WO2006005665A2

WO2006005665A2 - Method for reacting to changes in context by means of a neural network, and neural network used for reacting to changes in context

Info

Publication number: WO2006005665A2
Application number: PCT/EP2005/052859
Authority: WO
Inventors: Gustavo Deco; Martin Stetter
Original assignee: Siemens Aktiengesellschaft
Priority date: 2004-07-09
Filing date: 2005-06-21
Publication date: 2006-01-19
Also published as: WO2006005665A3

Abstract

Disclosed is a method for reacting to changes in context by means of a neural network (1). According to said method, input data is fed to an input layer (10) of the neural network (1) while output data is retrieved from an output layer (30) of the neural network (1). Several models which predefine one respective copy of the input data to the output data are stored in the neural network (1), only one model being active at the same time. The following steps are then repeated: the neural network (1) copies a piece of fed input data (2) to a piece of output data (3). Said piece of output data (3) is then retrieved from the output layer (30). If the retrieved piece of output data (3) is incorrect in a current context for the fed piece of data (2), the neural network (1) receives an error feedback message (4), whereupon the neural network (1) activates another model.

Description

description

A method of responding to contextual changes with a neuronal network and neural network to respond to contextual changes

The invention relates to a method for responding to text changes with a neural network, as well as a neuronal network, which can respond to context changes.

Context is understood to be an external meaning context which indicates which behavior makes sense in a particular situation.

For a stock price, the context could indicate whether the price is going up or down, and depending on that context, it would be a wise decision to buy or sell the stock. The context may also indicate that a change in the weather is imminent, and depending on this change in the weather, a positive or negative weather forecast would make sense. The context may also be a variable strategy of a player of a computer game, and depending on this strategy, a particular behavior of a character may be useful in the computer game (e.g., attack or defense).

The invention thus relates to a method of data processing in order to find a correct behavior for a given situation. Behavior is already understood here as the estimation of the situation, but the behavior can also include a decision or an action.

Intelligent software agents are known which are developed using methods of artificial intelligence. These are rule-based systems which have a database with rules and a logic for processing the rules. The rules are explicitly formulated and must be manually entered by the developer. It is a declarative programming model. An example of such intelligent agents are BDI agents (lieve-desire-intention). These collect sensory data, process it with a set of rules and choose a behavior.

A known development is the implementation of a fuzzy logic, which, however, also requires an explicit rule formulation. In psychology, other models of human behavior are known, which are also based on explicit rules.

The invention also relates to dynamic reinforcement learning, in which a system or agent receives feedback on its behavior, which serve to enable it to flexibly respond to context changes. If the system or the agent z. If, for example, a buy decision is made and the share price subsequently falls, the system or the agent can learn by means of appropriate feedback and choose a more correct behavior in the next situation. Behavior is a decision, a situation assessment or an action ver¬ stood. The evaluation of a situation can also be called a feeling. Dynamic gain learning is described in document [1].

Neural networks are known in particular as learning systems. Information about the current situation in the form of input information is fed to such a neural network. The input information is processed by the neural network. Subsequently, output information can be taken from the neural network. The output information describes the behavior of the neural network and thus represents a decision, a situation evaluation or an action instruction. A neural network can be trained so that it learns which output information is correct for a given input information. This is called synaptic learning. Synapses are the connections between the individual neurons, the elements of the neural network. By expressing the synaptic strengths, the neural network learns in this context the correct mapping of input information to output information.

The advantage of neural networks lies in the fact that the rules for mapping the input information to output information do not have to be specified explicitly and declaratively. Rather, the neural network learns an implicit rule representation from the data with which it is trained. The disadvantage here, however, is that the neural network can not react flexibly to context changes. A change in context, that is to say a change in the external context of meaning, constitutes the requirement for the neural network to convert the mapping of input information to output information with immediate effect. However, synaptic learning is an incremental, time-delayed process which does not allow a flexible and rapid response to context changes.

Thus, the task arises to indicate a method for the reaction to context changes and a data processing unit which can react to context changes.

In the method for responding to context changes with a neural network, input information is supplied to an input layer of the neural network. Furthermore, output information is taken from an output layer of the neural network. Furthermore, several models are stored in the neural network, each of which predefines a mapping of the input information onto the output information. Only one model can be active at the same time. The following steps are then repeated: The neural network forms an inputting information with the active model on an output information. This is then taken from the output layer. If the extracted output information for the supplied input information is incorrect in a current context, the neural network receives a false feedback, whereupon the neural network activates another model.

Preferably, an action is derived and executed from the extracted output information.

In a further development, the neural network contains exciting pulsed neurons. These form model pools, with each model being assigned at least one model pool. The model pools compete with each other, whereby an active model pool prevails in the competition.

According to one embodiment of the invention, the neuronal network contains inhibiting pulsed neurons. These form at least one inhibitory pool. The inhibitory pool exerts a global inhibition on the competing model pools. The false feedback activates the inhibiting pool. The activated inhibiting pool performs a complete inhibition of all model pools. The complete inhibition then deactivates the active model pool. After complete inhibition, another model pool is activated.

According to a development of the invention, sys- tems adapt to the exciting pulsed neurons of the active model pool. As a result, recurrent weights of the active model pool decrease. This leads to the fact that the active model pool is subject to the complete inhibition in the competition compared to the other model pools.

