Multiple artificial learning units are coupled to the projection hierarchy.

Abstract

This invention relates to a method in a system comprising at least one second artificial learning unit and a third artificial learning unit, comprising the steps of: inputting a first input value to at least one second artificial learning unit and obtaining an output value based on the input value from the at least one second artificial learning unit; temporarily storing situation data, the situation data including the first input value and / or a second output value of the at least one second unit; using the situation data as an input value for the third artificial learning unit, the third artificial learning unit generating a third output value in response to the input value; and checking whether a second output value of the at least one second unit satisfies one or more specified conditions based on the third output value. The invention also relates to a system for performing this method.

Description

Technical Field

[0001] This invention relates to a system of two or more coupled artificial intelligence units. Background Technology

[0002] Artificial intelligence is now playing an increasingly important role in countless application areas. This was initially understood to mean any automation of intelligent behavior and machine learning. However, such systems are typically designed and trained for specific tasks. This form of artificial intelligence (AI) is often called "weak AI," and is essentially based on applied computation and algorithms to simulate intelligent behavior in a fixed domain. Examples include systems capable of recognizing certain patterns, such as safety systems in vehicles, or systems capable of learning and enforcing certain rules, such as in chess. Meanwhile, these systems are largely useless in other domains and must be completely retrained for other applications, sometimes using entirely different methods.

[0003] For the practical implementation of such artificial intelligence units, neural networks are used, among other things. In principle, these networks replicate the function of biological neurons at an abstract level. Several artificial neurons, or nodes, are interconnected, capable of receiving, processing, and transmitting signals to other nodes. For example, for each node, functions, weights, and thresholds are defined to determine whether and to what extent a signal is transmitted to the node.

[0004] Nodes are typically considered hierarchically, ensuring that each neural network has at least one output layer. Prior to this, other layers can exist as so-called hidden layers, forming a multi-layered network. Input values ​​or features can also be considered layers. Connections between nodes in different layers are called edges, and these nodes are usually assigned a fixed processing direction. Depending on the network topology, it can be specified which node in one layer links to which node in the next layer. In this case, all nodes can be connected; however, for example, due to the learned weights with a value of 0, the signal cannot be further processed through a particular node.

[0005] The processing of signals by a neural network can be described using various functions. Below, we describe this principle for a single neuron or node in the neural network. From several different input values ​​arriving at the node, the network input is formed by a propagation function (which is also the input function). Typically, this propagation function consists of a simple weighted sum, where a corresponding weight is given for each input value. However, in principle, other propagation functions are also possible. In this case, the weights can be specified as the network's weight matrix.

[0006] A threshold-dependent activation function can be applied to the network input of nodes formed in this way. This function represents the relationship between the network input and the neuron's activity level. Various activation functions are known, such as a simple binary threshold function whose output is zero below the threshold and an identity above the threshold; a sigmoid function; or a piecewise linear function with a given slope. These functions are specified in the design of the neural network. The output of the activation function forms the activation state. Optionally, an additional output function can be specified, which is applied to the output of the activation function and determines the final output value of the node. However, often the result of the activation function is simply passed directly as the output value; that is, the identifier is used as the output function. Depending on the terminology used, the activation function and the output function can also be combined into a transfer function.

[0007] Then, the output value of each node is passed to the next layer of the neural network as the input value for each node in that layer, where the corresponding steps are repeated to process the data using the node's functions and weights. Depending on the network topology, there may also be backward edges between previous or subsequent layers, resulting in a recurrent network.

[0008] Conversely, a network can change the weights of each input value, thereby adjusting the output value and the overall function of the network; this is considered the "learning" of a neural network.

[0009] Therefore, error backpropagation is typically used in networks, which compares the output value with the expected value and uses this comparison to adapt the input value to minimize the error. Error feedback can then be used to adjust various network parameters accordingly, such as the step size (learning rate) or the weights of the input values ​​at nodes. Similarly, the input values ​​can be recalculated.

[0010] The network can then be trained in training mode. The learning strategy used is also crucial to the possible applications of the neural network. The following variants are specifically distinguished:

[0011] In supervised learning, an input pattern or training dataset is given, and the network's output is compared with the expected value.

[0012] Unsupervised learning leaves the discovery of correlations or rules to the system, thus only specifying the patterns to be learned. An intermediate variant is partially supervised learning, where datasets without predefined classifications can also be used.

[0013] In reinforcement learning or Q-learning, agents are created that can accept rewards and penalties for their behavior, and their behavior is adjusted by trying to maximize the rewards obtained.

[0014] A key application of neural networks is classifying input data or inputs into certain categories or classes, i.e., identifying correlations and assigning values. These classes can be trained on known data and are at least partially predefined, or they can be developed or learned independently by the network.

[0015] The basic operations and further specific details of such neural networks are well known in the field, for example, R. Schwaiger, J. Steinwender, Neuronal Netze programming with Python, Rheinwerk Computing, Bonn 2019.

[0016] General-purpose AI systems, if not trained for a single specific task, result in high-dimensional spaces, thus requiring exponentially growing training and testing datasets. Real-time responses quickly become impossible. Therefore, there is a general trend towards reducing the dimensionality and complexity of such systems. Different approaches are currently being taken. For example, complexity can be reduced by linking datasets, reducing degrees of freedom, and / or by providing the system with known knowledge. As another approach, related or interdependent datasets can be at least partially separated, for example, through methods such as principal component analysis. Data that is either non-prominent or prominent during network training can be eliminated by applying filtering methods to features, such as statistical tests like the chi-square test. Finally, the selection of training data itself can be addressed as an optimization problem within the AI ​​network. This involves combining training data in a way that allows new networks to be trained as quickly as possible.

[0017] More advanced methods include so-called "convolutional neural networks," which apply convolutions in at least one layer of a multi-layered fully connected network, rather than simple matrix transformations. For this purpose, the so-called "Deep-Dream" method is known, particularly in the field of image recognition, where the weights remain optimal during network training, but instead, the input values ​​(e.g., the input image) are modified to depend on feedback loops that depend on the output values. Thus, for example, content that the system believes it can recognize is faded out. The name refers to the creation of dreamlike images in this process. This allows us to trace the internal processes of the neural network and its direction.

[0018] It is evident that these methods still show significant differences from human intelligence. While databases, text files, images, and audio files can in principle be compared to how facts, language, speech logic, sounds, images, and sequences of events are stored and processed in the brain, a significant difference in human intelligence, for example, is that it links all this data with sensory and unconscious “soft” categorizations. Summary of the Invention

[0019] According to the present invention, a method for a system of coupled artificial intelligence units and a system for performing the method are proposed. Advantageous embodiments are the subject matter described below.

[0020] Specifically, a method is proposed in a system of at least a second and a third artificial intelligence unit, comprising: inputting a first input value to the at least second artificial intelligence unit and obtaining an output value from the at least second artificial intelligence unit based on the input value; temporarily storing situation data, the situation data including at least one of the first input value and a second output value from the at least second unit; using the situation data as an input value to the third artificial intelligence unit, the third artificial intelligence unit generating a third output value in response to the input value; and checking, based on the third output value, whether a second output value from the at least second unit satisfies one or more predetermined conditions. In this way, one of the units can act as a verification unit, which evaluates, monitors, or otherwise verifies solutions from one or more other units in the system, particularly for compliance with certain constraints that can be predetermined or formed in the verification unit.

[0021] Therefore, if the check shows that the output value does not meet at least one predetermined condition, the second output value can be discarded. Alternatively, the method may also include deciding whether to permanently store at least a portion of the current situation data, and transferring the situation data to be permanently stored to long-term storage.

[0022] According to an embodiment, the decision on whether to permanently store at least a portion of the current situation data may depend on the result of checking whether the second output value meets one or more predetermined conditions.

[0023] Optionally, the stored situation data may also include time information for storing input and / or output values, wherein, for example, the time information may include one or more time intervals to which one or more input and / or output values ​​are assigned. In this way, input and / or output values ​​can be used to form one or more time series, which can be processed together and allow for classification of situations, events, or other data.

[0024] A time period for temporarily storing situational data can be predetermined, and at least one AI unit can be configured to set or change the time period. For example, this could be a third unit that can change the time span based on its output value or other evaluations, but other AI units in the system may also shorten or lengthen the predetermined time span. Similarly, the time span can be flexibly specified, or it can have different sizes for different data.

[0025] According to an exemplary embodiment, the method further includes comparing the current output value of at least one second unit with an output value stored in situational data. For example, the comparison may be performed by a third unit. Such a comparison allows for recourse to previous output values, i.e., previous solutions, optionally along with associated input values, and thus can simulate experience-based learning. Furthermore, the comparison of output values ​​can be used to evaluate the quality of the output values ​​or determine whether the current solution is worse or better when one or more predetermined conditions are met.

[0026] The system may further include a first artificial intelligence unit, to which the first input value may subsequently be input and to which a first output value of the first artificial intelligence unit may be output. Based on the output values ​​of the first and / or third artificial intelligence units, one or more modulation functions may then be formed, which may then be applied to one or more parameters of a second artificial intelligence unit, the one or more parameters affecting the processing of the input value and the acquisition of the output value in the second artificial intelligence unit. Therefore, the processing of the input value in the second unit can be affected without directly specifying the parameters themselves.

[0027] In some embodiments, each AI unit may have an associated classification memory, wherein each AI unit classifies input values ​​into one or more classes stored in the classification memory when generating output values, wherein each class is constructed in one or more subordinate levels, and wherein the number of classes and / or levels in the classification memory of the first or third AI unit is less than the number of classes and / or levels in the classification memory of the second AI unit. This asymmetry between the memories of different units ensures that each unit performs a specific task. For example, the first unit may be implemented as a fast or coarse classification unit, while the second unit, with more complex memory, may be implemented as a deep analysis unit.

[0028] In any embodiment, one or more artificial intelligence units may include neural networks. All artificial intelligence units of the system may also be neural networks. However, depending on the nature of the function performed, one or more units may implement other forms of machine learning. For example, the first and second units may be implemented as neural networks, while the third unit does not form a neural network.

