Fuzzy parameterized KI module and operation method

Abstract

An artificial intelligence intelligent module, KI module, is configured to process one or more input parameters into one or more output parameters by an internal processing chain, wherein the internal processing chain is specified by one or more parameters, and wherein the KI module is configured to determine the one or more parameters from one or more stored values ​​stored in a memory, wherein a distribution module is provided, the distribution module is configured to extract single values ​​from one or more statistical distributions and to determine the one or more parameters therefrom, wherein at least one statistical characteristic parameter of each statistical distribution is related to at least one stored value. Method for eliminating stored values. Method for discretizing and compressing stored values. Method for evaluating the uncertainty of output variables.

Description

Technical Field

[0001] The invention relates to an artificial intelligence module, a KI module, in which the uncertainty of statements made can be quantified and at the same time learning parameters can be stored in a space-saving manner on hardware. Background Art

[0002] The operation of motor vehicles in road traffic requires reliable visual recognition of a large number of objects, such as traffic signs, lane markings or other traffic participants. In order to enable fully or partially automated driving of the vehicle, it is necessary to artificially simulate the recognition that has hitherto been undertaken by human drivers. Due to the high dimensionality of the problem, artificial intelligence modules (KI modules) are used for this purpose, which classify images according to the objects contained therein. These KI modules can in particular include neural networks.

[0003] For operational safety, it is important to at least identify situations in which the detection does not appear to function reliably in order to be able to take further suitable measures. Comprehensive testing of all possible such situations in the training phase of the KI module is not practical.

[0004] DE 10 2005 050 577 A1 discloses a method for automatically testing a neural network of a control unit. Summary of the invention

[0005] Within the scope of the present invention, an artificial intelligence module, namely a KI module, has been developed. The KI module is designed to process one or more input variables into one or more output variables via an internal processing chain. The internal processing chain is defined by one or more parameters. In this case, the term "parameter" is not limited to scalar values, but also includes vectors, matrices or functions from a function space, such as filters.

[0006] The KI module is designed to determine the one or more parameters from one or more stored values ​​stored in the memory. This means that a calculation rule exists such that a specific stored value stored in the memory is assigned to the value of the parameter.

[0007] A distribution module is provided which is designed to extract a single value from each of one or more statistical distributions and to determine the one or more parameters therefrom, wherein at least one statistical characteristic variable of each statistical distribution is dependent on at least one stored value.

[0008] The stored value may have been acquired, for example, during the learning process of the KI module and therefore corresponds to the value of the learned learning variable. However, the stored value may also have been previously input as a program. Combinations are also possible. Thus, for example, a stored value previously input as a program can be updated during a later renewed learning process of the KI module. The lossy compression and discretization of the stored values ​​described below is possible in a manner independent of their original source.

[0009] This means that in repeated cycles of the distribution module one and the same configuration of the stored values ​​is no longer always assigned to the same value of the parameter. Instead, the acquired values ​​of the parameter are statistically distributed around the value predefined by the stored values.

[0010] The training of the KI module is usually carried out with a large number of learning values ​​of the input variables, for which the learning values ​​of the output variables are respectively known. The stored values ​​from which the distribution module is parameterized are optimized step by step so that the learning values ​​of the input variables are mapped to the learning values ​​of the output variables with sufficient accuracy for the respective application. It is then assumed that the values ​​of the input variables occurring in real operation are also mapped to values ​​of the output variables that are meaningful for the application.

[0011] The optimization of the stored values ​​can be performed, for example, in such a way that the parameters generated by the distribution module minimize the value of a performance function (“loss function”). Such an optimization can, for example, seek a local or global optimum and / or a stable state. Introducing fuzziness into the determination of the parameters from the stored values ​​has the following effect: As described below, the stored values ​​can be compressed and discretized in a lossy manner without this leading to excessive artifacts in the output variable. In addition, the optimization can converge to stored values ​​in which the effect of the imposed fuzziness on the finally acquired value of the output variable is reduced.