Preferably, the adaptation of the synapses is implemented as short-term synaptic depression (STD). According to one embodiment of the invention, exciting pulsed neurons form rule pools. Each of these rule pools ver¬ switches each one of the input information with one of the output information. The rule pools compete with each other, with the active model pool supporting a selection of rule pools. The supplied input information activates a rule pool from this selection. The appropriately activated rule pool activates the output information to be taken. This will be taken out afterwards.

In a particular embodiment of the invention, the neural network includes interconnections formed by Hebbian learning.

The neural network for responding to context changes has an input layer to which input information can be supplied. Furthermore, there is an intermediate layer, by means of which the input information can be mapped onto output information. Furthermore, there is an output layer from which the output information can be taken. Finally, the neural network also has a model layer with which the image formation in the intermediate layer can be controlled as a function of a context.

According to a development, the model layer contains exciting pulsed neurons. Furthermore, the model layer contains several model pools, which consist of the exciting pulsed neurons. With the model pools, the image in the intermediate layer can be controlled as a function of a context.

According to one embodiment of the invention, the neural network contains inhibitory pulsed neurons. Furthermore, the neuronal network contains an inhibiting pool, which consists of the inhibitory pulsed neurons. The inhibiting pool is interconnected with the model pools. Furthermore, the inhibiting pool can be activated by an incorrect feedback if an extracted output information is incorrect. According to one development of the invention, the neural network has the model layer in a first module and the other layers in a second module.

In a particular embodiment, the input layer for each input information has an input pool, which consists of exciting pulsed neurons and can be activated by supplying the respective input information. Furthermore, the intermediate layer has rule pools, by means of which an input information can be interconnected with an output information. In this case, the rule pools consist of exciting pulsed neurons. Further, for each output information, the output layer has an output pool consisting of exciting pulsed neurons. Furthermore, the model pools are connected to the rule pools such that only a selection of rule pools can be activated depending on the activation of the model pools. Furthermore, a representative input pool can be activated by an input data input, which represents the input information supplied. Furthermore, a preferred rule pool can be activated by the representative input pool, which belongs to the selection of rule pools which are connected with an active model pool. Finally, a representative output pool can be activated by the preferred rule pool, wherein the representative output pool represents a removable output information.

The neural network may further comprise one or more additional input layers, interlayers, output layers, model layers or other layers. These additional layers may store, filter, rate, network or categorize the input information, the output information or other information, or may perform other functions. The advantage of the method according to the invention is that the rules for the behavior are extremely implicitly represented and do not have to be specified explicitly. Furthermore, it is possible to change the mapping of input information to output information even after a single false feedback. As a result, the method achieves great speed and flexibility in the response to context changes, which distinguishes it from conventional neural networks.

Furthermore, the invention describes a quantitative mathematical model with direct biological relevance, in that it specifically depicts the neurodynamic processes which were also recognized in biological research. The behavior thus results as an autonomous, emergent phenomenon and has clear advantages over conventional agents with declarative Regel¬ representation on.

In contrast to the declarative rule formulation in conventional BDI agents, the invention permits not only the implicit rule representation but also dynamic rule selection, which also includes a random element. The random element results from the emergent, free interaction of neurons and pools of the neural network.

In the following, the invention will be explained in more detail by means of exemplary embodiments, which are illustrated schematically in the figures. The same reference numerals in the individual figures indicate the same elements. In detail shows:

Fig. 1 shows a first embodiment of the neural network; Fig. 2 shows a second embodiment of the neural network; Fig. 3 weights of a first module; Fig. 4 weights of a second module; Fig. 5 shows the structure of a pool. FIG. 1 shows a neural network 1 with an input layer 10, an intermediate layer 20, an output layer 30 and a model layer 40. The neural network 1 is supplied with input information 2 and later with an output information 3. The input layer 10 has strong connections to the intermediate layer 20, which in turn has strong connections to the exit layer 30. The model layer 40 has strong connections to the intermediate layer 20.

A strong connection between two layers means that the layers have an above-average number of compounds or that the compounds are markedly above average. The connections are usually made via synapses, which connect neurons of one layer with neurons of the other layer. A strong connection of two layers thus means that there are particularly many synaptic connections between the layers or that the synaptic connections are particularly pronounced. The synaptic strength can be described by a weight w. Higher values for a synaptic weight w mean a stronger synaptic connection between the participating neurons.

The neurons of the neural network can be partially or completely linked in the layers and over the layers. In the case of full connectivity, full connectivity exists, meaning that each neuron is linked, linked or networked with each other neuron.

According to one embodiment of the invention, the layers are fed back together, that is, there are also strong connections in the opposite direction. This feedback leads to a shift in the equilibrium in the competition of the individual neurons or groups (pools) of neurons. The thickness of the forwardly directed compounds, ie from the input layer 10 via the intermediate layer 20 to the output layer 30 and from the model layer 40 to the intermediate layer 20, is expediently stronger than the strength of the backward-looking connections.

Through the intermediate layer 20, the supplied input information 2 is displayed on the extracted output information 3. The intermediate layer 20 has stored for this purpose different images, one of which is activated by the model layer 40 in each case. As a result, a rapid response to context changes is possible since the mapping in the intermediate layer 20 can be flexibly changed.