[0029] The input values ​​input to at least the first unit and / or the second unit can be measurements sensed by one or more sensors, data sensed by the user interface, data retrieved from memory, data received via a communication interface, and / or data output by the computing unit, or any combination of these possibilities.

[0030] A system is also proposed, comprising two or more artificial intelligence units and means (e.g., the sensor, interface, or others) for sensing input values ​​input to the first artificial intelligence unit and / or the second artificial intelligence unit as first input values; the system further comprises at least one user interface for outputting a total output value to a user, the total output value being formed based on output values ​​from one or more of the artificial intelligence units, and wherein the system is adapted to perform a method using any combination of the above method steps.

[0031] Other advantages and embodiments of the invention will become apparent from the description and accompanying drawings.

[0032] It should be noted that the names of the units “first,” “second,” and “third” are used here in the sense of reference identifiers, for units that have the specific functions described in each context, and are not necessarily to be understood as ordinal numbers. In particular, for example, in a system consisting of only two artificial intelligence units, for consistency, these two units may be referred to as “second” and “third” units.

[0033] It should be understood that, without departing from the scope of the invention, the above features and the features to be explained below can be used not only in the combinations indicated in each case, but also in other combinations or individually.

[0034] The present invention will be illustrated with reference to examples of embodiments shown in the accompanying drawings, and will be described below with reference to the drawings. Attached Figure Description

[0035] Figure 1 This illustrates a combination of two artificial intelligence units coupled together;

[0036] Figure 2 Various exemplary modulation functions are illustrated schematically;

[0037] Figure 3 This illustrates the application of a dropout method in two coupled neural networks according to one embodiment;

[0038] Figure 4 As shown Figure 1 The system shown includes an additional timer;

[0039] Figure 5 Schematic illustration as follows Figure 1 The system shown has associated classification memories;

[0040] Figure 6 An alternative system with three coupled artificial intelligence units is shown.

[0041] Figure 7An exemplary extended system with a third artificial intelligence unit and a projection plane is shown; and

[0042] Figure 8 Show Figure 7 A system with associated storage elements. Detailed Implementation

[0043] Figure 1 An exemplary embodiment of artificial intelligence units 110 and 120 with two links is shown, which will be described in more detail below. In the following explanation, the artificial intelligence units are exemplary designed as neural networks.

[0044] Therefore, a first artificial intelligence unit is provided, here in the form of a first neural network 110, which can be essentially used to process the input signal X. i The classification is performed, and the results of this classification influence a second artificial intelligence unit 120, which is a second neural network. Preferably, the results of the first neural network are not used as input values ​​to the second neural network, but rather used to influence the network's existing weights, step size, and function. In particular, these parameters of the second neural network can be influenced such that they are not completely redefined, but rather modulated or superimposed on the original parameters of the second network 120 based on the output signal of the first neural network 110. This means that the two neural networks are best operated independently, for example, training their own base values, but can be coupled by superposition. At this point, the two neural networks can be substantially similar to each other in design, but have significantly different levels of complexity, such as the number of layers and classifications. Furthermore, each neural network has its own memory.

[0045] In one possible embodiment, the first neural network 110 can be used as a classification network for coarse and rapid classification of input values, and then, based on the classification results, influences the second network by modulating its parameters. For this purpose, the first neural network may be a network with relatively few layers and a small number of classes K1, K2, ... Kn, which are preferably highly abstract to achieve coarse classification. For example, the first neural network can be limited to 10, 50, 100, or 500 classes, these numbers being, of course, only pertaining to coarse examples. In this respect, the training of the first neural network can be performed specifically and independently of further coupled neural networks. However, additionally or alternatively, a training phase coupled to one or more coupled neural networks may also be used.

[0046] Therefore, the first neural network is designed to provide a usable output in a short time, which can be used to meaningfully influence the second neural network. Weights and functions can be generated from the output value Output1 of the first neural network 110, which can be superimposed on the self-generated weights and functions of the second neural network 120. This means that the second neural network initially operates independently and does not entirely adopt the output value of the first network or the parameters derived therefrom. Furthermore, the second neural network 120 can initially be trained independently in the usual manner, thus possessing self-generated weights.

[0047] In this context, the second neural network can be designed to be significantly more complex than the first neural network, and in particular, can have more layers and / or memory classes. The degree of increased complexity of the second neural network compared to the first network can be determined differently depending on the application. Therefore, the input values ​​or input data used for the second neural network are preferably the same as those used for the first neural network, so that more complex analyses can now be performed using the same data. However, alternatively, the output values ​​of the first neural network can also be used at least partially as input values ​​for the second network. Specifically, for example, in cases where the complexity of the second network is significantly different, a second network can be provided to which the original input values, which are also used as input values ​​for the first network, are provided as input values, and furthermore, the output values ​​of the first network are used as input values ​​for the second network.

[0048] Figure 2 Examples of different modulation functions f are shown. mod This function can be used to superimpose one or more parameters of a second neural network. In principle, superposition or modulation can occur in any way. For example, when the modulation function f... mod_w The weight w applied to the node i2 In this case, it can be specified that the weighting matrix of the second network 120 is used as the independent variable of the modulation function, or it can be the weight w. i2 Each of these provides a one-dimensional (or even different) function. When the modulation function f mod_f_ One of the description functions applied to the second neural network, namely the transfer function f of network 120. trans2 Activation function f akt2 , propagation function or output function f out2 This can be accomplished by combining these two functions and modulating the function f again. mod_f It can be applied to only some or all of the relevant descriptive functions (e.g., all activation functions f applied to the second neural network 120). akt2 Modulation can be applied equally to all nodes in the network, or optionally to only a subset of nodes, or different modulation can be applied to each node. Similarly, for example, modulation can be applied individually or otherwise interleaved for each layer of the network.

[0049] Specifically, the modulation function f mod It can also be a time-dependent function, thereby changing the weights w of the second neural network in a time-dependent manner. i2 Or a function. However, a static modulation function used to modulate the second neural network is also conceivable. In this case, modulation is applied to the parameters of the second network 120, which were originally defined for the second network (e.g., a propagation function or activation function) or obtained independently during the training phase, such as adaptive self-generated weights.

[0050] Eight different time-dependent modulation functions in Figure 2 The following examples are shown. Example a) illustrates a simple binary step function where a value of zero is specified for a given time interval, followed by a value greater than zero. Here, the second value can theoretically be 1, but can also have different values, causing the original parameter to be additionally affected by a factor. In this way, weights can be turned on and off in a time-dependent manner, or weights can be amplified in a time-dependent manner. Example b) illustrates the opposite case, where a step function is predetermined to have a second value less than zero. Similarly, as an alternative to variations of examples a) and b), one can envision step functions comprising two distinct values ​​not equal to 0, thereby increasing or decreasing the level in a corresponding time-dependent manner.

[0051] Example c) illustrates a periodic modulation function, which can also be applied to any parameter of the second network, and in this way will periodically amplify or attenuate certain elements in a time-dependent manner. For example, different amplitudes and / or periods can be selected for different nodes and / or different layers respectively. Any periodic function, such as a sine function or even a discontinuous function, can be used at this point. Depending on the cascade type of the function with the self-generating function of the second network, only positive function values ​​can be selected, but negative function values ​​can also be chosen.

[0052] Example d) illustrates slow, continuous, and instantaneous increases and decreases in layer hierarchy. On the other hand, example e) describes short, approximately rectangular high-level hierarchy, or otherwise low function values, which may be arbitrarily zero. Similarly, example f) shows very short, irregularly distributed peaks or spikes, resulting in increases in layer hierarchy over a very short period. Here, the peaks have different amplitudes and can be positive or negative (relative to the base value). For variations from examples e) and f), there may be regular, periodic, and temporally completely irregular (e.g., randomly determined) distributions of peaks or amplifications. In this case, for example, short increases in layer hierarchy may occur within the time of a decision period of the second neural network, while longer, significant changes in layer hierarchy may extend over several decision periods.

[0053] Figure 2Example g) further illustrates damped oscillations, which can also be arbitrarily designed with different damping and amplitudes. Finally, example h) shows time series of different oscillations around a base value, specifically oscillations with different period lengths while the amplitude remains constant. Such combinations of different oscillations can also be designed as additive superpositions, i.e., beats.

[0054] Typically, any modulation function is conceivable, and Figure 2 The functions shown are for illustrative purposes only. In particular, any combination of the example functions shown is possible. It should also be understood that the baselines shown in all examples can operate at 0 or another base value, depending on the desired effect of the modulation function. For a pure cascade of modulation functions with corresponding modulators, a base value of 0 and a corresponding increase in the function value can be used to ensure that the corresponding nodes contribute to the processing only in a time-dependent manner and are turned off at other times. On the other hand, with a base value of 1, it is possible to, for example, in Figure 2 In the example in a), the modulation function applied to the weights first reproduces the self-generated weights of the modulated network as basic values, and then obtains correspondingly increased weights from higher values ​​in a stepwise manner. Therefore, such a function also acts on the modulation of functions such as activation functions.

[0055] As described above, a modulation function can be formed based on the output value of the first artificial intelligence unit, i.e., based on the first neural network in this example. The relationship between the output value and the resulting modulation function can be arbitrary. For example, this correlation can be generated at least partially during the joint training phase of the coupled networks. In other embodiments, the dependency between the modulation function and the output value of the first network can be predetermined. Optionally, it can also be determined that for certain output values, modulation by the second network does not occur initially.

[0056] Alternatively, in addition to applying the modulation function to the weights and functions of the second neural network, a coupling exit method can also be applied, such as... Figure 3 As shown. This is typically a neural network training process where, in each training cycle, only a subset of the neurons present in the hidden and input layers are used, while the rest are not used (“exit”). For this purpose, existing techniques typically set the exit rate based on the network's feedback error, which determines what percentage of the total network consists of neurons that have been shut down. Similarly, instead of neurons, some edges or connections between neurons may be turned off.