[0012] This in turn indirectly has the effect that the processing by the KI module in real operation is more resistant to disturbances in the input variables, such as noise. During training, the KI module is situated on a vibration table and learns the vibrations.

[0013] The mandatory fuzzification of parameters opens up the possibility of discretizing the stored values ​​in a meaningful way and compressing them in a lossless way in this way. Until now, the stored values ​​have been stored with high precision, for example as 32-bit floating point numbers. When the mandatory fuzzification of the limiting factor for precision is now a parameter in the internal processing chain, the relative deterioration of the precision of the stored values ​​is insignificant. Training under mandatory fuzzification is therefore a statistically motivated method to find out where storage space can be saved for the stored values ​​without a significant loss of quality in the end result.

[0014] Finally, the values ​​of the output variable also scatter when the same value of the input variable is supplied to the KI module multiple times. This scatter is a direct, easily understandable measure of the uncertainty of the statements made by the KI module.

[0015] In a particularly advantageous embodiment of the invention, the distribution module contains at least one table, referred to below as a codebook table, having a plurality of elements and is designed to determine the one or more parameters as linear combinations of these elements. The coefficients of the linear combination are single values ​​extracted from a statistical distribution. If, for example, the parameter is a scalar, then the scalar itself is not trained as a stored value, but therefore a plurality of stored values ​​are given, like the elements in the codebook table.

[0016] This does not lead to the desired saving of storage space, but instead to a doubling of the space requirement. However, it is already known that statements about how large the entropy of the parameters is can be derived from the representation of the parameters as a linear combination. If the storage values ​​are determined by the training such that the distribution over the following parameters has a high entropy, for example a mean distribution, these parameters can be set to zero.

[0017] In a particularly advantageous embodiment of the invention, the internal processing chain is a deep neural network. In such a neural network, between each two layers for each pair of two neurons, the connection between the two neurons should be equipped with weights as parameters, wherein one neuron belongs to one layer and the other neuron belongs to the other layer. The number of parameters required is correspondingly large, and therefore the demand for storage space is also correspondingly large. At the same time, it can be seen from the corresponding applications that not all weights are equally important for the final result. By decomposing the scalar weights into linear combinations composed of different elements, a statement about the entropy of the parameters is now provided, making it possible to explain the rationale for the selective lossy compression of the weights.

[0018] Advantageously, the codebook table contains an odd number of digits for this purpose, which are arranged in a symmetrical manner about the average value. For example, the codebook table can contain five digits, which are arranged in a symmetrical manner around 0. The distribution module can then include a separate codebook table for each layer of the neural network. However, the codebook table is also meaningful for other contents, ideally such contents in which the expected value is zero. It is not mandatory for this that the codebook table contains an odd number of digits. A symmetrical arrangement is also optional.

[0019] In another particularly advantageous embodiment of the invention, the internal processing chain is a neural network with at least one convolutional layer, i.e. a "convolutional neural network". Such a network is particularly suitable for reducing the dimensionality of high-dimensional data, in particular image data, provided as input variables by means of the stepwise application of filters, in order to ultimately achieve a classification of the image or an object contained therein.

[0020] Advantageously, the codebook table in such a neural network contains filters. In this context, a filter is understood in particular to be an operator that recognizes and / or emphasizes one or more features in the input data. The filter can advantageously generate output data that have a smaller dimension than the input data; however, this is not mandatory.

[0021] For example, the codebook table can contain between 7 and 11 filters. In contrast to deep neural networks, the decomposition into linear combinations of a plurality of filters does not necessarily lead to a doubling of the storage space requirement: the filters in the codebook table can be stored, for example, in the form of a matrix with integer values, which requires significantly less storage space than a matrix with real values. Only the stored values ​​and the codebook table should still be stored in a real-valued manner, from which the distribution module forms the coefficients of the linear combination. The distribution module can include a separate codebook table for each convolutional layer of the neural network.

[0022] In another particularly advantageous embodiment of the invention, the memory for the storage values ​​is designed to record integer indices. At the same time, at least one lookup table is provided for converting each index into one or more discrete storage values. In this way, the corresponding storage values ​​can be stored in a losslessly compressed manner, which storage values ​​do not have to be set to zero from the beginning.

