Adaptive control device, neuron-like base element and method for operation of said device

FIELD: information technology.

SUBSTANCE: device includes a plurality of neuron-like base elements (BE) which are grouped into modular units such that outputs of like BE are orthogonal to inputs of other BE and are separated by a dielectric film with a semiconductor layer. The neuron-like BE has a multipoint input for reception and summation of electrical signals, a working output and a signal processing unit, having a threshold voltage former, a comparator and two normalised voltage formers.

EFFECT: simple design of the device, broader functional capabilities thereof and a more complex category of problems solved.

10 cl, 7 dwg

 

The invention relates to analog-to-digital control devices and can be used to create complex multivariate systems of automatic control of different objects and processes that allows an object to alter its response depending on the changing nature of the external influencing factors, pattern recognition systems, robotics, and simulation of the human brain.

According to the General theory of systems, control device regularly or as needed generates control actions on a process. Self-organizing control device based on the principles of optimization tasks for which the input information is processed in accordance with the set of rules in order to receive the output control signals, adequately solving the task.

The best example of self-organization is the human brain. For the realization of devices that operate by analogy with neural networks, created numerous technical solutions, mainly in the field of computer design.

The work of adaptive systems is organized using artificial neural network, in one way or another modeling the nervous system of living organisms and functioning on the basis of Naropa is one of the elements. Device for simulation of neural networks can be analog-to-digital, hardware or computer. The first group includes devices in which the models of individual neurons and neuronal connections are implemented in the form of electronic items that use multiple analog or digital basis. Common drawbacks of these devices are their hardware complexity and high cost, with the result that it is impossible to create models of large neural networks with characteristics of versatility, difficulties in the formation and growth of complex neural structures. The second group may include devices in which the model neurons or model neuronal connections are made different functional electronics, such as optoelectronic model. General shortcomings of these models are insufficient flexibility, the complexity of the restructuring of the network, the complexity of the changes in the functional characteristics of neurons and neuronal connections. Another disadvantage inherent in the majority of such devices is the poor state of production technology components. The third group include universal and specialized programmable computing device that are used to model neural networks. When it has reached what is flexibility changes in neuronal connections and characteristics of neurons, you can simulate large neural network. The main disadvantages of such devices are cumbersome software system and poor performance.

There are a large number of neural networks of different types (see Artificial Neural Networks: Concepts and Theory, IEEE Computer Society Press, 1992; J.Anderson, An introduction to neural networks, Chapter 2, MIT, 1999), the General lack of which is the difficulty of determining the weights of synaptic connections and their correction, the way jobs are either grossly simplifies the real processes, or so complicates the design that its actual implementation becomes hardly possible.

Known neuron-like element on which you can build a device for solving a class of problems the evaluation of the functioning of open complex systems, assessing the degree of optimality of different solutions (EN 2295769, publ. 2007). The system has limited functionality, in particular incapable of learning.

Known a number of technical devices for the management of complex technological processes (Neural network, EN 66831, 2007; Domain neural network, EN 7208, 2008; Modular computing system, EN 75247, 2008; Associative memory, RU 77483, 2008 and others), common faults which are limited functionality and inability to self-learning, because switching neuronal connections and information transfer between n what ionami is based on the deterministic tables, relations, and for a number of devices you want to use external memory.

Known dynamic synapses for signal processing in neural networks. The system includes a set of interconnected base of neurons and is designed to handle signals in a sequence of pulses. The essence of the device is the feedback between presinapticheskimi and postsynaptic neurons and the use of inhibitory neurons that continuously changes the procedure for signal processing the synapse, so that the second presentation of the same signal synapse takes into account presinapticescuu picture and gives a different picture on the way out. Dynamic change of the conductive network structure with feedback results in an increase in the computing power of the network at the neural level and increase the ability of the network to dynamic training (US 6643627, 2003).

In the known device the correction information signal is carried out on the local level presents - postsynap that is only effective for small volume. Increasing the number of layers up to 4 or more useful properties of the dynamic synapse dramatically reduced (see figa patent specification). Class solved similar tasks system is limited, the best results were obtained for filtering noise in multivariate signals and speech Analytics, for example if it detects the same words, spoken by different people with different accents.

Known nanotech neural network and the way of organizing its work in many aspects of modeling biological brain and to understand the processes of interaction of neurons (US 7426501. Nanotechnology neural network methods and systems, 2008). The device contains a set of artificial neurons with antrostomy, balanced and able to move in a special medium (dielectric solution). Neighboring adaptive synaptic elements separated by gaps across which can be applied electrical field, which establishes electrical connection. Communication can strengthen, weaken or even dissolve in the dielectric solution for forgetting incorrect responses to input signals that provides a fast neural network training. For this purpose, the gap between adjacent neurons can be equipped with the appropriate logic.