[0057] In an exemplary embodiment, this partial disconnection of neurons and / or edges can now also be used in the second neural network, where the exit parameter is now used not based on the error feedback of the network itself, but in time-correlated modulation that depends on the output value of the first neural network. In this regard, for example, the exit rate of the second neural network can be determined based on the output value Output1 of the first neural network 310, and then that output value can be applied to the second network. The figure again illustrates... Figure 1 Two coupled networks 310 and 320 are shown, but the neurons or nodes 326 and 328 of the second network 320 are schematically indicated as circles. Here, connecting edges are not shown, and the arrangement of the neurons shown is not intended to have any binding relationship with their actual topology. A portion of the available neurons are now deactivated and therefore not used due to the exit rate. Active neurons 326 of the second network are shown shaded in the figure, while unfilled neurons are intended to represent exited neurons 328.

[0058] In a general approach, the coupling termination described in this paper can also be understood as a modulation function f, using a modulation function with weights of 0 or 1, or for example, the output function of each node. mod This can be determined based on the output value of the first network, specifically which neuron 326 or 328 is turned off, or simply by specifying the rate and using a random function to determine which neuron is turned off. In this respect, the exit rate can also be determined again based on the output value Output1 of the first network 310. In this regard, the exit modulation function can also optionally induce a time-dependent shutdown, which would, for example, correspond to... Figure 2 The cascade of the exit function and modulation function is shown. Similarly, a series of mode shutdowns, as demonstrated in previous training, can be employed, such as cyclic mode variations, for shutdown in the second neural network 320.

[0059] Generally, decoupling can improve the operating speed of a neural network. It also prevents adjacent neurons from becoming too similar in behavior. The decoupling described above can be used during the joint training phase of two coupled networks and in already trained networks.

[0060] To ensure that coupled neural networks complement each other in a meaningful way, it can be determined which neural network dominates the entire system at any given time. The network whose output value determines the output of the entire system can be designated as the dominant network or the dominant network. In the following, it is assumed that at any given time, in a set of networks consisting of two or more coupled networks, only one network is dominant, and therefore the output of the dominant network equals the output of the entire system. However, other embodiments are also conceivable in principle, such as specifying rules describing how the output value of the dominant network is processed into the final total output value when there is more than one dominant network.

[0061] In an exemplary embodiment, a timer or timing element may be implemented for this purpose, defining a time specification for one or more of the coupled neural networks. In this case, the time specification is preferably understood as a maximum value or time limit after which the output value of the corresponding network must appear, so that the output can also appear earlier. The output value of the network is then evaluated no later than the time specified for the particular network has expired. The timer can thus control and / or change the advantage between the coupled networks on the basis of a fixed time specification.

[0062] Exemplary embodiments of this type are as follows Figure 4 As shown. Here, the formation and coupling of the two neural networks 410 and 420 can correspond to Figure 1 The example has already been described. Timer 440 now ensures that the output of the first neural network 410 is evaluated no later than a predetermined time after a predetermined time defined by a predetermined time parameter value. The required time can, for example, be derived from the input value X. i The time at which data is fed into the corresponding network begins to be measured. Therefore, the selection of the predetermined timing parameters for the network can be performed particularly dependent on the complexity of the network, so that a usable result can be practically expected within the predetermined time. In, for example, the previously described example, where the first neural network 410 is preferably formed by a network with a small number of hidden layers and a small number of classifications, a correspondingly shorter time can also be chosen for this first network. Similarly, when selecting the timing parameters of the network, other considerations can be taken into account, such as the available hardware that has a decisive influence on the computation time of the network, and / or the application region considered for the coupled network. Furthermore, the predetermined timing parameters can be variable and can, for example, be modified or redefined based on results from at least one of the coupled neural networks. It should be understood that such a timing specification should at least include the time period required as the minimum time required to traverse the corresponding networks 410, 420 once. Figure 4 In the example, a 30ms time span is specified for the first network so that it dominates during the process's execution, from 0ms at the start of the process to 30ms. However, other suitable values ​​can certainly be chosen for this time span.

[0063] During the time period specified by the time parameter of the first network 410 (here, 30ms), the first neural network will process the input value X in the usual manner. i After a predetermined period of time, the output Output1 of the first neural network 410 can be used to generate a function for superimposing or modulating the weights and functions of the second neural network itself. Furthermore, the output value of the first neural network can also be processed independently, either as an alternative to influencing the second network 420, and for example, as a fast output for the entire system.

[0064] Once the modulation function f mod_f f mod_w The second neural network 420 has been applied, and the timer 440 can start a new timing measurement, now applied to the second timing parameters predetermined for the second neural network 420.

[0065] In this respect, the second neural network 420 can also optionally and independently utilize, even through the obtained modulation function f mod_f f mod_w Input value X before modulation i Thus, for example, the input values ​​can be provided to the second neural network 420 even before the start of the second predetermined time period, and processed accordingly there. After the first time period has elapsed, the input values ​​are then processed by applying the corresponding modulation function f. mod_f F mod_w The parameter values ​​and functions of the second neural network are then superimposed. In this regard, one or more modulation functions can be formed for different parts of the second neural network 420, such as the weights, output function, propagation function, and / or activation function of the second neural network. In the case of forming a second neural network 420 that is significantly more complex than the first neural network 410, for example by having significantly more layers and nodes and / or by having a higher number of memory classes, the second neural network will require a relatively higher computational workload and therefore also more time. Therefore, in this case, the second time period can be chosen to be correspondingly longer.

[0066] In this regard, alternatively, each of networks 410, 420 can continue to process and evaluate input values ​​continuously, even when it is determined that another network is the dominant network in the entire system based on the current time span. Specifically, in the example of the two coupled networks shown, the first network can continuously evaluate input values ​​even when the second network is dominant, so that the output value of the entire system can correspond to the output value of the second network after the second time period has passed and the second network has found a solution. In this way, a fast classification network such as the first network 410 described herein, which always evaluates available input values, can also perform short-term interventions to guide the found output values ​​to find their way into the total output. These embodiments will be described in further detail below.

[0067] Because of this time control performed within a predetermined time period in the timer, the entire system can make decisions earlier and, for example, can take action without the need for the final evaluation and detailed analysis of the second neural network. As an example, a situation in an autonomous driving system can be considered to be evaluated by such a system with at least two coupled networks. An early classification of "hazard" can be achieved through the first unit or the first neural network, which, without involving any further evaluation of the nature of the hazard, may already lead to an immediate reaction, such as reducing vehicle speed and activating braking and sensor systems. Simultaneously, based on the classification, i.e., under the influence of the modulation of the first network's output value, the second neural network performs a more in-depth analysis of the situation, thereby causing further reactions or changes in the entire system based on the output value of the second network.

[0068] Alternatively, it is conceivable not to specify a time limit for each coupled network, but only for one network (or, if more than two networks are coupled, only for a subset of the coupled networks). For example, in the example above, a timer could be applied to the first fast classification neural network, while the second network has no fixed time constraint, and vice versa. Such embodiments can also be combined with further methods for determining the currently dominant network, which will be described in further detail below.

[0069] In all embodiments with inserted timers, it can be specified that the output value of the neural network currently having the active timer is used as the output of the entire system. Due to the time required for the network to reach a first solution for a given input value, there is a certain waiting time during which previous output values ​​(of the first or second network) are still available as the total output value.

[0070] If a timer is defined only for certain coupled networks, such as a timer that is only valid for the first network, then it can be defined such that the output of the entire system usually always corresponds to the output of the second network, and is only replaced by the output of the first network when the timer of the first network is active, i.e., the predefined time period is actively running and has not yet expired.

[0071] In systems with two or more networks, proper synchronization between networks can be achieved by aligning predetermined time spans and altering timers, especially when several networks with different tasks need to achieve a result simultaneously, which in turn affects one or more other networks. Similarly, synchronization can be achieved between several separate monolithic systems, each comprising several coupled networks, by adjusting predetermined time periods and sequences. In this case, the systems can be synchronized, for example, through time alignment, and then operate independently but synchronously according to their respective timer specifications.

[0072] In addition to changing the dominant neural networks in the system based on timers, each neural network can also, alternatively, make decisions to transfer dominance in a cooperative manner. This could mean, for example, that the first neural network in the system processes the input values ​​and reaches a certain first solution or a certain output value.

[0073] As the timer's help center shifts, it can be specified here that the output value of the total network corresponds to the output value of the current dominant network in each case.

[0074] For this purpose, for example, changes in the input values ​​can be evaluated. As long as the input values ​​remain substantially constant, the dominance distribution between the coupled networks can also remain substantially constant, and / or can be determined solely based on timers. However, if the input values ​​change abruptly, a predetermined dominance can be established that overrides other dominant behaviors of the coupled networks. For example, for a suddenly changed input value, it can be determined that the dominance will initially revert to the first neural network in any case. This also restarts an optional current timer for that first neural network and executes the sequence as described above. Significant changes in the input values ​​may occur, for example, if sensor values ​​detect a new environment, or if a previously evaluated process has been completed and a new process is now to be triggered.

[0075] Thresholds can be specified as significance thresholds, which can be used to determine whether a change in the input value should be considered significant and result in a dominant change. Individual significance thresholds can also be predetermined for different input values ​​or for each input value, or a general value, such as a percentage deviation, can be provided as the basis for evaluating the significance of changes in the input value. Similarly, in addition to fixed significance thresholds, there can be thresholds that can change as needed or adaptively, or they can be functions, matrices, or patterns on which the significance of changes can be evaluated.

[0076] Optionally or additionally, the advantage between coupled networks can be altered based on the output values ​​found by each network. For example, according to an embodiment, the first neural network can evaluate input values ​​and / or changes thereto. In this case, an importance threshold can be pre-determined for the category that can be used by the first classification neural network in each case, such that if the first neural network detects a significant change in the category of the input data, the advantage is immediately transferred to the first neural network to quickly reassess the situation and, if necessary, react. In this way, it also prevents the second neural network from unnecessarily continuing analysis for an extended period without deeply considering the change, even if the first fast classification network detects a significant change in the input.

