21d) is superimposed (210) and a reaction (210 22, 22a-22d) which produces the test pattern (21, 21a-21d) in the outputs (13a-13d). Associated AI module (1) and computer programs.

Description

The present invention relates to methods by which the monitoring of KI modules for correct operation can be improved, as well as to a suitably equipped KI module.

State of the art

Driving a vehicle in traffic is a complex process that needs to be learned by people through hands-on experience in driving lessons. Accordingly, it is difficult to grasp this control task in program code for an autonomous control system. Therefore Artificial Intelligence, AI, is used for this purpose, which, analogous to the learning process of humans, derives from unconscious rules for the control task from learning data and then translates the input quantities determined by a sensor of the vehicle into the correct output variables for an actuator system.

Typically, an AI module receives a variety of inputs that translate into a variety of outputs. The translation takes place in a strongly meshed and massively parallel internal processing chain. Here, for example, connections between the nodes of a neural network are free parameters. In the learning phase, those values are determined for the free parameters with which predetermined learning values for the input variables are translated into the corresponding learning values for the output variables, ie, for example, an image containing a scene with a stop sign is classified as an image containing a stop sign. It is then assumed that in real operation (deployment phase) real values for the input variables are translated correctly, ie, for example, real images of traffic situations are correctly classified as to which objects they contain. The US 8,468,109 B2 exemplifies massive parallel neural networks.

Control tasks for participation in traffic are safety relevant. The US 8,953,436 B2 discloses a neural network for vehicles in which tasks may be delegated from a faulty part of the network to a working part, for example due to a hardware defect.

For the detection of whether there is an error, the reveals US 7,937,343 B2 a Monte Carlo method with which the behavior of a neural network can be plausibilized on the basis of random samples so that a given confidence can be assumed to work correctly.

The US Pat. No. 6,473,746 B1 discloses another method by which it is possible to check whether a neural network is properly trained for a specific safety-relevant control task.

Disclosure of the invention

Within the scope of the invention, a method for training an artificial intelligence module, KI module, has been developed. The AI module translates input variables into output variables through a parameterized internal processing chain. During training, the parameters of this processing chain are determined in such a way that learning values of the input variables are translated to the corresponding learning values of the output variables.

For example, in an AI module whose internal processing chain comprises a neural network, the connections between nodes at different levels of the network belong to the parameters of the processing chain. The connections may be occupied, for example, with numerical probabilities.

The learning values of the input variables are superimposed with a predetermined test pattern, and a reaction is determined which causes the test pattern in the values of the output values supplied by the internal processing chain.

In particular, an overlay means any addition of the test pattern to the values of the input variables which can also be carried out during the operation of the AI module without impairing the primary function of the AI module.

For example, in a particularly advantageous embodiment of the invention, the input variables of the KI module comprise image data, and the output variables of the KI module can then include a statement about the occurrence of at least one object in the image data. The test pattern can then be added to the image data as an additional object, for example. In this case, for example, the location in the image at which the test pattern is inserted can be selected such that objects to be recognized by the KI module are not covered in the image. However, the test pattern can also be inserted into the image with such a degree of transparency, for example, that it still reveals an underlying object. This embodiment has the advantage that the function of the KI module can be monitored with respect to a classification or evaluation of image data.

In a further particularly advantageous embodiment of the invention, the input variables of the KI module may comprise audio data, and the output variables of the KI module may then include a statement about the occurrence of at least one specific noise in the audio data. The test pattern may then be superimposed, for example, in a frequency range that is not used by a sound that is to be detected by the KI module. However, the test pattern can also be mixed into the audio data, for example, with such a strength that the noise to be recognized is still recognized in spite of frequency overlap with the test pattern, analogous to the partially transparent insertion into an image. This embodiment has the advantage that the function of the KI module can be monitored with respect to a classification or evaluation of audio data.

It was recognized that the overlay with the test pattern at the time of training creates a benchmark that can be used to monitor the proper functioning of the AI module at any time during the subsequent operation of the AI module. This monitoring is possible without interrupting the ongoing operation of the AI module. In particular, in the case of safety-critical control tasks, the monitoring can thus be constantly active, without the AI module having to be configured redundantly for this purpose alone.