According to one embodiment, the input layer 10, the intermediate layer 20, the output layer 30, and the model layer 40 contain groups (pools) of neurons. It is also possible that only some of these layers contain pools. The pools of each layer may contain excitatory pulsed neurons. Exciting pulsed neurons are activated by pulses from other exciting pulsed neurons and send out pulses to other neurons themselves. The activity of a pool of pulsed neurons can be modeled using a mean-field approximation.

The pools are attached to the biology of the human brain ange¬. In the human brain, large and homogeneous populations of neurons that receive a similar external input, are mutually coupled, and probably together act as a single entity, forming groups (pools). These pools can be a more robust processing and coding unit because their instantaneous population average response, as opposed to the time average of a relatively stochastic neuron in a large time window, is better suited to the analysis of fast changes in the real world. If one neuron can activate another one, this means that there is a strong connection between the two neurons. The same applies in the event that one pool activates another pool. This means that there is at least one strong connection of at least one neuron of the first pool with a neuron of the second pool.

FIG. 5 shows that the neurons 101 of a pool 100 are also strongly connected to one another. As shown in FIG. 5, there may be a partial connection, but also a complete networking of the neurons 101 of the pool 100. The neurons 101 of the pool 100 are linked via strong connections 102, thus contributing to the activation of the pool 100 in mutual support. The greater the activity of the neurons 101, the greater the activity of the pool 100.

The activity of the neurons 101 can be described by mathematical models. Further information on the mathematical modeling of pools in the development of neural networks as well as different mathematical models for pulsed neurons, as used in the exemplary embodiments are u. a. from the writings [2], [3] and [4] known.

If multiple pools 100 compete in a layer, it means that the connections between the pools 100 in the layer are weak. This means that the pools 100 in one shift do not support each other. If the pools 100 are now also subject to a global inhibition, only one pool 100 with the greatest activity will be able to prevail in the competition.

In the context of the invention, neurons 101 are always artificial neurons 101. These partially or completely model a particular type of neuron known from biology. The modeling can be done by an electronic circuit or a mathematical model. gen, which is calculated by a data processing system. By connecting two pools 100, it is meant that the neurons 101 of these pools 100 are strongly or weakly connected to one another, ie that, for example, B. many or few, strong or weak synaptic connections between the

Neurons of the one and the neurons 101 of the other pool 100 exist.

The strong connections 102 between the neurons 101 of the pool 100, shown in FIG. 5, serve for the self-amplification of the pool 100. The strength of these strong connections 102 corresponds to z. For example, the synaptic strengths between the neurons 101, which are referred to within the pool 100 as recurrent weights.

According to the invention, the neural network 1 stores a model for each context, ie for each class of situations. The model indicates how the input information should be mapped to the output information. The model thus describes which behavior is correct in a particular situation. Different situations may require different behavior. This is always the case when the context changes from one situation to the next. This means that in the second situation different output information must be selected than in the first situation. Different situations can therefore be based on a different context. With the models, the neural network 1 now attempts to specify, for each context, for each class of situations, how the input information is to be mapped to the output information.

If the neural network 1 is then confronted with a changed situation on which another context is based, then it is only necessary to change its model. According to one embodiment of the invention, the intermediate layer 20 contains rule pools 21 which compete with each other, whereby only one activated rule pool 23 can enforce itself. The rule pools 21 each represent a possibility of interconnection of an input information with an output information. The input control information 2 is interconnected with the output information 3 to be extracted by the activated rule pool 23.

This interconnection consists of strong connections between neurons of the input layer 10, neurons of the control pool 23 and neurons of the output layer 30. Interconnection thus means that neurons, pools or layers are not only linked to one another, which is already present in a full interconnection, but instead that the weights of the links are strong.

According to one embodiment of the invention, the model layer 40 contains model pools 41 that compete with each other, whereby only one active model pool 42 can prevail. The active model pool 42 is interconnected with a selection 22 of rule pools 21. This means that the rule pools 21 in the selection 22 are supported in their competition with the other rule pools 21 as long as the active model pool 42 is activated. The active model pool 42 thus determines which

Rule pools 21 can be activated, and thus, how the zu¬ guided input information 2 is mapped to the output information 3 to be taken. Which rule pool 21 is activated in the selection 22 depends on the supplied input information 2. Which output information 3 is taken depends on the activated rule pool 23.

The above-mentioned models are thus formed in each case by a model pool 41 and by the rule pools 21 connected to the model pool 41 as well as their interaction. According to one embodiment of the invention, the neuronal network 1 contains an inhibiting pool 50, which is formed from inhibitory, pulsed neurons. The inhibiting pool 50 is connected to the model layer 40 and exerts a global inhibition on the competing pools 41. This is necessary so that only one active model pool 42 can prevail in the competition.

The inhibiting pool 50 is stimulated here by the model pools 41. This means that the active model pool 42 stimulates the inhibitory pool 50 by its activity and thereby indirectly inhibits the other model pools 41. According to the same principle, competition in the other layers can also be implemented.

If one pool activates or excites another, this means that strong connections exist between the neurons of the two pools. The excitation or activation involves excitatory compounds, eg. For example, strong, excitatory, synaptic connections. In the case of inhibition are inhibitory compounds, eg. B. inhibitory, synaptic connections.