[0023] If, for example, the codebook table contains n elements and the lookup table contains m entries, the discretization of the stored values ​​thus results in that only n*log2(m) bits per stored value must be stored at the conventional 32-bit locations. In the case of n=5 and m=2, for example, 5 bits per stored value are still required, which corresponds to a compression to less than one sixth of the original storage space requirement. In addition, there is a one-time storage requirement for the codebook table and the lookup table.

[0024] When the lookup table already contains a vector consisting of multiple values ​​at the position of the scalar value, it becomes even larger, however only log2(m) bits of storage space are still required for each stored value.

[0025] According to what has been described above, the invention also relates to a method for operating the described KI module. In this method, a set of stored values ​​is determined by using the forced fuzziness introduced by the distribution module in such a way that the KI module converts a large number of learned values ​​of the input variable into the associated learned value of the output variable with sufficient accuracy. Subsequently, the distribution module is repeatedly used to determine the individual values ​​of the parameter, using the acquired stored values. As soon as the fluctuation amplitude of the parameter exceeds a predefined threshold value, the associated stored value is set to zero.

[0026] As explained above in the context of linear combinations of elements of codebook tables, strongly fluctuating parameters ultimately contain only very weak information. Accordingly, only a small amount of information is lost when the associated stored values ​​are reset to zero. On the other hand, a disproportionately large amount of storage space is saved by zeroing the stored values ​​("pruning").

[0027] The invention also relates to another method for operating the described KI module. This method can be applied alternatively to the last-mentioned method or also in combination therewith. As in the last-mentioned method, a set of stored values ​​is determined in the following manner, using the forced fuzziness introduced by the distribution module: the KI module converts a large number of learning values ​​of the input variables into the corresponding learning values ​​of the output variables with sufficient accuracy. Subsequently, these stored values ​​are clustered into clusters. For each cluster, the stored values ​​are recorded in a lookup table. In combination with the last-mentioned method, this method is only applied to stored values ​​that have not yet been set to zero.

[0028] It is already known that clustering, for example using the k-means algorithm, is a statistically well-motivated and at the same time very flexible way and method for discretizing stored values. In particular, a fixed grid is abandoned. It is therefore absolutely possible that the cluster is significantly denser in a certain range of the parameter space than in other ranges of the parameter space.

[0029] In a particularly advantageous embodiment of the invention, the following statistical distribution is a continuous relaxation of a discrete random variable, from which individual values ​​are extracted in the determination of the parameters, wherein the continuous relaxation is also referred to as a "concrete distribution". This concrete distribution X ~ Concrete(α,γ) of the (n-1)-dimensional random variable X is defined in such a way that its density D(x) is given by:

[0030]

[0031] where α ∈ (0,∞) n and γ ∈ (0, ∞). α is the location of the distribution and γ is the temperature parameter. This specific distribution is distinguished by the fact that its density can be written in closed form and it can be easily reparameterized.

[0032] In this case, the temperature parameter γ is advantageously selected during training for the purpose of creating the desired blurring such that a linear combination of the elements of the codebook table always contains at least two coefficients different from zero, ie is blurred. The corresponding stored value for the position α is estimated.

[0033] According to the above description, the invention also relates to a further method for operating a KI module, in which the same value of the one or more input variables is supplied to the KI module several times, and the scatter of the obtained values ​​of the one or more output variables is evaluated as a measure of the uncertainty of the statements made by the KI module.

[0034] It has been found that forced ambiguity in the parameters of the internal processing chain of the KI module represents a good test for this purpose: whether the value of the input variable is mapped essentially without doubt to a similar value of the output variable in each cycle or whether this assignment is as if "unresolved" and can already be caused by equal weighting due to the ambiguity.

[0035] Regardless of whether forced fuzziness is also exploited for space-efficient storage of stored values, the newly acquired information about uncertainty alone can already be exploited multiple times.