Such a function monitoring is particularly advantageous when the internal processing chain of the AI module is highly meshed and parallelized. This is the case, for example, in a particularly advantageous embodiment of the invention, in which the internal processing chain of the KI module comprises at least one neural network. The effects of errors in hardware components on the correct translation of the input variables into output variables has hitherto been difficult to predict in the case of such KI modules since it is only during the course of training that it becomes clear which hardware components are actually used in which operating situations. The susceptibility of a specific function of the AI module to defects in the hardware can in particular depend on how many parallel hardware components the execution of exactly this function is distributed according to the training. If the execution is distributed over many hardware components, an error in each of these hardware components can affect the correct execution, but on the other hand, a single error in a hardware component can be partially or completely compensated by the intact hardware components. If the function is performed essentially only by a hardware component, only an error in this one hardware component can affect the correct execution; but if this happens, it can not be compensated by other hardware components. With the method described here, a systematic test of all involved hardware components can be performed.

Among a hardware component in this context are all units involved in the calculation, e.g. Arithmetic units, and / or memory modules or the like, understood. As an example of the hardware for performing the calculation of a neural network, for example, a GPU, a CPU, an FPGA, an ASIC or any other arithmetic unit can be understood. As hardware components for a GPU, here e.g. Visual Processing Units, Shader Processing Units, Tensor Processing Units, etc.

In a particularly advantageous embodiment of the invention, when training the AI module, the parameters of the processing chain are adapted to the effect that the reaction caused by the test pattern approaches a desired reaction. The correct function of the AI module on which the AI module is trained then also includes the translation of the test pattern to the known desired response. This translation should also be carried out during the later operation of the AI module with consistent quality. If this quality deteriorates, for example if the test pattern is no longer detected or only with a worse confidence than at the time of the training, it can be concluded that the AI module is defective. In this case, for example, a gradual deterioration can already be detected, so that countermeasures can be taken in good time before the AI module completely fails in the control task carried out by it.

The reaction caused by the test pattern can alternatively or also be stored in combination as a reference for a later functional test of the KI module. Thus, for example, the test pattern can only be superimposed by the learning values of the input variables if the actual training has already been completed, ie if the parameters of the processing chain have already been determined and are no longer changed. Then the function monitoring is completely independent of this actual training. The AI module is then not specifically trained for the recognition of the test pattern, however, a change in the quality with which the test pattern is recognized is still a strong indicator that a hardware defect is present. Saving the reaction as a reference without specific training on the recognition of the test pattern, for example, be useful if an existing process for training an AI module has been certified to a safety-relevant task and the inclusion of the test pattern in the training invalidate the certification.

In a further particularly advantageous embodiment of the invention, at least two test patterns are selected such that the first test pattern is predominantly processed by a part of the processing chain which is executed on a first hardware component of the KI module, while the second test pattern is mainly processed by a part of the processing chain that is executed on a second hardware component of the AI module.

It has been recognized that, especially in massively parallel and meshed processing chains, the existing hardware components in the operation of the AI module, ie in the deployment phase after the training, sometimes be exploited significantly differently. If the KI module is designed, for example, to recognize traffic signs in image data, then in real operation the vehicle comparatively seldom encounters the requirement no. 129, which warns that the road leads to an unsecured shore. For example, a defect in a specific hardware component can have a particular effect on the recognition of this traffic sign, but not on the detection of the priority signs and speed restrictions which occur much more frequently in everyday life. Thus, the defect may go unnoticed for quite some time and have a fatal impact at the next occasion that the detection of the No. 129 sign is required. The same applies, for example, to the detection of a red flashing light or an approaching train, even on an unrestricted level crossing.

By now specifically controlled by the choice of the test pattern, which hardware component of the AI module is checked for correct function, it can thus be ensured in particular that the function monitoring actually extends to all existing hardware components. An error can not be "hidden" in the manner described before the review.

For this purpose, for example, in an AI module designed to detect objects in image data, alternatively or also in combination with the appearance of the test pattern, the location at which the test pattern is inserted into the image data can be varied. The parallelization of such an AI module may be organized, for example, such that certain areas of the image are assigned to certain input nodes of the processing chain according to a fixed scheme.

The invention also relates to a method for functional testing of an AI module which translates input variables into output variables through a parameterized processing chain.

In the operation of the KI module, a given test pattern is superimposed on the real input variables, and a reaction is determined which causes the test pattern in the output variables.