According to one embodiment of the invention, the inhibiting pool 50 is also strongly connected to the other layers or there are one or more inhibiting pools which are interconnected with the other layers, d. H. have an inhibiting influence on them.

If the extracted output information 3 was false, the neural network 1 must be informed that its current map or model does not correspond to the current situation or context. For this purpose, the inhibiting pool 50 is activated by a false feedback 4, whereby the global inhibition becomes complete

Inhibition of all model pools 41 is reinforced. As a result, the active model pool 42 is deactivated with all other model pools 41. fourth. The active model pool 42 is then in competition with the other model pools 41, thereby activating another model pool 41.

The false feedback 4 can z. B. be implemented by activating the neurons of the inhibitory pool 50 via the external synaptic inputs.

So that the active model pool 42 is subject to complete inhibition in the competition, according to an embodiment of the invention, synapses of the exciting, pulsed neurons of the active model pool 42 are adapted such that the self-amplification of the active model pool 42 decreases over time , By reducing the synaptic strengths of the exciting, pulsed neurons of the active model pool 42, the recurring weights of the active model pool 42 are also reduced. Such a method is known from specialist literature as "short-term synaptic depression (STD)". However, other methods may be used.

After the active model pool 42 is inferior in the competition, another model pool 41 is activated, the z. For example, a selection 24 of rule pools 21 is supported. So now another picture is chosen, which takes into account the context change. In the case of the next input information 2 supplied, the previously activated rule pool 23 would be in competition with the rule pools 21 from the selection 24, since these are now supported by the model layer 40. As a result, the next supplied input information 2 is interconnected with another removable output information 3.

According to one embodiment of the invention, the input layer 10 also has input pools 11 which compete with each other, whereby only one input pool 11 can ever prevail in the competition. In this case, each input information can be assigned its own input pool 11. It However, preprocessing may also take place, so that from the input information certain features are extracted, which are subsequently represented by an input pool 11. Thus, an input pool 11 for angular shapes and another input pool for round shapes can be activated.

In one variant of the invention, the output layer 30 also has output pools 31 which compete with one another, whereby again only one output pool 31 can prevail in the competition.

FIG. 1 shows the case that an active model pool 42 supports a selection 22 of rule pools 21. Furthermore, a representative input pool 12 is activated by the supplied input information 2. This activates the activated rule pool 23 within the selection 22. This in turn is interconnected with a representative output pool 32, for which reason it is activated. Subsequently, the output information 3 is taken, which is represented by the representative output pool 32.

According to one embodiment of the invention, the layers are excited by at least one non-specific pool, which consists of exciting pulsed neurons. This device does not receive any specific inputs from one of the layers and contributes with spontaneous pulses to the formation of realistic pulse distributions. The neurons of the non-specific pool are not correlated with the other layers or pools, i. H. they are not specifically activatable, as they may be networked with the other layers / pools, but not with strong connections.

The neural network 1 can be designed as an attractor-recurrent autoassociative neural network. Such networks are described in the cited documents. In this case, the synaptic connections from the input layer 10 to the intermediate layer 20, from the intermediate layer 20 to the output layer 30, and from the model layer 40 to the intermediate layer 20, respectively, are made stronger than in the return direction.

According to one embodiment, it is assumed that the synaptic strengths between the neurons of the neural network 1 are designed as if they had been formed by Hebbian learning. This means that the synaptic strength between neurons, which have a correlated activity, is strong, whereas the synaptic strength between uncorrelated neurons is weaker. Accordingly, the synaptic strengths between the excitatory pulsed neurons within one of the excitatory pools, ie an input pool 11, a rule pool 21, an output pool 31 or a model pool 41, with a strong weight of z. B. w = 2.1 executed. The inhibitory neurons in the inhibitory pool 50 or other inhibitory pools may be interconnected at a weight of w = 1, which contributes to non-oscillatory behavior. The inhibiting neurons may also be connected to all exciting neurons of the respective module or the entire neural network 1 with the same weight w = 1. By contrast, the weight between two exciting pulsed neurons, which lie in different exciting pools of the same layer, is preferably weak. The neurons of the excitatory pools are furthermore preferably connected to neurons in the non-selective pool with a forward synaptic strength of w = 1 and a reverse synaptic strength, which likewise falls weakly.

According to one embodiment of the invention, each rule pool 21 is connected to an input pool 11, a model pool 41, and an output pool 31 as if the connection were through

Hebbian learning would have been trained while practicing different behaviors. The synaptic strength the connections between the model pools 41 and the Regel¬ pools 21 is z. At w = 1.1.

The interplay between the layers corresponds to the multi-area interconnection in the human brain. The neural network 1 can also have further layers. These additional layers can form the functions of specific brain areas. In this way, the functionality of the neuronal network 1 can be considerably expanded. It is conceivable z. As the filtering of input information by modeling ei¬ ner selective attention, as well as the implementation of work or long-term memory functions. The further layers can be constructed in the manner described and interact with one another, or else implement other known methods.

The neural network 1 can be designed as a neurodynamic network, in particular in the form of a neurodynamic network of pulsed neurons. This may include the use of known neural networks, multilayer

Perceptrons, SOMs (seif organizing maps) etc. include. The pulsed neurons can z. B. be formed as so-called spiking or as so-called pulse-coding neurons.