[0077] In all the examples above, the output value of the entire system can be further used in any way, such as as a direct or indirect control signal for the actuator, as data stored for future use, or as a signal passed to the output unit. In all cases, the output value can also be further processed by additional functions and evaluations, and / or combined with further data and values.

[0078] Figure 5 Show again as Figure 1 The simplified embodiment shown has two unidirectionally coupled networks 510, 520, and thus, classification memories 512, 522 for each network are now schematically illustrated. The classification type K used is... i This is initially secondary and will be described in more detail below. In particular, the sizes and structures of the two classification memories of the first network 512 and the second network 522 can be significantly different, thus forming two neural networks with different speeds and focuses or centers of gravity. Therefore, for example, as already briefly described, the interaction between a fast, coarse classification network and a slower, but more detailed analysis network can be achieved to form a coupled overall system.

[0079] In this example, the first neural network 510 consists of a relatively small number of classifications K1, K2, ..., K n The network can also be formed following, for example, a flat hierarchical structure, thus performing classification in only one dimension. Preferably, such a first network 510 can also be formed with a relatively simple topology, i.e., having a relatively small number of n neurons and hidden layers. However, in principle, the network topology can be largely independent of classification.

[0080] The second neural network 520 can then have significantly larger and / or more complex classification systems. For example, such as Figure 5 As shown, the memory 522 or basic classification can also be hierarchically structured across multiple levels 524. The total number m of classes K1, K2, ..., Km in the second network 520 can be very large, particularly significantly larger than the number n of classes used by the first neural network 510. For example, the number of classes m and n can differ by one or more orders of magnitude. Therefore, an asymmetric distribution of the networks throughout the system is achieved.

[0081] The input values ​​can then be rapidly classified using the fast classification of the first neural network 510. An abstract summary class is preferably used for this purpose. In one example, the classification of the sensed situation (e.g., based on sensor data such as image and audio data) can then be initially performed by the first neural network 510 as “large, potentially dangerous animal” without performing any further analysis for this purpose. This means, for example, that there is no further classification by animal species (wolf, dog) or as a dangerous predator in the first network, but only based on the broadest possible range of general characteristics, such as size, detection of teeth, attack posture, and other features. This data essentially corresponds to the output “danger” and can then be optionally passed to an appropriate external system for an initial and rapid response, such as a user alarm system or a specific actuator in an automated system. Furthermore, the output Output1 of the first neural network 510 is used to generate the described modulation function for the second neural network 520.

[0082] The same input value X i For example, the sensor values ​​are also fed to the second neural network 520. In this case, the input values ​​can be input immediately, i.e., substantially simultaneously with the first network, or delayed, in which case they are input before the modulation function is applied, or only when the modulation function is applied, i.e., when the result of the first network is available, depending on the embodiment. Preferably, they should not be provided to the second neural network later, especially in time-sensitive processes, to avoid delays. The second neural network then also computes the solution, and the initial self-generated weights of the second network and its basis functions (e.g., the specified activation and output functions) can be superimposed based on the modulation function formed by the output values ​​of the first network. This allows the iterative work of the second network to omit a large number of possible variations in cases where there is no time for critical situations (e.g., dangerous situations) detected quickly by the first network. Although the analysis of the second neural network is slower, as mentioned above, possible responses can already be performed on the basis of the first neural network. This corresponds to the first instinctive response in biological systems. Compared to the first network, the hierarchical structure and significantly larger memory of the second network allow for precise analysis of the input values, such as the detailed classification mentioned in the example as "dog" class, corresponding breed, behavioral characteristics indicating dangerous or harmless situations, etc. If necessary, once the second neural network produces a result, the previous reactions of the entire system can be overridden, for example by downgrading the first classification "danger" again.

[0083] In general, for such coupled global systems with asymmetric classification, it is conceivable, for example, that the fast classification first network 510 class K... nThe primary function is to perform abstract classifications, such as new / known situations, dangerous / non-dangerous events, interesting / uninteresting features, necessary / unnecessary decisions, etc., without going into depth. In this respect, the first classification does not necessarily correspond to the final result ultimately found by the second unit 520. However, the two-stage classification performed by at least one fast and one deep analysis unit thus allows for emotional or instinctive reactions from the overall AI system. For example, if an object is identified as potentially a snake through image recognition, then the "worst-case scenario" is best determined by the first classification, regardless of whether that classification is correct. In the case of human intelligence, evolutionary knowledge and instinctive reactions can be replaced by a fast first classification based on pre-programmed knowledge, so appropriate default reactions (keeping distance, initiating movement, activating increased attention) can also be performed by the entire system and its actuators. The additional modulation by the second learning unit based on this first classification can then be understood as an analogous to an emotion-related superposition, i.e., a fear response corresponding to an analysis of a situation that automatically initiates a different understanding than a harmless one. The superposition of the parameters of the second neural network, performed by the modulation function, can thus lead to necessary shifts to other classification spaces that would otherwise not be reached or not immediately reached by default.

[0084] Therefore, such systems can be used in a wide range of applications, including all applications where critical decision-making occurs. Examples include driving systems, rescue or warning systems for different types of hazards, surgical systems, and tasks that are typically complex and non-linear.

[0085] In the embodiments described so far, only two AI units are coupled together. However, this idea is also applicable in principle to more than two units, so that, for example, three or more AI units can be coupled in a suitable manner, thereby determining which unit can modulate the parameters of one or more specific other units. Figure 6 An example is shown in which three neural networks 610, 620, 630 (and / or other artificial intelligence units) can be provided, wherein the output value of the first network 610 generates a modulation function for the weights and / or functions of the second network 620, and wherein the output value of the second network generates a modulation function for the weights and / or functions of the third network 630. In this way, an arbitrarily long chain of artificial intelligence units can be formed, which influence each other in a coupled manner through superposition.

[0086] Similar to the previous example with two neural networks, in one embodiment, all coupled networks can receive the same input value, and processing can only be coupled through the modulation of the respective networks. However, embodiments are also conceivable where, for example, in Figure 1A third neural network is provided after the two neural networks, which receives the output values ​​of the first and / or second networks as input values. Optionally, the functions and / or weights of the third neural network can also be modulated by, for example, a modulation function formed by the output values ​​of the first network. These can be the same as or different from the modulation function formed for the second network. Optionally, for example, the output values ​​of the third network can be used to form an additional modulation function, which is then recursively applied to the first and / or second networks.

[0087] It is understandable that various further combinations of the corresponding coupled learning units are possible, wherein at least two connected units are coupled by forming a modulation function for describing the parameters of the unit, particularly for neural networks used for the weights and / or functions of the network. As the number of coupled units increases, more complex variations in modulation and coupling are conceivable.

[0088] As noted at the outset, the embodiments described herein are examples relating to neural networks, but in principle, they can also be transferred to other forms of machine learning. In this context, all variations are considered, wherein at least a second artificial intelligence unit can be influenced by a first artificial intelligence unit through superposition or modulation based on output values. Modifications to the weights and functions of the neural network by superposition of modulation functions from the examples above can be replaced by appropriate modulation of any suitable parameters controlling or describing the operation of such learning units. In each example, the term "learning unit" may be replaced by a specific instance of a neural network; conversely, the neural network of the exemplary embodiments can also be implemented in a generalized manner as an artificial intelligence unit, even if not explicitly stated in the corresponding examples.

[0089] In addition to neural networks, examples include evolutionary algorithms, support vector machines (SVM), decision trees, and special forms such as random forests or genetic algorithms.

[0090] Similarly, neural networks and other artificial intelligence units may be combined. In particular, for example, the first neural network in the previous example, described as a fast classification unit, can be replaced with any other artificial intelligence unit. In this context, a fast, coarse classification method particularly suited to the features can also be selectively chosen. However, the output value of such a first learning unit can then be applied in the same manner as described for two neural networks to form a modulation function for a second artificial intelligence unit, which can again be a neural network.

[0091] In addition to or alternatively to the variations described, systems comprising two or more AI units can be further improved by using at least one AI unit, for example, to evaluate or validate the results of another AI unit, or to perform further tasks as shown below. For example, in principle, a system could be formed with two AI units, where one subsequently evaluates and selectively influences the results of the other, for example, by determining the overall output of the system. However, more complex variations are also possible, such as systems comprising three AI units, where two are modulatedly coupled to each other as described above, and a third unit is now added as a validation unit.

[0092] In this regard, as an exemplary embodiment, a system including a first neural network 710 and a second neural network 720 can be considered as a combination Figure 1 or Figure 5 As described, a third neural network 730 is added to it. This third network 730, or more generally, a third artificial intelligence unit, can also have its own classification memory. Such an extended system, such as Figure 7 As shown. Additionally, a projection surface 750 can now be provided. The projection surface 750 can be a software and / or hardware unit or a combination of several such units. In particular, the projection surface can form a complex of several units, which may also include an artificial intelligence unit and its memory elements. The projection surface 750 can thus form a central unit in which the outputs of at least a second unit (and optionally other units) are processed and linked, such as a multiplexer unit, a unit that generates sequences from data, and a unit that imprints identifiers or labels on stored data or sequences after positive or negative evaluation, for example, when comparing new data with stored data, this unit can simplify decision-making. These functions can also be performed wholly or partially by program modules. The projection layer may also include input and / or output units, such as screens or acoustic units, enabling communication with a user or a user supporting, for example, the training phase of the system, and enabling evaluation, for example, the current processing state.

[0093] A storage element may be associated with the projection surface 750 as a projection memory, wherein data from the projection surface may be stored, at least temporarily. For example, the storage period of this data can typically be defined, but may also be defined by one of the units, such as a third verification unit. According to embodiments, the projection layer memory can essentially be used as a short-term memory whose contents can be checked, erased, overwritten, and / or transferred to other memories, such as memory elements of the corresponding neural network or unit, as needed. Therefore, the projection memory can be implemented, for example, as a ring memory, wherein in each case, after a certain number of entries or a certain amount of data, the memory is "full," and thus the previous data is overwritten from the beginning, corresponding to a ring structure.