In this case, the KI module may have been trained in particular with the method described above. However, this is not mandatory. For example, the response to the test pattern in the new state of the AI module, in which the hardware is likely to be error-free, can be determined and stored as a reference, for example, immediately following the definition of the parameters of the processing chain. During operation of the KI module, the reaction of the test pattern can then be compared with the previously determined and stored reference.

In a particularly advantageous embodiment of the invention, the reaction is compared with a predetermined desired reaction. For example, if the triggering of the target reaction was part of the training of the AI module, a deviation of the actual response from the target response is a particularly strong indicator that a hardware defect is present.

In accordance with what has been described above, in a further particularly advantageous embodiment of the invention, a statement about the correct function of one or more hardware components is run on which the part of the processing chain of the KI module runs, of which the test pattern is predominantly processed. In this way, an error can be located in the processing chain.

In a particularly advantageous embodiment of the invention, a test pattern is selected which does not occur in the operation of the KI module in the real input variables of the KI module. For example, an image as a test pattern may be a matrix code that does not appear anywhere in public. For example, a noise as a test pattern may be in a frequency range that is not occupied by everyday sounds, and / or may include a non-publicly known tone sequence. In this way, it can be ensured that, for example, an object or noise to be detected by the AI module is not erroneously identified as a test pattern and possibly removed from the processing chain downstream of the AI module. For example, if a traffic sign is identified as a test pattern, the reaction of an autonomous vehicle to this test pattern might fail.

After the above, the invention also relates to an AI module, which Input variables are translated into output variables by a parameterized internal processing chain.

The internal processing chain is designed to translate at least one test pattern presented in the input variables into a predetermined reaction in the output variables, wherein the test pattern does not occur in the operation of the KI module in the real input variables of the KI module.

Training the AI module with a test pattern that does not occur in the input variables of the AI module expected during operation thus leaves behind a "fingerprint" that can be detected in the interconnection of the processing chain in the AI module.

The corresponding parameters of the processing chain need not necessarily be generated on the AI module itself, but can come from any source and be retrofitted, for example, as a software update. Thus, the invention also relates to a computer program having parameters for the internal processing chain of an AI module which, when loaded into an AI module, upgrade the AI module to an AI module according to the invention. Furthermore, the invention also relates to a machine-readable data carrier on which the computer program is stored, and to a downloadable product containing the computer program.

The methods according to the invention do not necessarily require a change of the hardware of the AI module, or of other hardware in the processing of the input variables or the output variables upstream or downstream of the AI module. Thus, for example, the methods may be fully implemented in software that represents a self-salable product. Thus, the invention also relates to a computer program having machine-readable instructions which, when executed on a computer, cause the computer to perform a method according to the invention and / or an AI module to an AI module according to the invention upgrade. The computer may be, for example, a computer that performs control tasks in a vehicle or, for example, a computer that is used in the manufacture or maintenance of a vehicle. Furthermore, the invention also relates to a machine-readable data carrier on which the computer program is stored, and to a downloadable product containing the computer program.

Further measures improving the invention will be described in more detail below together with the description of the preferred embodiments of the invention with reference to figures.

list of figures

• 1 Ausführungsbeispiel des Verfahrens 100 zum Trainieren eines KI-Moduls 1;
• 2 Beispielhafte Verwendung zweier Testmuster 21a und 21b, die in verschiedenen Hardwarekomponenten 1a und 1b des KI-Moduls 1 verarbeitet werden;
• 3 Ausführungsbeispiel des Verfahrens 200 zur Funktionsprüfung eines KI-Moduls 1;
• 4 Beispielhafte Testmuster 21a-21d zur umfassenden Funktionsprüfung eines KI-Moduls 1, das Bilddaten als Eingangsgrößen 11a-11d verarbeitet und hierin Objekte 31-33 erkennt.

It shows:

• 1 Embodiment of the method 100 to train an AI module 1 ;
• 2 Exemplary use of two test patterns 21a and 21b that are in different hardware components 1a and 1b of the AI module 1 are processed;
• 3 Embodiment of the method 200 for functional testing of an AI module 1 ;
• 4 Exemplary test patterns 21a - 21d for comprehensive functional testing of an AI module 1 , the image data as input variables 11a - 11d processed and herein objects 31 - 33 recognizes.