Further notes on the implementation of the layers, pools, connections and neurons as well as their interplay can be found in [2], [3] and [4]. The dynamics of the respective neurons, pools and layers can be implemented by one or more concepts of the following group: mean-field approach, integrate-and-fire approach, approach to pulse-coding neurons, multicompartment approach and Hodgkin's -Huxley approach. Reference is made to the cited documents.

Intelligent agents which use the method according to the invention or the neural network can be used for neurocognitive process control, in particular for technical processes, for neurocognitive driver assistance and for neurocognitive robot control.

The invention is based on the principle of influenced competition (Biased Competition). On the one hand, an active model, which is formed by the active model pool 42 and by the interconnections of the rule pools 21 which belong to the selection 22, acts on the mapping of input information to output information in a top-down manner. On the other hand, the input information acts as an opposing influence (bottom-up), which activates certain neurons and pools. The actual interconnection of the supplied input information 2 with the extracted output information 3 develops from an interplay of these two influences. The neural network 1 thus achieves a high degree of flexibility, which in its behavior, ie. H. in its mapping of input information to output information flexible to context changes, d. H. changed situations, can react.

Since the neural network 1 directly depicts the neurodynamic processes in the human brain, it is of direct biological relevance. This is based on the exact simulation of the biological processes in the brain. For example, the pulsed neurons may be integrated-and-firing neurons

(integrate and fire). There may be talk of neurocognitive modeling, i. H. The human behavior is modeled directly by the involved neuronal processes.

This results in a number of medical-clinical applications. Thus, the neural network and method according to the invention can be used as a model for clinical trial results. First, experiments are conducted with a subject to examine behavior in situations of changing context. Subsequently, the method according to the invention with a neural network becomes the same Test conditions carried out. By comparing the results, hypotheses about the functioning of the brain can be checked directly.

The same applies to pathological disorders. Thus, in the simulation with the method according to the invention, modified chemical conditions, eg. As metabolic disorders or ver¬ changed concentrations of neurotransmitters, be taken into account. By means of this inverse modeling, the behavior, for example of a schizophrenic patient, can be simulated by the method according to the invention. From the results and the comparison with test results with schizophrenic patients, hypotheses concerning the nature of the disorder in the brain can be checked again. The same applies to other disorders.

An intelligent agent which implements the method according to the invention or the neural network according to the invention can be used in an extremely versatile manner. For example, such an agent could be used for chart analysis. Here, an attempt is made to conclude from the history of a stock price on its future development. In this application, the neural network 1 is fed as input information 2, for example, the price history of a share during the last six months. The extracted output information 3 now represents a purchase or sales recommendation. If the recommendation turns out to be false, the neural network 1 receives a false feedback 4. In this way, the neural network 1 can react flexibly to context changes, for example, when the mood on the stock exchange changes. At the next input information input 2, e.g. If, for example, the price of another share during the past six months, the neural network 1 can already take into account the changed mood on the stock market, ie the changed context. The invention is particularly suitable for complex connections that can not be detected by a declarative rule model. These include developments on the stock markets or the weather forecast. The neural network 1 is figuratively in a position to make a gut decision. This is because _/ that the extracted output information 2 can also be interpreted as a feeling. The feeling is based on an intuitive interpretation of the input information, whereby intuitively means that the input information is mapped to the output information by implicit, not explicitly represented rules. In this context, the extracted output information 3 gives as feeling a general direction of trade. Based on the feeling that the stock price will fall, different choices can be made, such as holding a stock position, reducing it, or selling it all. If there is a feeling that the share price is going to rise as the extracted issue information 3, the new purchase as well as the stocking or holding of an equity position are possible as actions. The behavior of the intelligent agent is thus made more flexible by the intermediate step of situational evaluation with the aid of a feeling as extracted output information 3.

Intelligent agents with flexible, human-like behavior can also be used elsewhere, for example as characters in training simulations and computer games.

The invention relates inter alia to the field of cognitive flexibility and behavioral change. The dynamic reinforcement (Dynamic Reinforcement) supports the Ler¬ nen. In one embodiment of the invention, the neuronal network 1 as an image in the intermediate layer 20 forms associations between input information and a reward or punishment. Due to the interplay with the model layer 40, these associations can already be reversed after a false feedback 4. The false feedback 4 always follows when a reward expectation has been taken as output information 3 and this does not occur or punishment occurs. The false-feedback 4 can also take place when it is taken as output information 3 that the neural network 1 expects a punishment, but this fails or a reward occurs. In this way, the neural network 1 learns when to reward, and when to expect punishment, and responds with a feeling as output information 3 to be extracted, representing this maintenance, and from which an appropriate action can be derived ,

Figure 2 shows a specific embodiment of the invention. The neural network 1 is here divided into a first module 5 and a second module 6. The second module 6 is a

Input information 2 supplied and an output information 3 taken. If the output information is incorrect, the first module 5 is supplied with a false feedback 4.

The first module 5 has a model layer 40. The second module 6 has an input layer 10, an intermediate layer 20 and an output layer 30.