[0094] The third neural network (agent) 730, which acts as the verification unit here, now evaluates the data present in the projection plane 750. This can be, for example, input data, output data from one of the other two networks 710, 720, and / or other data. For example, output data (optionally with additional information) can also be processed in the projection plane 750. Furthermore, the data can be processed in some way or, for example, simplified before the projection plane. In response to the processing of data by the third unit 730, which may correspond to classification or evaluation, various actions can be performed that affect the final output value of the entire system and the actuators, output interfaces, and other related signals, and / or actions that affect the further behavior of the entire system, particularly the further behavior of other units included in the system. These actions will be described in more detail below.

[0095] In this respect, the data in projection plane 750 can also be correlated with each other to generate at least one more piece of information. For example, a specific output value of the second network 720 can be indicated by a corresponding assignment or sub-parameter from a specific input value that also exists in projection plane 750, and the two values ​​are thus linked. Furthermore, time parameters can exist, such as indicating at what time a specific output value is output. Therefore, such time parameters can include absolute time indications, but can also include relative time indications that depend on the current evaluation time or another reference time. Similarly, time periods can be specified instead of fixed points in time, and one or more input and / or output values ​​of the network are assigned to these time periods. In addition to or as an alternative to actual time parameters, sequences can be associated with at least a portion of the data so that, even without an explicit time specification in the projection plane, it can be identified in which sequence multiple current data values ​​were generated or processed, for example, in the form of designated numbers of output values ​​in the projection plane.

[0096] In this way, data existing in the projection plane 750, such as input and output values, can form a time series. Optionally, such a series can also be labeled to, for example, determine a certain time period or a specific output value belonging to a defined series. For further processing, storage, or evaluation, the series formed in this way can then be processed as a whole. Optionally, different input and output values ​​belonging to the same time period or series can then be processed together, for example, compared with each other. By placing data over time in the projection plane and grouping them into specific processes and sequences that can be accessed again, a memory-like structure becomes possible. Figure 7 Timer 740 is also shown again, according to the combination Figure 4 The ideas described can also be used in this system to dominate between control units 710 and 720.

[0097] Figure 8 Show again Figure 7The system is shown in the diagram, where memory elements for all system components are also illustrated. The third neural network 830 can also be allocated a classification memory 832, which can be designed to have relatively few layers and classes Kj, as shown here. However, in principle, more complex designs are also possible. In this example, it is assumed that the number of layers and / or classes Kj of the third neural network 830 is significantly smaller than that of the second neural network 820 provided as an analysis unit, i.e., one or more orders of magnitude smaller. In the schematic diagram, the memory 812 of the first classification network 810 is designed to be similarly complex to the memory 832 of the third verification network 830; however, this is not necessarily predetermined. The memory may also differ further. However, in variations where the three networks 810, 820, 830 are assigned tasks as described above, a significant asymmetry between memory size and complexity will generally prevail, where the second network 820, designed as an analysis network, has a larger and more complex classification memory 822.

[0098] The projection surface 850 is also provided with a memory 852, which can be designed as a volatile or non-volatile memory element, such as a ring memory or another short-term memory, as already combined. Figure 7 The data to be processed in the projection surface 850 is stored in the projection memory 852. The storage duration of the data and the selection of the data to be stored can be designed in very different ways. For example, a fixed storage period may initially be specified. After this time has elapsed, the data in the projection memory 852 can be discarded and / or overwritten. Additionally or optionally, a part of the system, such as a third neural network 830, can make a decision about whether to transfer at least partially the data in the projection memory 852 to another element of the system. To prevent data loss, this decision can be made before the predetermined storage period expires. For example, the third neural network 830 can decide that some or all of the data stored in the projection memory 852 should be transferred to another memory element for long-term storage. For example, this could also be one of the memory elements 812, 822 of the neural network, particularly the memory 822 of the second neural network 820, which, as an analysis unit, can perform the basic part of result determination. By storing other units in the system that will later evaluate new situations in the memory, this allows for the merging of already evaluated situations in a manner similar to experience. However, in principle, a separate memory module (not shown) can also be provided as long-term memory, wherein one of the units can be pre-defined or fixedly scheduled to determine which unit can access the data stored there and to what extent. For example, it can be determined that only the second unit 820 (and in the case of a larger system, such as a further detailed analysis unit) accesses the long-term memory, and / or the third unit 830 regulates access for evaluation.

[0099] For example, by combining input values ​​such as acoustic data, image data, sensor data, and text, the data present in the projection surface 850 can correspond to an image of the current situation, and then by combining them with the output values ​​of various units, an explanatory image of the situation can be obtained.

[0100] Output values ​​from at least one of the other networks 810, 820 can now be provided to the projection surface 850, as indicated by the arrows in the figure. Optional inclusion of input values ​​is not shown here. The third neural network 830 receives data from the projection surface 850 as input values ​​and thereby forms a third output value. In this regard, all or only a portion of the data present in the projection surface 850 or the projection memory 852 can be used as input values ​​for the third network 830. Based on the obtained solution or output value, the third network 830 can then trigger various further actions. For example, based on the third output value, a determination can be made regarding whether the output values ​​of the first network 810 and / or the second network 820 are within a valid range or parameter, i.e., whether they are acceptable as a valid solution. If it is determined here that the current output value of the first network and / or the second network (determined as the total output value of the system in the preceding example) is outside the permissible range, based on this evaluation, the reaction of the entire system can be prevented or stopped so that the previously obtained first / second output values ​​of the first / second network are not passed to, for example, an actuator or interface. Output values ​​that are evaluated as unacceptable or invalid can then be discarded, but they can also be stored along with the evaluation so that they can be referenced in later cases, such as through comparison. In this way, the evaluation of later cases can be simplified or accelerated because solutions that have already been determined to be invalid are not sought.

[0101] Conversely, it is possible not only to check the permissibility of output values ​​and solutions, but also to check whether they particularly meet certain conditions or specifications, representing the most desirable possible solution. Output values ​​found in this way can then be preferably used as the output of the entire system, or they can be stored in long-term memory, for example, so that they can be quickly retrieved in the future as the best solutions found. Similarly, corresponding evaluation parameters can be provided for output values ​​evaluated as particularly unfavorable or favorable, and these evaluation parameters can also be stored and / or further transmitted in association with these values. These evaluation options can be implemented individually or in combination.

[0102] For example, the previously described decision to transmit and store data from the projection surface 850 can be made based on this evaluation of the third unit 830.

[0103] In another embodiment, the input values ​​of the system used as input values ​​for the first and second units may initially be stored only in the projection surface or its associated memory 852. This can be considered as an unprocessed image of the current situation. Meanwhile, the first and second units begin to process these input values ​​as previously described, i.e., using various means such as alternating dominance between units and modulating the determined parameters and functions of the second unit.

[0104] The solutions found, i.e., the output values ​​of the first and / or second meshes, can then be transferred to the projection plane, where they can either overwrite the corresponding input values ​​or be stored and linked together. For example, depending on the variation, any newer result obtained from the same input values ​​may overwrite an older result. In other cases, results can be transferred to the projection plane while retaining older solutions so that, for example, a comparison of current and previous solutions can be used to evaluate whether a later or earlier result better satisfies the specifications and boundary conditions of the third mesh. As an example of linking, it is possible, in particular, with respect to elements that are simultaneously superimposed or also superimposed, such as images.

[0105] Optionally, the advantages between the previously described coupling units can also be considered so that, for example, only the output values ​​of the currently dominant network are stored in the projection plane or its memory. Similarly, time parameters can be associated with these values. For example, time periods of equal or different lengths can be defined, and corresponding input and output values ​​can be appropriately assigned to those time periods to reflect the temporal progression of the situation. For example, input values ​​can be stored in the first time period, while the output values ​​of the first classification unit are stored in the next time period, followed by the output values ​​of the second unit. Furthermore, improved or at least modified output values ​​can then be stored in additional sections. For each block of output values, a time period can be predetermined; however, they can optionally be marked as belonging together to clarify the order of identification.

[0106] In exemplary embodiments, control over data storage to or from the projection memory can also be undertaken, at least in part, by multiple units of the system. For example, a scenario has been described where, in a system with a defined unit dominance, an examination is performed to see if the input value has changed beyond a predetermined level, thus assuming a new scenario. In this case, dominance can be passed to a first classification unit to create a new coarse classification of the input value. Simultaneously, the first unit can then signal the projection memory whether the data stored there (corresponding to the previous scenario) should be transferred to another memory element, such as long-term memory, or whether it can be overwritten later. Similarly, depending on the current processing operation, the first and / or second units can adjust the storage duration in the projection memory in response to different scenarios or objectives. For example, a longer storage duration can be set in the projection memory when a long, in-depth solution search is required, while rapid decision-making can lead to rapid changes in the stored data.

[0107] Optionally, one of the units may have the priority to determine, for example, the storage duration, so that, for example, the decision of the first unit to discard previous data in the projection memory can be checked or prevented by the second unit, so that, for example, the corresponding data is still stored in the case of duplicate input values. Furthermore, it is conceivable that, at least to a limited extent, one unit can modify the classification memory of another unit and, for example, create new categories. However, protected areas may also be defined where all default values ​​that must not be changed or deleted by any unit are stored.

[0108] like Figure 4 or Figure 7 In the system shown, if a timer is provided that can measure time intervals and transmit that information to the units, the timer can, for example, monitor a specified storage duration. Thus, the storage time can also be coordinated or synchronized with the distribution of advantages among different units of the system. However, alternatively, another part of the system can be provided for this purpose. In principle, a third network or a third learning unit can also change and redefine the predetermined storage duration. In this context, different specifications for the storage duration can also be defined for different values. For example, it can be determined that the output value of the first fast classification unit is stored in the projection plane for a short time, thus defining a long first storage duration specification for it, while the output value of the second unit has a longer storage duration specification. Alternatively, only the output value of the first unit can be stored until the output value of the second unit is available. Optionally, the third unit's check of the output value can be waited for. If it is determined that the output values ​​of the second unit are invalid because they do not correspond to the specified boundary conditions, the storage of these values ​​can be canceled, while the coarse output values ​​of the first unit are retained.