To 1 will be in step 105 of the method exemplified here 100 first a test pattern 21 . 21a - 21d selected, not in the operation of the AI module 1 normally expected inputs 11a -11d occurs. In step 110 becomes the test pattern 21 . 21a - 21d on the learning values 15a - 15d the input variables 11a - 11d impressed. In an iterative process are now in step 120 the parameters 14 the internal processing chain 12 of the AI module 1 adapted so that the in step 130 from the input variables 11a - 11d generated output variables 13a - 13d best possible their learning values 16a - 16d correspond.

The test pattern 21 . 21a - 21d generated in the output variables 13a - 13d a reaction 22 . 22a - 22d , In step 122 within the learning process 120 become the parameters 14 the internal processing chain 12 adapted to the reaction 22 . 22a - 22d best possible a target reaction 23 equivalent. That way, the test pattern becomes 21 . 21a - 21d into the AI module 1 learned that it will be recognized in later operation, if it again the input variables 11a - 11d is superimposed. If this recognition fails, a hardware defect is likely.

Alternatively or in combination, in step 150 the reaction 22 . 22a - 22d for reference 24 be stored. Will the test pattern 21 . 21a - 21d later submitted again and the reaction changes 22 . 22a - 22d , can thus also without training in the processing chain 12 even be closed to a hardware defect.

2 shows how by using two exemplified test patterns 21a and 21b two hardware components exemplified here 1a and 1b can be monitored for function. The internal processing chain 12 consists, exemplified here, of two parts 12a and 12b , each on different hardware components 1a and 1b of the AI module 1 to run. The first test pattern 21a is chosen so that it is essentially from the first part 12a the processing chain 12 , and thus on the first hardware component 1a of the AI module 1 , for a first reaction 22a is processed. The second test pattern 21b is chosen so that it is essentially from the second part 12b the processing chain 12 , and thus on the second hardware component 1b of the AI module 1 , to a second reaction 22b is processed. The two reactions 22a and 22b will be at the output of the AI module 1 merged and as a contribution to the output variables 13a - 13d output.

3 shows an embodiment of the method 100 for functional testing of the AI module 1 , Analogous to 1 will be in step 205 first a test pattern 21 . 21a - 21d selected in the expected input variables 11a - 11d of the AI module 1 does not occur. In step 210 will this test pattern 21 . 21a -21d the real inputs 11a - 11d superimposed. The input variables 11a -11d will be in step 220 through the AI module to be tested 1 with its internal processing chain 2 in output quantities 13a - 13d translated. In these outputs 13a - 13d is a post 22 . 22a - 22d included on the test pattern 21 . 21a - 21d declining.

This reaction 22 . 22a - 22d will be in step 230 determined and can now, individually or in combination, be further processed in various ways. According to step 240 will the reaction 22 . 22a - 22d with a target reaction 23 compared that to the test pattern 21 . 21a - 21d corresponds. According to step 250 will the reaction 22 . 22a - 22d with a previously stored reference 24 compared. According to step 260 will the reaction 22 . 22a - 22d , and thus the result of the test, a specific hardware component 1a . 1b assigned to the corresponding test pattern 21 . 21a - 21d within the AI module 1 has been processed substantially.

4 shows an example observable from a vehicle in the direction of travel scenery, the input variables 11a - 11d an AI module 1 can form. The scenery includes a roadway 31 , a traffic sign 32 and a vehicle 33 , The same scene, which originates for example from camera data, are different test patterns according to the partial images ad 21a - 21d superimposed, which differ in their positions within the scenery.

Is now, for example, the parallelized and meshed processing chain 12 of the AI module 1 organized so that different image parts of different hardware components 1a . 1b of the AI module 1 are processed, so by the use of different test patterns 21a - 21d Ensured that all these hardware components 1a . 1b be included in the functional test.

Here comes the test pattern 21a - 21d each in the expected input variables 11a - 11d , ie in the real scenery, not before. In this way it is ensured that not about the traffic sign 32 falsely as a test pattern 21a - 21d is identified, so that may be a reaction of the AI module 1 downstream control system on the traffic sign 32 omitted.