The input layer 10 includes two input pools 15 and 16 which compete with each other. If a first object is supplied, the input pool 15 is activated. If a second object is supplied, the input pool 16 is activated.

The intermediate layer 20 has a selection 22 and a selection 24 of rule pools. The selection 22 contains rule pools 25 and 26, the selection 24 contains rule pools 27 and 28. The output layer 30 contains output pools 35 and 36. The model layer 40 contains model pools 45 and 46. Within each tier, the pools compete with each other. The competition in the model layer 40 is supported by an inhibiting pool 50, which consists of inhibitory neurons, and can be additionally excited by the false feedback 4. The competition in the layers of the second module 6 is aided by an inhibiting pool 60 which is connected to these layers and exerts a global inhibition on the contained neurons. The first module 5 is assigned a non-specific pool 70 whose activity is not influenced by the layers of the two modules. Similarly, the module 6 contains a non-specific pool 80 whose activity is not affected by the layers. The non-specific pools 80 and 70 contribute with spontaneous pulses to the formation of realistic pulse distributions.

In a specific embodiment of the invention, the second module 6 is implemented with 1600 pulsed neurons (here pyramidal cells) and 400 inhibiting pulsed neurons. The first module 5 can be implemented with 1000 exciting pulsed neurons and 200 inhibitory neurons. These numbers are only an example. The ratio of excitatory pyramidal cells to inhibitory neurons may be e. B. 80:20 amount. According to one embodiment of the invention, all neurons are fully interconnected. The pools can z. B. from 80 or 100 neurons are formed.

In a specific embodiment, all the pools of the first module including the inhibiting pool 50 and the non-specific pool 70 are networked with each other. Furthermore, all the pools of the second module 6, including the enclosing pool 60 and the non-specific pool 80, are networked with one another.

The networking of the pools takes place via synaptic connections between the neurons contained in the pools. The strength of the synaptic connections between the individual pools for the first module is shown by way of example in FIG.

A synaptic strength of w _w stands for a weak synaptic strength, for example of w _w = 0.878. A sy- The naptic strength of w _s stands for a strong synaptic strength (for example of w _s = 2,1). From Figure 3 shows that the model pools 45 and 46 are interconnected only weakly. The connections of the neurons within one of the two pools are strong. In this way, the neurons of the two pools form two teams that compete with each other. The connections from the model pools 45 and 46 to the non-specific pool 70 are made in neutral strength (w = 1). In the opposite direction, the connections of the non-specific pool 70 to the model pools 45 and 46 are weak.

FIG. 4 shows examples of synaptic strengths of the connections between the pools of the second module, including the inhibiting pool 60 and the non-specific pool 80

Pools of the layers are each self-crosslinked with a strong synaptic strength w _s . The connections to the other pools the respective same layer fall ^"on the other hand from weak (w _w). This also serves the competition. Of particular importance are the weights or synaptic strengths Wff and Wf]. _3, the synaptic strength is the Wff Staer ¬ ke of a feed forward, for example in the amount of Wff = 2.1 The synaptic strength Wf] ₃ is the synaptic strength of a backward connection (feedback), for example Wf] ₃ = 1, 7. It now appears that the input pool 15 is strongly connected to the rule pools 25 and 27 (with Wff) .In the reverse direction there is also a strong, but somewhat weaker synaptic strength (with Wf] ₃ ) lies between the input layer 10 and the intermediate layer 20 a recurrent

Connection before. This means that a dynamic interplay develops between the two layers. Although the coating 10 exerts the greater influence on the intermediate layer 20, processes in the intermediate layer also have a recurring effect on the input layer. In this way, selective processing of input information can be implemented. The rule pools 25 to 28 in turn have strong forward directed connections to the output pools 35 and 36. Again, backward, recurrent connections exist (with w _w ). The input layer 10 is thus recurrent with the

Intermediate layer 20, and the intermediate layer 20 recurrently connected to the Ausga layer 30. The synaptic strengths of FIG. 3 and FIG. 4 correspond to the values which would have to be expected by training the neural network 1 by means of Hebbian learning.

The model pool 45 is strongly connected to the rule pools 25 and 26 (synaptic strength, for example, w = 1.1), whereby it forms the selection 22. The model pool 46 is strongly connected to the control pools 27 and 28 (synaptic strength for the example w = 1.1), whereby it forms the selection 24. Illustratively, the input pool 15 is interconnected by the rule pool 25 of the selection 22 with the output pool 35. Likewise, the input pool 16 is connected to the output pool 36 via the rule pool 26 of the selection 22. The interconnection of the input pools 15 and 16 with the output pools 35 and 36 takes place via the control pools 27 and 28 of the selection 24 exactly opposite. Thus, in an anecdotal perspective, the selection 22 and the selection 24 provide two different maps of input information to output information. Which of the two images is used depends on which of the two model pools 45 and 46 is activated.

The output pool 35 and the output pool 36 may each represent a situation score (eg, positive / negative), a decision (eg, buy / sell), or an action (eg, go forward / backward).

According to one embodiment of the invention, the supplied input information 2 is transmitted to all neurons of the neural network 1, for example via in each case 800 connections from outside the network Network supplied. In addition to the input information, background noise can also be supplied via these connections, which represents a spontaneous firing of neurons outside the network.