[0109] Alternatively, the third AI unit can further modify the detailed modulation between the first and second units based on its evaluation, or form its own modulation function, which can be applied to the parameters and functions of the second network. Thus, the modulation function can be applied to the second unit, i.e., the second network in the embodiment, by both the first and third AI units, thereby again affecting their output values. Similarly, based on the evaluation of the output values ​​that have already occurred, the third unit can, for example, redefine or modify the exit rate already described for the nodes of the second neural network to affect processing. For example, this method can also be used to improve processing speed if finding a solution without exits takes too long. All these possibilities can occur in parallel and are complementary to the coupling and modulation between the first two units.

[0110] The system can be designed in such a way that the third unit does not determine the validity or invalidity of the results and output values ​​on its own, but rather in conjunction with other units of the system, for example by influencing processing parameters, storage time, and other components.

[0111] When the third neural network is designed as an evaluation or verification unit, this can, for example, include classifications that primarily comprise normative features such as prohibitions, priorities, and similar values. By having such a third unit evaluate the outputs of other units, it can be ensured that the entire system only allows solutions that satisfy these norms. These can be simple, fixed boundary conditions of the system; however, they can also be further developed into an ethical-like module that keeps the development and decision-making of the entire system aligned with the values ​​and rules prescribed. These boundary conditions, particularly those given by the classifications in the third network, can therefore preferably be fixed and stored without being altered by the system. However, in other embodiments, it is also conceivable that the system learns these classifications at least partially on its own, i.e., through unsupervised learning, thereby at least partially forming its own value system or learned set of boundary conditions. Embodiments in which a basic system with predefined, immutable boundary conditions can also be used, which can then be supplemented during the training phase or during operation and / or by external data input.

[0112] The constraints provided by the third unit can be applied to the joint training phase of the coupled network and the later evaluation phase of the trained system. Optionally, several individual presets corresponding to several completed classification groups can be provided in the classification memory of the third unit. One of these groups can then be selected as needed, for example, depending on the situation at hand. The identification of the specific situation at hand and the assignment of the classification group to be applied can again be based on the results of the first and / or second units. For example, in this way, different risk preferences or "basic sentiments" of a system can be achieved. In this context, basic settings can also be predefined and changed only in certain situations. It is also conceivable that, during the training and operation phases of the coupled system, new classification groups with additional or flexible boundary conditions can be actively formed for the third unit from the invariant basic settings.

[0113] As an example, in the case of autonomous vehicles or aircraft (such as aerial drones) used for rescue personnel, riskier driving methods can be permitted as long as there are no passengers, especially if the vehicle needs to reach a predetermined location quickly. Therefore, after a rough identification of situations, such as those that can be handled by the first unit (e.g., an "alarm situation"), a suitable classification group can be selected for the third unit, and the solutions or output values ​​of all units can then be evaluated based on that classification group. Thus, basic constraints, such as avoiding accidents, can be further respected, while other constraints (such as sharp turns, accepting damage, or others) can be relaxed. Once the system detects a new situation, such as passengers in the vehicle, a different classification group may be applied to the third unit, now potentially focusing more on the well-being of the passengers, or even the rescued casualties. In extended embodiments, further standard catalogs can also be created that can be used for classification in specific situations, such as for load transport, during firefighting, during reconnaissance flights or trips, and other situations. If a situation can be classified as such a known situation, then the third entity can be constrained to maintain the validity of boundary conditions and remain passive as long as no contradictions occur. However, if more complex or unknown situations arise, which could lead to damage or other undesirable consequences, the third unit can also intervene more actively in the second unit's search for solutions, and for example, by specifying a new search space, changing or modulating the parameters of the second unit, or otherwise supporting the finding of a suitable solution.

[0114] Therefore, in this way, the framework conditions of the entire system, such as prohibitions, priorities, and value systems, are located in the memory of the third unit, preferably hard-coded. Processing acceleration can also be achieved by excluding certain solutions. The third unit can actively intervene in the solution search of the second unit through actions such as rewards / penalties or inducing new step sizes. Therefore, the output value of the second unit is also influenced by the third unit through a special type of feedback.

[0115] As another option, if already combined Figure 4 or Figure 7 As described, the third learning unit can influence the determination of the dominance of individual units in the system, which is implemented in a system consisting of several (e.g., three in this case) coupled units. For example, the third unit can examine whether a specified time parameter for a dominance shift leads to a meaningful result, or whether different distributions or specifications of time periods should be specified. This can be used, for example, to flexibly address situations requiring coarser classification than usual, such as situations requiring decisions to be made in a shorter timeframe. Therefore, as previously described, signals from the third unit to the timer module can be used to set one or more new timing parameters for each coupled network as needed, and then further determine dominance based on said parameters. Similarly, after evaluating input values ​​over time, the third unit can determine whether the input values ​​and the situation being evaluated change very rapidly, or whether the situation remains quasi-statistically constant over a long period, and specify different time parameters for processing based on this.

[0116] Figure 7 and Figure 8 The described components and processing steps are, of course, not limited to the illustrated embodiment with three neural networks. In particular, the basic idea of ​​the evaluation unit can, in principle, be applied to any system with at least two artificial intelligence units.

[0117] Another element of an intelligent system comprising multiple coupled artificial intelligence units can be a language unit. In this context, language is preferably understood here as internal communication within the system, and optionally also as communication with other identical or similar systems. Specifically, language units can be used to reflect and process situations represented in abstract form by the system's input and output values. For example, in the case of a unit learning in an unsupervised manner, classifications can and should be formed independently without prior specification; thus, the system's own classifications (recognition of situations, objects, and conditions) can be understood as a language. In this context, language units can also be provided with their own associated memory and can optionally be configured as neural networks. According to embodiments, it may be convenient for the speech memory to be closely associated with, for example, a second analysis unit of the system (according to the preceding example embodiment) or to be formed as part of that unit. If the speech unit has its own associated memory, that memory can be structurally and content-similar to the memory of the second unit and can optionally communicate with the second unit, for example, to exchange newly acquired classifications. External connections can also be provided for inputting and outputting data from the speech memory.

[0118] Preferably, such a unit can convert, for example, memory categories and contents stored on a projection surface for long-term storage into abstract terms, while also forming syntax and semantics, which then allows for logical application and processing of the content. This language can be used for communication between various learning units within the system, allowing for abstract problem-solving without directly addressing physical input values.

[0119] In all embodiments, it will be understood that individual elements, such as memory elements, neural networks, interconnections between units, and further, may be implemented differently than those shown herein. For example, of course, there may be additional memory elements not shown in these diagrams, or some or all of these memories may be in the form of a single physical memory element, for example, appropriately subdivided by addressing.

[0120] The various sub-methods described above can (especially in combination) lead to systems that act jointly and autonomously. Such systems can cover a wider range of applications compared to intelligent systems trained only in a single specialized domain. In traditional systems, data processing is equivalent to finding the minimum of error or the maximum of success in a multidimensional space. The more dimensions this space has (through classification, options, hierarchy), the more likely the system is to get stuck in local minima or maxima. On the other hand, the coupling and influence of the first unit on the second unit, as previously described in the form of modulation functions applied to weights, step size, functions, etc., allows jumps in the search space to initially unchecked and essentially irrelevant regions that would otherwise never be reached, depending on the defined step size. In doing so, the task can also be changed in a short time, as the relevance of the new solution space is quickly checked in each case.

[0121] For example, if a potentially dangerous but unidentifiable animal is detected from the input values, and neither the first nor the second unit can find a matching image pattern for that animal, the system can now jump to noise analysis with a modulation step size (e.g., randomly induced), where it finds matching noise recorded from the unidentifiable animal. The second unit, modulated in this way, can then check whether the found solution can be applied to previously unclassifiable images, which corresponds to associative performance.

[0122] Similarly, the introduction of projection layers can model associative behavior; for example, the current decision (in the form of an output value) can be compared with previous decisions and optionally evaluated. Therefore, the evaluation system is not based on abstract learning ability based on positive or negative rewards, but rather on associative processes. For example, if the system does not find a better solution in a jump domain, it can jump back to the domain of the previous best solution, which is determined by the evaluation in the projection layer and optionally stored, and can initiate a new jump domain. In this way, for example, modulation of the first unit can always be performed from a suitable starting point found based on the evaluation of the third unit.

[0123] As a possible application example, a user-specific AI system should be considered. Ideally, such a system could develop intelligent behavior in the sense of hard AI by coupling multiple AI units (including, among others, the described modulated feedback and at least one evaluation unit with corresponding storage capacity). Such a system should preferably be able to freely correlate and independently classify problems. Furthermore, user-specific behaviors should be possible so that the AI ​​system can respond individually to the user, i.e., specifically detect and / or learn what interests, traits, emotions, feelings, personality characteristics, and what level of knowledge the user possesses.

[0124] In this process, such data, along with other externally collected data, can be added to the system. Updates to the entire system are also possible, such as changes to the rating system or classification within a certain scope. However, ideally, mechanisms should be in place to completely prevent data from being exported from the system, especially since it operates at a highly personal level. Therefore, personal data should not be disclosed to outsiders or accessed at will. For example, in this context, it is also conceivable that the AI ​​system operates primarily offline, i.e., without connecting to external communication networks or other interfaces. To update the system or load background knowledge and other data, a time-limited secure connection can then be established, for example, one that can be entirely controlled by the user. For instance, the source of the added data can be specified, and the user can be given the option to consent to the connection.