[0036] Thus, for example, in at least partially automated driving, information about the vehicle environment acquired by multiple sensor types can be fused in a motivated manner by weighting the different contributions in accordance with the respective uncertainties. Based on the fused information, for example, at least one actuator and / or physical warning device of the vehicle can then be actuated.

[0037] The desired trajectories of other road users also often factor into the trajectory planning of the at least partially automated vehicle or robot. By taking into account the corresponding uncertainty of the desired trajectory, the planning can be made more robust and sudden maneuvers can be avoided. If the neural network predicts, for example, that an oncoming vehicle will continue straight ahead at an intersection, the high uncertainty of this prediction can, for example, lead to the vehicle or robot being slowed down as a precaution. For this purpose, for example, at least one actuator and / or physical transmission device of the vehicle can be controlled to output corresponding instructions to the driver.

[0038] The uncertainty of a particular prediction can also be taken into account early in the route planning in order to avoid unnecessary detours. If, for example, it is predicted only with relatively high uncertainty that a particular parking space will still be free at a certain time, another parking possibility can be activated from the outset in which the uncertainty is lower. For this purpose, for example, at least one actuator and / or physical transmission device of the vehicle can be activated in order to output a corresponding command to the driver.

[0039] In another particularly advantageous embodiment of the invention, an image is selected as the input variable and a classification of the image is selected as the output variable. In imaging, there are multiple sources of uncertainty. Therefore, the desired information is not always present with the same intensity in the image and can also be completely or partially superimposed due to quality fluctuations in the image recording.

[0040] When, for example, a medical image is classified according to the presence or degree of manifestation of a disease, the doctor can therefore be informed of the certainty with which the decision was made, in addition to the decision of the algorithm. With a certain uncertainty, the image can be reported as "undecided", for example. For this purpose, a physical transmission device can be controlled, for example, to output a report to the doctor. The doctor then knows that he should obtain a better-quality image or should choose another imaging method.

[0041] In at least partially automated driving, it may happen that the vehicle encounters a specific traffic sign for the first time. Thus, the traffic sign may be changed by the legislator, for example, or a specific traffic sign that only appears in foreign countries is unknown. In such a case, a warning may be output that the classification is uncertain. It may also happen that traffic signs have been deliberately manipulated in such a way that the neural network is confused, such as by coloring or applying labels (“adversarial examples”). For example, a physical transmission device may be actuated in order to output a warning to the driver. For example, a switching device may also be actuated to transfer control to the driver, and / or at least one actuator of the vehicle may be actuated, for example, to preventively slow down the vehicle.

[0042] Similarly, in automated quality control, for manual post-control, a test object is selected from the image starting from a certain uncertainty. For this purpose, for example, a marking device can be actuated and / or a conveyor device can be actuated, for example, in such a way that the test object is conveyed to a work station for manual post-control, wherein the marking device marks the test object accordingly. In automated facial recognition, a person can be bypassed to a human supervisor starting from a certain uncertainty, for example by correspondingly actuating a visual guidance system, and / or an exit can be automatically locked, for example by correspondingly actuating a door system.

[0043] In another particularly advantageous embodiment of the invention, the control variable is output to a recording device for recording the input variable in order to change at least one physical parameter of the recording so that the uncertainty of the statement is reduced after the new recording. The recording device is thus controlled in a modified manner in order to record the input variable with a better quality with regard to the evaluation by recording with the modified physical parameter.

[0044] Thus, for example, the intensity and / or color of the lighting can be changed in the image recording, whereby parts of the image can then be introduced into saturation, but the dependent information is more clearly highlighted. Sensors and data processing can also be operated in an energy-saving mode under normal circumstances, while in questionable situations more computing time and energy consumption are invested in more precise identification. In at least partially automated driving, additional sensors can be connected in real time in questionable situations, which should not be in operation continuously because their use is associated with high energy consumption or high wear, for example. If the smoke detector does not detect a certain signal with its photoelectric sensor, it can also switch on a laser or a camera, for example, in order to avoid false alarms.