The states in which the neural network 1 can stabilize can also be referred to as global attractors. These are each composed of individual attractors for each pool. When a pool wins in the competition, the activity in that layer converges to the attractor concerned. The global attractor is thus composed of a combination of activated pools in the input layer 10, in the intermediate layer 20, in the output layer 30 and in the model layer 40. The optional recurrent connections between the layers give rise to controlled competition in one, several or all layers. This leads to an autonomous, emergent and highly flexible behavior of the neural network 1.

In a concrete application scenario, the course of a stock price during the last six months is fed as input information 2. As a result, the input pool 15 is activated. Furthermore, the model pool 45 is activated. He represents a model that assumes a positive mood on the stock exchange. The model pool 45 controls the competition in the intermediate layer 20 in such a way that the pools 22 in the selection 22 can compete in the competition. Within the selection 22, the rule pool 25 is activated, which is strongly connected to the input pool 15. The rule pool 25 in turn activates the output pool 35 via a strong connection, which represents the output information "buy share", which is taken from the neural network as output information 3. While model pool 45 is active, its self-boosting is depleted, that is, the synaptic weights of the model pool 45's internal connections decrease over time. When the removed Issue information 3 was incorrect because the stock price fell after the purchase, the inhibitory pool 50 is activated via a false feedback. This then amplifies its global inhibition to a complete inhibition of the model pools 45 and 46, which can also extend to the second module 6. Subsequently, the formerly active model pool 45 is in competition with the model pool 46, since the self-amplification of the model pool 45 has exhausted itself compared to the self-boosting of the model pool 46. Therefore now the model pool 46 wins in the competition. The model pool 46 now supports the selection 24 in the intermediate layer 20. If now the same price course is performed as input information 2, the input pool 15 is again activated. This now activates the rule pool 27, since this belongs to the selection 24, which is supported by the model pool 46. The rule pool 27 in turn activates the output pool 36. This represents the output information "Do not buy stock", which is subsequently taken as output information 3. The neural network 1 has thus adapted its behavior in one step to the changed context, for example a change in mood on the stock market.

Suitable values for the synaptic starches may deviate from the stated values and can be determined or optimized in the experiment. The general procedure for this is described in the document [2].

The embodiments shown in the figures are aus¬ finally listed as possible examples.

In particular, the readout of the output information 3 can also be realized differently. For example, it is conceivable that this takes place from each layer or from the respective pools.

Furthermore, a plurality of input, output, intermediate or model layers are conceivable. Different input layers 10 could represent different features (color, shape, size, location, motion, etc.) of the input information.

Different model layers 40 could have different aspects of a model (different dimensions of a model)

Context). For example, In a computer game, a model layer 40 could store whether the player is practiced or untrained, as well as another which strategy the player is currently following. Thus, the model could represent a context that is composed of different aspects.

Different issue stories 30 could represent different aspects of issue information (e.g., buy / sell, urgency, security of recommendation).

In these scenarios, the intermediate layer 20 could also be implemented in multiple layers. In any case, it would have the task of networking the additional layers in a meaningful way.

Additional and existing layers can be organized both according to the principle of competition and according to the principle of cooperation. Competition: Within a shift certain characteristics or feature groups compete with each other for representation. This produces a weighting map (salicency map) as an emergent process, so that certain features are more intensely represented than others. A context-dependent information selection is obtained. This can be implemented by competing pools.

Cooperation: Features can also be linked dynamically to feature groups or categories. One feature may also favor the representation of another feature. This can be implemented by pools that are connected to each other with strong weights and thus terstützen. For example, several pools can be activated simultaneously in the input layer 10 and thereby simultaneously represent several properties of an input information.

Additional layers can be implemented according to the principle of influenced competition and biased competition and cooperation: Through the connection between layers, a layer can direct the competition process in one or more other layers. This process can be recurrent, so that successively and dynamically an ever better matching of different feature spaces of the different layers arises with each other through this mutual steering process. In particular, because it covers only a partial aspect of the environment, each representation inevitably contains ambiguities. Influenced competition represents a mechanism by which the various layers can resolve ambiguities in the respective other feature spaces by the information of their particular feature space. Each representation evolves before the context of all other representations. Cooperation can then bind different characteristics to groupings, that is, relate them to one another.

The establishment of relationships can be done in a dynamic way for (a) present characteristics among each other,

(b) current features with other feature spaces,

(c) current features with past values, other features and feature spaces, and (d) the overall condition with future but expected features. In particular, the inclusion of the past can take into account the causal character of the signals.

Dynamic data from technical systems can be fed into the neural network 1 as input information after pre-processing, if necessary for dimensional reduction. This can extract various features (eg Independent Composites). nents or nonparametric Merkraalvektoren analogous to self-organizing feature maps) in one or more input layers, some of which may also be equipped with a persisten¬ th activity (working memory function). Optimization of the neural network can be achieved by biologically motivated learning rules (eg Hebb rule or spike time dependent plasticity) with which cost functions can also be set up to evaluate how well a dynamic task is solved.

Quoted literature

[1] Sutton, R.S. and Barto, A.G. (1998): "Reinforcement

Learning, "Bradford Book, pp. 87-160.