[0125] An initial training phase can be provided for the system, during which learning communications with external parties occur, along with predefined training datasets and / or data not from actual users of the system. This training phase can provide a general, basic environment for topics, knowledge, experience, and expertise, so that external data can be used only in specific situations later. Predefined communication characters and an initial depth of association for general states of learning processes can also be set. Furthermore, problem identification and appropriate responses to situations, as well as associative communication processes, can be trained during the training phase.

[0126] After the initial training phase, which can be performed by the manufacturer, is completed, a second training phase can be performed by the user, for example. Here, for instance, time parameters can now be set for the user. By learning from and adapting the system to the coupled network, the previously initially set communication characters can be adapted to the user (through mirroring or supplementation). Similarly, character features and interests previously set for a general user can now be adapted to a specific end user. Furthermore, in exemplary embodiments, particularly those including visual evaluation, the projection surface can generate a current image of its state and display it on a screen. This overlaid image then allows external users or trainers to evaluate it during the operational phase, particularly the training phase, and they can use it to determine how the system is generally assessing the current situation. In this way, how the system is working can be seen at an early stage, or intervention can be made if necessary to correct or modify certain aspects, thereby accelerating the training phase.

[0127] After these training phases, the system is best ready for use. However, supplementary training phases can be used later.

[0128] According to embodiments, the AI ​​system can have different interfaces to register environmental conditions and user-performed actions, as well as the user's emotions and mental state. Various sensors can be used for this purpose, such as cameras, microphones, motion sensors, infrared sensors, "artificial noses," ultrasonic sensors, and any other sensors. These can be arranged individually, distributed, and / or in combination in suitable moving or static objects to provide the most comprehensive analysis possible. Furthermore, other interfaces can be provided for the AI ​​system to communicate with the user, such as speakers for voice output or screens and other display devices for visual display and text representation.

[0129] In one possible embodiment, an object is provided in which such an AI system is integrated for the user. This can be a mobile object, such as a technological device (e.g., a smartphone), but more specifically it can also be a furniture object or an everyday object, such as a lamp, vase, screen, mirror, or other object that already occupies a fixed position in the home. The system's task is to provide the user with an artificial, personal, intelligent companion. The system establishes the user's identity and communicates with them, for example, through voice and / or images, if it is a screen or a projector installed in the room. Therefore, the output is connected to an interface (such as a speaker, screen, or projector).

[0130] Based on the aforementioned inventive elements, this system can categorize situations, introduce and associate stored and learned knowledge. Its purpose is to provide inspiration, offer advice, alleviate loneliness and low mood in users, and act as a coach or professional advisor / problem solver. Applications include: as a leisure companion (helping to alleviate boredom, stimulating the urge to talk, entertaining people, and providing assistance in life); as an inspiration for intellectual, scientific, and artistic impulses; as a coach or advisor providing psychological or intellectual support, especially for those with mental illness; as an advisor for various daily situations (fashion, hygiene, health, duties, care); as a personal secretary, capable of creating and using a wide range of knowledge databases; as a playmate for the most diverse games; and other tasks.

[0131] In particular, such a system can adapt to the user in the short term, such as adapting to their current mood, and in the long term, adapting to the user's character type. Among other things, information processed by the projection surface and subsequently stored in long-term memory can be used for this purpose. Ideally, the system could be equipped with codes, biometric user identification (image, voice, fingerprint, tone of voice, or other features), and other access control options for this purpose.

[0132] Preferably, such a system and the described method steps can realize an ethical system. For example, a personal intelligent companion can encourage satisfactory and useful actions for the user and his / her environment; it can draw attention to ethical issues related to adapting to the user's type, personality, and emotions, such as spreading certain virtues (helpfulness, generosity, kindness, courage, wisdom).

[0133] By establishing appropriate boundary conditions in the third assessment unit, the system can avoid harm, distress, and stress not only to the user but also to all those affected by their decisions. In situations where a serious erroneous decision is anticipated, the personal assistant can initiate discussions, particularly debating the consequences of certain course of action, and offering constructive suggestions for alternatives. The priority is not prescribing action, but rather how the action should or could ideally be performed. The personal assistant can identify dilemmas and point them out to the user, while simultaneously seeking alternative solutions or the most advantageous available solutions. In particular, the possibility of the personal intelligent companion providing support solely through pre-programmed and learned assessments in the user's artwork is significant. In such cases, considerations can already be tailored to the user. The system's associative capabilities play a crucial role here.

[0134] As another example of an embodiment, a "smart mirror" is described. A mirror is typically provided at the entrance or in the bathroom area. The input and output interfaces described in the preceding general examples, such as various sensors, can be easily integrated into the mirror. A variety of possibilities can be realized in such an AI system by using objects that users pass through briefly but frequently.

[0135] For example, appropriate cameras, microphones, motion detectors, ultrasonic sensors, artificial noses, and infrared sensors can be used to collect various information about users, including general conditions and user habits, without requiring users to actively input this information. This is particularly useful in entrance areas where access control can be implemented.

[0136] In this way, the intelligent AI system can alert users to clothing issues and provide clothing recommendations; for example, it can indicate the expected weather and necessary appliances if it detects that they have been forgotten. The items carried by the user can be recorded and identified. If necessary, issues can be clarified in a conversation with the user (via voice or other input methods), such as whether something is needed, or whether it has been forgotten or lost.

[0137] It can also include comments recorded by the microphone, allowing users to actively support these processes, for example, by commenting on situations or objects or by actively pointing them out for recording. Over time, the AI ​​system can use this method to know almost all the items, clothing, and their whereabouts in the apartment. If something is being searched, or if the user has questions about their clothing, food inventory, book inventory, etc., the system can provide hints. For example, it can tell the user they entered the apartment wearing glasses, thus determining that the glasses must be inside the apartment.

[0138] Through conversation, the AI ​​system can also manage appointment calendars, lists, and other daily life support tools. Therefore, the system is particularly suitable for the elderly, patients, or those who typically experience some form of limitation in their daily lives.

[0139] By appropriately evaluating the detected data, such as facial expression recognition and evaluation of speech pitch or other data, the system can detect a user's emotions in a short period of time and provide corresponding instructions based on this, such as if someone wants to rush to start a long journey. For example, in the above embodiment example, the detected emotions can be included in the evaluation performed by an evaluation third neural network.

[0140] In this scenario, the detection options and dialogue are not necessarily tied to an object, i.e., the smart mirror in this case. Therefore, to prevent users from having to engage in lengthy conversations with the AI ​​system in inappropriate locations, the system can initiate or continue the conversation with the user via speakers, microphones, and other devices distributed appropriately throughout the apartment. Furthermore, the components of the AI ​​system itself can be distributed across multiple modules and interconnected, for example, via suitable wireless or wired communication interfaces.

[0141] For all the AI ​​systems described here, it is best suited to use data that is collected and stored, especially personal data, and is subject to strict protection. To this end, an identification system can also be integrated, which can reliably identify users from images, voice, but also from motion features, assessments of voice emphasis, or any other biometrics. This prevents the disclosure of personal information to guests or other unauthorized persons during conversations.

[0142] The following further outlines other possible embodiments that can be implemented individually or in any combination of each other, as well as in conjunction with the foregoing embodiments:

[0143] For example, according to one embodiment, a method can be implemented in a system of at least two artificial intelligence units, the method comprising inputting input values ​​to at least a first artificial intelligence unit and a second artificial intelligence unit, thereby obtaining a first output value of the first artificial intelligence unit. Based on the output value of the first artificial intelligence unit, one or more modulation functions can be formed and then applied to one or more parameters of the second artificial intelligence unit. In this respect, the one or more parameters can be parameters that affect or influence in some way the processing of the input values ​​and the acquisition of the output values ​​in the second artificial intelligence unit. Furthermore, the output values ​​of the second artificial intelligence unit are obtained. These can represent, for example, the modulated output values ​​of the second unit.

[0144] In this way, two AI units are coupled together without using direct feedback of input or output values. Instead, one unit is used to influence the function of the second unit by modulating certain function-related parameters, resulting in a novel coupling that leads to different results or output values ​​than in traditional learning units. Furthermore, by processing input values ​​in both coupled units, results can be obtained in less time or with greater depth of analysis than in traditional systems, thus improving overall efficiency. In particular, it enables rapid classification of the problem at hand and consideration of rapid changes.

[0145] In an exemplary embodiment, at least one artificial intelligence unit may include a neural network with multiple nodes, particularly a learning unit that applies a modulation function. In this case, one or more parameters may be at least one of the following: weights of the nodes of the neural network, activation functions of the nodes, output functions of the nodes, and propagation functions of the nodes. These are fundamental components of the neural network that determine how data is processed within the network. Instead of defining new weights or functions for the nodes, the modulation function can be used to superimpose existing self-learning and / or predefined functions of the modulated network, depending on the results of the first artificial intelligence unit. Therefore, this application of the modulation function can also occur, in particular, outside the training phase of the network, and thus enable active coupling of two or more networks in the processing of input values.

[0146] According to an exemplary embodiment, each AI unit may be allocated a classification memory, wherein each AI unit performs the task of classifying input values ​​into one or more classes stored in the classification memory, wherein each class is constructed in one or more subordinate hierarchies, and wherein the number of classes and / or hierarchies in the first classification memory of the first AI unit is less than the number of classes and / or hierarchies in the second classification memory of the second AI unit. By making the classification memories of the two coupled AI units asymmetric in this way, parallel or time-dependent evaluation of input values ​​with different objectives can be performed, such as a combination of rapid classification of input values ​​and in-depth, slower analysis of input values.

[0147] Alternatively, or in addition to the asymmetric design of the classification memory, the complexity of the first AI unit and the second AI unit can also be designed differently, so that, for example, the first AI unit has a significantly lower complexity than the second AI unit. In this regard, for the case of neural networks, for example, the first neural network can have substantially fewer nodes and / or layers and / or edges than the second neural network.