[0045] On mobile embedded devices, small neural networks can be run that consume fewer resources. For example, stored values ​​can additionally be compressed according to the invention. Smaller neural networks can then be used for most requests, and only in doubtful cases do input variables need to be forwarded to larger neural networks in the "cloud".

[0046] In another particularly advantageous embodiment of the invention, the uncertainty of the statement exceeding a predetermined threshold value is evaluated as follows: an anomaly is present in the input variable. In this way, for example, "adversarial examples" can be identified, that is, intentional manipulation of traffic signs, for example with the goal that the traffic sign is not recognized by the at least partially automated vehicle or is recognized as another sign.

[0047] The described innovations can be implemented completely or partially in the software of an embedded system, a KI module and / or a computer. Changes to the hardware are not mandatory in order to achieve the advantages provided by the invention. The software is in this respect a product that can be purchased independently, which can also be marketed as an update or expansion (Add-On) to an existing system. Therefore, the present invention also relates to a computer program that contains machine-readable instructions that, when executed on a computer and / or an embedded system and / or a KI module, upgrade the computer and / or the embedded system and / or the KI module to a KI module according to the invention and / or promote the implementation of the method according to the invention. Likewise, the present invention relates to a machine-readable data carrier having such a computer program.

[0048] Further measures for improving the invention are described below in conjunction with the description of preferred exemplary embodiments of the invention with reference to the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 An exemplary sketch of a KI module 1 is shown;

[0050] Figure 2 An exemplary schematic diagram of a method 100 for operating a KI module 1 is shown, wherein meaningless stored values ​​15 are eliminated by means of the KI module;

[0051] Figure 3 An exemplary sketch of a method 200 for operating a KI module 1 is shown, with which storage values ​​15 are discretized and compressed in a lossy manner;

[0052] Figure 4 An exemplary schematic diagram is shown for operating a method 300 for operating a KI module 1 , with which uncertainty 19 of a statement of the KI module 1 is evaluated. DETAILED DESCRIPTION

[0053] according to Figure 1An input variable 11 is recorded by a recording device 3 , whose physical recording parameters can be controlled via a control variable 31 . An internal processing chain 12 of the KI module 1 maps the value of the input variable 11 to a value of an output variable 16 .

[0054] The internal processing chain 12 is specified by means of parameters 13. Figure 1 As shown in , the internal processing chain 12 is a deep neural network, and the parameters correspond to the following weights, with which the connections between neurons of adjacent layers are equipped. These parameters 13 are partially or completely redefined by the distribution module 2 for each application.

[0055] In the memory 14 of the KI module 1, an integer index 15a is stored, under which the discrete storage value 15 is stored in the lookup table 17. The term "discrete" does not mean that the value must be an integer, but only that only a limited number of values ​​are used. The storage value can also be a floating point number, for example.

[0056] In the distribution module 2, individual values ​​22 are called from a statistical distribution 21, which is parameterized with stored values ​​15. These individual values ​​22 are used as coefficients to linearly combine the elements 23a-23c stored in a codebook table 23 to form the required parameters 13. The internal processing chains are configured accordingly to the values ​​of the parameters 13. This specifies which values ​​of the output variable 16 the values ​​of the input variable 11 are mapped to.

[0057] The parameters 13 determined by means of the distribution module 2 are valid only once. If the same value of the input variable 11 is again supplied to the KI module 1, the parameters 13 are recalculated and other values ​​are assumed, since they are random variables. Accordingly, the values ​​of the output variable 16 have a scatter 18, from which the uncertainty 19 of the statement made by the KI module 1 can be derived again.

[0058] For example, the weight w of layer s in a neural network for the connection from neuron i to neuron j is s，jj can be shown as a linear combination of the elements in the codebook table 23, which contains five numbers, k=[-0.3, -0.1, 0, 0.1, 0.3]:

[0059]

[0060] Among them, a s,ij,k are single values ​​from statistical distribution 21 and the sum of these single values ​​is 1: .

[0061] a s,ij,k In this example, the concrete distribution is known at the location corresponding to the stored value 15. The blurring of the distribution is dependent on the temperature parameter γ of the concrete distribution.