[2] Szabo, M., Almeida, R., Deco, G. and Stetter, M. (2004): "(Cooperation and biased competition model can explain attendant filtering in the prefrontal cortex", Eur. J. Neurosci., Vol. 9, pp. 1669-1677.

[3] Brunei, N., and Wang, X. J. (2001): "Effects of neuromodulation in a cortical network model of object working memory dominated by recurrent inhibition", Comput. New rosci. 11: 63-85.

[4] Koch, C., and Segev, I (Eds.) (2001): "Methods in Neuronal Modeling: From Synapses to Networks," MIT Press, Cambridge, MA, chapters 1-5.

Claims

claims

1. A method for responding to contextual changes with a neural network (1),

- wherein an input layer (10) of the neural network (1) input information is supplied;

- wherein an output layer (30) of the neural network (1) output information is taken; - wherein in the neural network (1) a plurality of models are stored, each of which predefines a mapping of the input information to the output information;

- where only one model is active at a time;

- wherein the following steps are repeated: - the neural network (1) forms a supplied input information (2) with the active model on an output information (3), which is taken from the output layer (30); if the extracted output information (3) for the supplied input information (2) is incorrect in a current context, the neural network (1) receives an incorrect feedback (4), whereupon the neural network (1) activates another model ,

2. The method of claim 1, wherein taken from the

Output information (3) an action is derived and executed.

3. Method according to one of the preceding claims, - in which the neural network (1) contains exciting pulsed neurons,

in which the exciting pulsed neurons form model pools (41), each model being assigned at least one model pool (41), - in which the model pools (41) compete with one another, in which an active model pool (42) prevails in the competition ,

The method of claim 3, wherein the neural network (1) contains inhibitory pulsed neurons, - in which the inhibitory pulsed neurons form at least one inhibitory pool (50),

in which the inhibiting pool (50) exerts a global inhibition on the competing model pools (41),

in which the false feedback (4) activates the inhibiting pool (50),

in which the activated inhibiting pool (5) performs a complete inhibition of all model pools (41) in which the complete inhibition deactivates the active model pool (42), - in which a different model pool (41) is activated after complete inhibition ,

5. The method according to claim 4,

in which synapses of the exciting pulsed neurons of the active model pool (42) adapt,

in which recurrent weights of the active model pool (42) sink,

- in which the active model pool is subject to complete inhibition in the competition compared to the other model pools (41).

6. Method according to claim 5, in which the adaptation of the synapses is implemented as short-term synaptic depression (STD).

7. The method according to claim 3,

in which exciting pulsed neurons form control pools (21),

in which each control pool (21) in each case connects one of the input information with one of the output information,

- in which the rule pools (21) compete with each other, in which the active model pool (42) supports a selection (22) of rule pools (21) in which the supplied input information (2) activates a rule pool (23) from the selection (22) of rule pools (21), in which the activated rule pool (23) activates the output information (3) in which the activated output information (3) is taken.

8. The method of claim 1, wherein the neural network (1) contains interconnections, which are ausge¬ forms by Hebb'sches learning.

9. A neural network (1) for responding to context changes, having an input layer (10) to be fed with input information; with an intermediate layer (20), by means of which the input information can be imaged onto output information; - With an output layer (30) at which the Ausgabeinforma¬ tions are removed;

- With a model layer (40), with which the image in the intermediate layer (20) is controllable depending on a context.

10. neural network (1) according to claim 9,

in which the model layer (40) contains exciting pulsed neurons, in which the model layer (40) contains a plurality of model pools (41) which are composed of the exciting pulsed neurons,

in which, with the model pools (41), the image in the intermediate layer (20) can be controlled as a function of a context.

11. neural network (1) according to claim 10, in which the neural network (1) contains inhibiting pulsed neurons,

in which the neural network (1) contains an inhibiting pool (50) consisting of the inhibitory pulsed neurons,

in which the inhibiting pool (50) is connected to the model pools (41), in which the inhibiting pool (50) can be activated by an incorrect feedback (4) if a retrieved output information (3) is incorrect.

The neural network (1) of claim 9, wherein a first module comprises the model layer (40) and a second module comprises the other layers (10, 20, 30).

13. neural network (1) according to claim 10,

in which the input layer (10) has an input pool (11) for each input information, which consists of exciting pulsed neurons and can be activated by supplying the respective input information, in which the intermediate layer (20) has control pools (21) through which in each case an input information can be interconnected with an output information, and wherein the control pools (21) consist of exciting pulsed neurons, - in which the output layer (30) has for each output information an output pool (31) which consists of exciting pulsed neurons in which the model pools (41) are connected to the rule pools (21) such that, depending on the activation of the model pools (41), only a selection (22, 24) of rule pools (21) can be activated,

- In which by a supplied input information (2) a representative input pool (12) can be activated, which represents the input information (2) supplied, - in which by the representative input pool (12) a be¬ preferred rule pool (23) can be activated, which to the Selection (22) of rule pools (21), which are interconnected with an active model pool (42),

in which a representative output pool (32) can be activated by the preferred rule pool (23), the representative output pool (32) representing a removable output information (3).

14. Neural network according to claim 9,

- which contains one or more additional input layers (10), intermediate layers (20), output layers (30), model layers (40) or other layers, the additional layers filtering the input information, the output information or other information, rate, network or categorize or take over other functions.