[0148] In a possible embodiment, the application of at least one modulation function allows for the time-dependent superposition of the parameters of the second artificial intelligence unit. This modulation function may include one of the following characteristics: a periodic function, a step function, a function with a transiently increasing amplitude, a damped oscillation function, a beat frequency function as a superposition of several periodic functions, a continuously increasing function, or a continuously decreasing function. Combinations or time series of such functions are also conceivable. In this way, the relevant parameters of the learning unit can be superimposed in a time-dependent manner, such that, for example, the output value “jumps” into the search space due to modulation, which would not be possible without superposition.

[0149] Optionally, the second artificial intelligence unit may include a second neural network with multiple nodes, wherein applying at least one modulation function causes at least a portion of the nodes to deactivate. This type of deactivation can also be considered as an "exit" based on the output value of the first artificial intelligence unit, and may also provide new search regions for exploration in classification, as well as reduce computational overhead and thus accelerate the execution of the method.

[0150] In an exemplary embodiment, the method may further include determining the currently dominant artificial intelligence unit in the system, and forming the overall output value of the system from the output value of the currently dominant unit. In this way, two or more networks in the system can be meaningfully coupled and synchronized.

[0151] In this regard, for example, the first AI unit can be identified as the dominant unit at least before one or more output values ​​of the second AI unit are available. In this way, it can be ensured that the system is decision-safe at all times, that is, the system's response is possible at all times (after the first AI unit has run for the first time), even before all existing AI units in the system have fully classified the input values.

[0152] In this scenario, the comparison between the current input value and the previous input value can be further applied by at least one artificial intelligence unit of the system, wherein if the comparison results in a deviation exceeding a predetermined input threshold, the first artificial intelligence unit is designated as the dominant unit. In this way, it can be ensured that a response is immediately made to substantially changed input values ​​(e.g., sensor detection of new situations) through a new evaluation of the input value.

[0153] Alternatively, the current output value of the first artificial intelligence unit can be compared with its previous output value. If the comparison results in a deviation exceeding a predetermined output threshold, the first artificial intelligence unit is identified as the dominant unit. By calculating the deviation in the output value, for example, in the case of a deviation class compared to the previous run, meaningful changes in the input value can also be detected indirectly, thus making the new classification meaningful.

[0154] In some embodiments, the system may also include a timer that stores one or more predetermined time periods associated with one or more artificial intelligence units, the timer being arranged to measure the elapsed time of the predetermined time period associated with one of the artificial intelligence units each time. For example, such an element creates the possibility of synchronizing different units of the system and controls when the output value of a particular unit is expected or further processed. Thus, the timer can be used to define an adjustable waiting period for the entire system during which a decision should be available as the entire output value of the system. This time can be, for example, a few milliseconds, such as 30 or 50 milliseconds, and can be particularly dependent on the computing units and the existing topology of the existing computing units (processors or other data processing devices).

[0155] Therefore, for example, once the AI ​​unit is identified as the leading unit, measurements can begin on one of the AI ​​units for a specified, predetermined time period. In this way, it can be ensured that the unit develops a solution within the predetermined time, or alternatively, even suspend data processing.

[0156] In one possible embodiment, if a first time period has elapsed as predetermined for the first AI unit within a timer, the second AI unit can be designated as the leading unit. This ensures that the response based on the first AI unit is possible before the input values ​​are analyzed by other AI units, with the data subsequently analyzed in more detail by the second unit.

[0157] In any embodiment, the input value may include one or more of the following: measurements detected by one or more sensors, data detected by the user interface, data retrieved from memory, data received via a communication interface, and data output by a computing unit. Therefore, it can be, for example, image data captured by a camera, audio data, location data, physical measurements such as speed, distance measurements, resistance values, and any values ​​typically captured by suitable sensors. Similarly, data may be input or selected by a user via a keyboard or screen, and may optionally be associated with other data such as sensor data.

[0158] It should be understood that the above examples can be combined in any way. For example, in any of the embodiments described, combinations such as... Figure 4 or Figure 7 The timer described. Similarly, in all examples, the learning unit may include a classification memory, such as a combination of... Figure 5 or Figure 8 As described as an example, all these variations again apply to the coupling of two or more AI units.

Claims

1. A method in a system of at least a second artificial intelligence unit (720, 820) and a third artificial intelligence unit (730, 830), comprising: inputting a first input value to at least one second artificial intelligence unit (720, 820) and obtaining an output value based on the input value from the at least one second artificial intelligence unit (720, 820), wherein the input value is acoustic data, image data, sensor data and text; at least temporarily storing situation data, the situation data comprising the first input value and / or the second output value of the at least one second unit; using the situation data as an input value for the third artificial intelligence unit, the third artificial intelligence unit (730, 830) producing a third output value in response to the input value; and based on the third output value, checking whether the second output value of the at least one second unit fulfils one or more predetermined conditions; wherein the system further comprises a first artificial intelligence unit (710, 810), wherein one or more of the artificial intelligence units comprise a neural network, the method further comprising: The first input value X i is also input to the first artificial intelligence unit (710, 810) and a first output value of the first artificial intelligence unit is obtained; one or more modulation functions (f mod_f , f mod_w ) are formed on the basis of the output values of the first artificial intelligence unit (710, 810) and / or on the basis of the output values of the third artificial intelligence unit (730, 830); applying the one or more modulation functions (f mod_f , f mod_w ) to one or more parameters of the second artificial intelligence unit, the one or more parameters affecting the processing of input values and the obtaining of output values in the second artificial intelligence unit, each of the artificial intelligence units (810, 820, 830) is assigned a classification memory (812, 822, 832), wherein when producing the output value, each artificial intelligence unit classifies the input value into one or more classes (K1,... Kj, Km, Kn) stored in the respective classification memory (812, 822, 832), each class being structured in one or more subordinate hierarchies, and the number of classes and / or hierarchies in the classification memory of the first artificial intelligence unit (812) or the third artificial intelligence unit (832) is smaller than the number of classes and / or hierarchies in the classification memory (822) of the second artificial intelligence unit.

2. The method according to claim 1, wherein if the checking indicates that the output value does not fulfil at least one predetermined condition, discarding the second output value.

3. The method according to claim 1 or 2, further comprising deciding whether at least some of the current situation data should be permanently stored, and transferring the situation data to be permanently stored to a long-term memory.

4. The method of claim 3, wherein, The decision whether to permanently store at least some of the current situation data depends on the result of testing the second output value against one or more predetermined conditions.

5. The method of claim 1, wherein, Storing the situation data further comprises storing timing information for the input value and / or output value.

6. The method of claim 5, wherein, The timing information comprises one or more time intervals associated with one or more input values and / or output values.

7. The method of claim 1, wherein, The situation data is temporarily stored for at least one predetermined time period, and wherein at least one of the artificial intelligence units (710, 720, 730, 810, 820, 830) is adapted to set or change the time period.

8. The method according to claim 1, comprising comparing the current output value of the at least one second unit with output values stored in the situation data.

9. The method of claim 1, wherein, At least one of the artificial intelligence units (710, 720, 730; 810, 820, 830) comprises a neural network having a plurality of nodes, and wherein the one or more parameters are at least one of: weights of nodes of the neural network, activation functions of nodes, output functions of nodes, propagation functions of nodes.

10. The method of claim 1, wherein, The first input values comprise at least one of: measured values detected by one or more sensors, data detected by a user interface, data retrieved from a memory, data received via a communication interface, data output by a computing unit.

11. A system for performing the method according to any one of claims 1 to 10, the system comprising: two or more artificial intelligence units; means for obtaining input values input to the first artificial intelligence unit and / or the second artificial intelligence unit as first input values; at least one user interface for outputting a total output value to a user, wherein the total output value is formed based on output values from one or more of the artificial intelligence units.

12. A system for processing, the system comprising: at least a second artificial intelligence unit (720, 820) and a third artificial intelligence unit (730, 830), wherein the second artificial intelligence unit (720, 820) is arranged to receive first input values and to output an output value based on the input values, wherein the input values are acoustic data, image data, sensor data and text; further comprising a projection surface (750, 850) adapted to at least temporarily store situation data, the situation data comprising the first input values and / or the second output values of the at least one second unit; and wherein the third artificial intelligence unit (730, 830) is arranged to receive the situation data as input values and to produce a third output value in response to the input values; wherein the system is further adapted to check, based on the third output value, whether the second output values of the at least one second unit meet one or more predetermined conditions; wherein one or more of the artificial intelligence units comprise a neural network; wherein the system further comprises a first artificial intelligence unit arranged to receive the first output values and to form an output value based on the first input values; wherein the system is further arranged to form one or more modulation functions based on the output value of the first artificial intelligence unit and / or based on the output value of the third artificial intelligence unit, and to apply the formed one or more modulation functions to one or more parameters of the second artificial intelligence unit, the one or more parameters affecting the processing of input values and the obtaining of output values in the second artificial intelligence unit; wherein the system is further arranged to form one or more modulation functions based on the output value of the first artificial intelligence unit and / or based on the output value of the third artificial intelligence unit, and to apply the formed one or more modulation functions to one or more parameters of the second artificial intelligence unit, the one or more parameters affecting the processing of input values and the obtaining of output values in the second artificial intelligence unit; wherein a classification memory (812, 822, 832) is associated with each of the artificial intelligence units, wherein each of the artificial intelligence units is arranged to classify the input value into one or more classes (K1,..., Kj, Km, Kn) stored in the classification memory (812, 822, 832), each class being structured in one or more subordinate levels, and the number of classes and / or levels in the classification memory of the first artificial intelligence unit (812) or the third artificial intelligence unit (832) is smaller than the number of classes and / or levels in the classification memory (822) of the second artificial intelligence unit m.

13. The system of claim 12, Further comprising at least one sensor element and / or at least one detection unit, wherein The value detected by the sensor element and / or the detection unit is used at least in part as the first input value X i .

14. The system according to claim 12 or 13, further comprising a storage element (852) of the projection face, wherein, The case data is stored.

15. The system of claim 12, further comprising at least one output module for outputting an output value to a user, wherein, The output module comprises at least one of a display screen, a touch screen, a loudspeaker, a projection module.

Images