[0062] Figure 2 An exemplary embodiment of the method 100 is shown, with which meaningless stored values ​​15 are eliminated. First, in step 110, the internal processing chain 12 of the KI module 1 is trained, which is decomposed into substeps 111 to 116. In substep 111, the values ​​of the codebook table 23 and the stored values ​​15 are initialized. In substep 112, a minibatch is selected from the learned values ​​11a of the input variable 11 and from the learned values ​​16a of the output variable 16. In substep 113, individual values ​​22 are extracted from the statistical distribution 21 parameterized with the stored values ​​15, with which parameters 13 for the configuration of the internal processing chain 12 are determined in substep 114.

[0063] In substep 115 , the numerical deviation of output variable 16 generated by internal processing chain 12 from learned value 11a of input variable 11 from learned value 16a of output variable 16 is determined. Thus, in substep 116 , stored value 15 is updated with the goal of minimizing this deviation.

[0064] If step 110 is to converge to a final valid stored value 15, the values ​​13a-13c of the parameter 13 are repeatedly determined in step 120 using the distribution module 2. Since the parameter 13 is a random variable, its individual values ​​13a-13c differ from cycle to cycle. The fluctuation amplitude associated with this is evaluated in step 130. As long as the fluctuation amplitude exceeds a predefined threshold value (true value 1), the associated stored value 15 is set to zero in step 140.

[0065] Figure 3 An embodiment of a method 200 is shown, by which a stored value 15 is discretized and compressed in a lossy manner. The method 200 may begin with step 210, which is similar to Figure 2 The same as step 110 in Figure 3 However, the method 200 can also be connected to the method 100 in such a way that the final result of the stored value 15 from the method 100 is supplied to step 220 of the method 200 .

[0066] In step 220, the stored values ​​15 are grouped into clusters 15a-15c. In step 230, the stored values ​​15 are converted into a lookup table 17 for each cluster 15a-15c, respectively.

[0067] If, for example, Figure 1 In the numerical example described above, the value p has been learned for the stored value s,ij =[0.22, 0.32, 0.7, 0.77, 0.65], using these values ​​to obtain the value for the number c from the codebook table 23 s,k =The random coefficient a of the linear combination of [-0.3, -0.1, 0, 0.1, 0.3] s,ij,k The exemplary clustering process then results in two clusters to which these values ​​[0.2, 0.7] for the lookup table 17 are assigned. The mentioned values ​​p s,ij The values ​​are discretized in the lookup table 17 by assigning the corresponding index [0,0,1,1,1] to the values. Each such index 15a therefore only requires 1 bit of storage space.

[0068] When parameter 13 is required, then the values ​​a s,ij,k is always drawn from a specific distribution. However, for an efficient implementation on hardware one can also sample a from the class distribution s,ij,k , where then for each parameter 13 the value a s,ij,k Only one of them is equal to 1 and all other values ​​are zero. This is an approximation of a specific distribution.

[0069] Figure 4 An exemplary embodiment of the method 400 is shown, with which the uncertainty 19 of the statements of the KI module 1 is evaluated. In optional step 305, an image is selected as input variable 11, and in optional step 310, a classification of the image is selected as output variable 16. Instead of an image, any other higher-dimensional input data can also be selected.

[0070] In step 320 , the same value of input variable 11 is supplied to KI module 1 multiple times. In this process, individual values ​​16 b - 16 d of output variable 16 are formed, which are scattered with respect to one another. This scatter 18 is evaluated in step 330 as a measure of uncertainty 19 for the statement of KI module 1 .

[0071] In step 340 , this uncertainty 19 is used in order to thereby determine a manipulated variable 31 for recording device 3 with the goal of qualitatively improving the physical recording of input variable 11 and subsequently reducing uncertainty 19 of output variable 16 obtained therefrom.

[0072] Alternatively or in combination therewith, it is checked in step 350 whether the uncertainty 19 exceeds a predefined threshold value. If this is the case (true value 1), it is concluded in step 360 that an anomaly is present in the input variable 11. This anomaly is notified to the user in step 365.