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 8468109 B2 [0003]
• US 8953436 B2 [0004]
• US 7937343 B2 [0005]
• US 6473746 B1 [0006]

Claims (15)

Method (100) for training an artificial intelligence module, KI module (1), which converts input variables (11a-11d) into output variables (13a-13d) through a parameterized internal processing chain (12), wherein the parameters (14) the processing chain (12) are determined (120) such that learning values (15a-15d) of the input quantities (11a-11d) are translated to the matching learning values (16a-16d) of the output quantities (13a-13d) ( 130), characterized in that the learning values (15a-15d) of the input quantities (11a-11d) are superimposed with a predetermined test pattern (21, 21a-21d) (110) and a response (22, 22a-22d ) is determined (140), which causes the test pattern (21, 21a-21d) in the output values (13a-13d) supplied by the internal processing chain (12). Method (100) according to Claim 1 , characterized in that the parameters (14) of the processing chain (12) are adapted (122) such that the reaction (22, 22a-22d) caused by the test pattern (21, 21a-21d) conforms to a desired reaction ( 23) approximates. Method (100) according to one of Claims 1 to 2 , characterized in that the reaction (22, 22a-22d) caused by the test pattern (21, 21a-21d) is stored as reference (24) for a later functional test of the KI module (1) (150). Method (100) according to one of Claims 1 to 3 , characterized in that at least two test patterns (21a-21d) are selected such that the first test pattern (21a) is predominantly processed by a part (12a) of the processing chain (12) which is mounted on a first hardware component (1a ) of the KI module (1), while the second test pattern (21b) is predominantly processed by a part (12b) of the processing chain (12) which is located on a second hardware component (1b) of the KI module (1). is performed. Method (200) for functional testing of an artificial intelligence module, KI module (1), which converts input variables (11a-11d) into output variables (13a-13d) through a parameterized internal processing chain (12), characterized in that, during operation of the KI module (1), a predefined test pattern (21, 21a-21d) is superimposed on the real input variables (11a-11d) (210) and a reaction (22, 22a-22d) is determined (230) which causes the test pattern (21, 21a-21d) in the outputs (13a-13d). Method (200) according to Claim 5 , characterized in that the reaction (22, 22a-22d) is compared with a predetermined desired response (23) (240). Method (200) according to one of Claims 5 to 6 , characterized in that the reaction (22, 22a-22d) is compared with a previously determined and stored reference (24) (250). Method (200) according to one of Claims 5 to 7 , characterized in that from the reaction (22, 22a-22d) a statement about the correct function of one or more hardware components (1a, 1b) is derived (260) on which that part (12a, 12b) of the processing chain (12) of the KI module (1) from which the test pattern (21, 21a-21d) is predominantly processed. Method (100, 200) according to one of Claims 1 to 8th , characterized in that a test pattern (21, 21a-21d) is selected (105, 205), which in the operation of the KI module (1) not in the real input variables (11a-11d) of the KI module (1 ) occurs. Method (100, 200) according to one of Claims 1 to 9 , characterized in that the input variables (11a-11d) of the KI module (1) comprise image data, wherein the output variables (13a-13d) of the KI module (1) provide information about the occurrence of at least one object (31-33) in the image data. Method (100, 200) according to one of Claims 1 to 10 , characterized in that the input variables (11a-11d) of the KI module (1) audio data, wherein the output variables (13a-13d) of the KI module (1) include a statement about the occurrence of at least one specific noise in the audio data , Method (100, 200) according to one of Claims 1 to 11 , characterized in that an AI module (1) is selected whose internal processing chain (12) comprises at least one neural network. Artificial intelligence module, KI module (1), which converts input variables (11a-11d) into output variables (13a-13d) through a parameterized internal processing chain (12), characterized in that the internal processing chain (12) is designed to at least one test pattern (21, 21a-21d) presented in the input quantities (11a-11d) can be translated into a predetermined response (22, 22a-22d) in the output quantities (13a-13d), the test pattern (21 , 21a-21d) in the operation of the KI module (1) does not occur in the real input variables (11a-11d) of the KI module (1). Computer program containing parameters (14) for the internal processing chain (12) of an AI module (1) which, when loaded into an AI module (1), converts the AI module (1) into an AI module (1). 1) after Claim 13 upgrade. A computer program containing machine readable instructions which, when executed on a computer, cause the computer to perform a method according to any one of Claims 1 to 12 and / or an AI module (1) to an AI module (1) Claim 13 upgrade.

Images