Claims

1. An artificial intelligence (KI) module (1) for classifying images, the KI module being designed to process one or more input variables (11) into one or more output variables (16) via an internal processing chain (12), wherein an image is selected as the input variable (11) and the classification of the image is selected as the output variable (16), wherein the internal processing chain (12) is defined by one or more parameters (13), and wherein the KI module (1) is designed to determine the one or more parameters (13) from one or more stored values ​​(15) stored in a memory (14), characterized in that A distribution module (2) is provided, which is constructed to extract individual values ​​(22) from one or more statistical distributions (21) and to determine the one or more parameters (13) therefrom, wherein at least one statistical characteristic variable of each statistical distribution (21) is related to at least one stored value (15), wherein the distribution module (2) contains at least one code book table (23) having a plurality of elements (23a-23c) and is constructed to determine the one or more parameters (13) as a linear combination of the elements (23a-23c), wherein the coefficients of the linear combination are the individual values ​​(22) extracted from the statistical distribution (21).

2. The KI module (1) according to claim 1, characterized in that The internal processing chain (12) is a deep neural network.

3. The KI module (1) according to claim 2, characterized in that The codebook table (23) contains an odd number of digits, which are arranged symmetrically about a mean value.

4. The KI module (1) according to claim 1, characterized in that The inner processing chain (12) is a neural network having at least one convolutional layer.

5. The KI module (1) according to claim 4, characterized in that The codebook table (23) contains filters.

6. The KI module (1) according to any one of claims 2 to 5, characterized in that: The distribution module (2) contains a separate codebook table (23) for each layer of the neural network.

7. The KI module (1) according to any one of claims 2 to 5, characterized in that: The memory (14) for the storage values ​​(15) is designed to record integer indices (15a) and at least one lookup table (17) is provided for converting each index into one or more discrete storage values ​​(15).

8. The KI module (1) according to claim 7, characterized in that A separate lookup table (17) is provided for each layer of the neural network.

9. A method (100) for operating a KI module (1) according to any one of claims 1 to 8, comprising the following steps: A set of stored values ​​(15) is determined (110, 111-116) in such a way that the KI module (1) converts a plurality of learned values ​​(11a) of the input variable (11) into associated learned values ​​(16a) of the output variable (16) with sufficient accuracy; repeatedly determining (120) a single value (13a-13c) of the parameter (13) using the distribution module (2) using the stored value (15); and As soon as the fluctuation amplitude of the parameter (13) exceeds a threshold value which is predefined (130), the associated stored value (15) is set to zero (140).

10. A method (200) for operating a KI module (1) according to any one of claims 7 to 8, comprising the following steps: A set of stored values ​​(15) is determined (210; 111-116) in such a way that the KI module (1) converts a plurality of learned values ​​(11a) of the input variable (11) into associated learned values ​​(16a) of the output variable (16) with sufficient accuracy; The stored values ​​(15) are aggregated (220) into clusters (15a-15c); and For each cluster, the storage value (15) is recorded (230) into the lookup table (17).

11. Method (300) for operating a KI module (1) according to any one of claims 1 to 8, characterized in that The same value of the one or more input variables (11) is supplied to the KI module (1) multiple times (320) and a scatter (18) of the obtained values ​​(16b-16d) of the one or more output variables (16) is evaluated (330) as a measure of the uncertainty (19) of the statement made by the KI module (1).

12. The method (300) according to claim 11, characterized in that: A control variable (31) is output (340) to a recording device (3) for recording the input variable (11) in order to change at least one physical parameter of the recording so that the uncertainty (19) of the statement is reduced after the new recording.

13. The method (300) according to any one of claims 11 to 12, characterized in that: The uncertainty (19) of the statement that exceeds (350) a predefined threshold value is evaluated (360) as follows: an anomaly is present in the input variable (11).

14. A computer program product, wherein the computer program comprises machine-readable instructions, which, when executed on a computer and / or an embedded system and / or a KI module (1), cause the computer and / or the embedded system and / or the KI module (1) to implement the method (100, 200, 300) according to any one of claims 9 to 13.

Images