The known device solves some of these issues, and that the claimed technical solution, but in other ways, difficult for physical implementation and practical use.

Known technical nervous system EN 2128857, publ. 1999. The device contains the technical brain, the distributor output signals, the controller power supply receptor, the active device training, Retz who ptory output signals technical brain, receptors of external systems, power distribution and training. The neural network is constructed in the form of a matrix of adjustable resistor elements. At each step of the study define the situation in which the ratio of the deflection control signal to the allowable deviation of the signal in this situation was the greatest, and conduct external adjustment of the conductance matrix elements of the technical brain for this situation. The system has limited functionality, difficult to manufacture and contains a large number of external controls to correct reactions, which limits its ability to adapt.

Closest to the claimed technical solution for the intended purpose and essential features of a multi-layer perceptron device (US 5220641, Multi-layer perceptron circuit device, 1993). The device chosen for the prototype.

The architecture of an adaptive neural network a multilayer perceptron, like the invention of US 5220641, is used most often. Typically, each neuron builds a weighted sum of its inputs, with an amendment in the form of term and then skips the value of the activation transfer function, and thus the device generates an output signal value. Training recurrent neural network essentially consists in the study of surface errors and locating global is th minimum.

The device prototype contains many basic neural elements, the multiple ties between which is provided by the special topologies, in which items are grouped in modular units so that the outputs be one block are orthogonal or at an angle to the inputs b of another block that can be established between the output of one neuron-like element and the inputs of many other elements of the synaptic connection. Each of the synapses receiving signals from one of the lines of input signals, performs calculation of the weight signal with a prescribed value and then transmits the result of the calculation of the weight on one of the lines of output signal. Note, as in the prototype, the main element is the synapse, note that here and below the line input synapse prototype is a line of output neuron element according to the traditional view, and the output line for synapse prototype is a line input for traditional neuron and for said neural element. Each synapse contains the block count weight. Learning is built on the back-propagation algorithm errors. The same principles are embedded in multi-layer neural network on MOS inverters for US 5347613, publ. 1994, which announced a new five-step method of learning, but it is based on the well-known ways to the correction of the weights of links, and training is a modification of the method of teaching according to the invention prototype.

General shortcomings of systems with self-learning by the variation calculation and correction of the weights of links are unacceptably large processing times of the configuration input signals to obtain adequate control signals.

According to the prototype in order to train a neural network solution to any problem, it is necessary to repeatedly adjust the weight of each element in such a way as to decrease the error is the discrepancy between actual and desired output. At the same time, learning is growing much faster than the network size. The prototype disadvantages are the complexity of the device due to the presence of a large number of regulatory elements and limited functionality.

To eliminate deficiencies in the multifunctional adaptive control unit it is advisable to abandon the correction of the weights of links, including back-propagation algorithm errors. The invention used a different principle: instead of infinite adjustment of the weights of links continuously pave new connections between neurons, "forgetting" old and those clauses that take into account previous events and implement the "correct" response to the changing configuration of input signals. Such extensive PU is e-learning a more promising for adaptive systems and more correspond to the biological nature of intelligence. According to some estimates, the human brain has the potential in the hundreds of terabytes, which involved about half, and its content changes at an average rate of hundreds of kilobytes per second.

According to the invention the control unit can be implemented using a simple node - neuron-like basic element that can form an electrical connection with its neighbors and with some elementary properties of biological neurons: the ability to get excited, tired, repair and fed energy to transmit a signal without attenuation. To implement such a device, you must create a multilayer neural network with a very large memory, which is achieved by the special topology of the neural network, described below.

It should be noted that the Korean company Samsung has announced a substrate for semiconductor manufacturing, the thickness of which is only 0.08 mm Is 20% less than 0.1 mm substrate, which Samsung released earlier http://habrahabr.ru/blogs/hardware/14025/). The capacity of the human brain is of the order of 400-500·1012bytes. You can show that you have already mastered the technology of microelectronics in the same volume, which takes the human brain (approximately 1.5 DM3=1500 cm3), the structures of the claimed device is VA can be placed not less than 5.5 terabytes of information, or about 1% of the capacity of the brain.

An object of the invention is the creation of an adaptive control device that combines simplicity of design with functionality, creating the neural elements (artificial neurons), based on which you can build highly trainable neural network, a method of organizing the work of the adaptive control device with neural network.

The first task is solved in that in the known device prototype that contains many neuron-like base elements, each of which has input and output, able to receive and sum the electrical signals from sensors or from other EB and go to an excited state with the generation of an electrical signal, which can be transferred to many other EB or Executive mechanism, grouped in M>1 flat modular units for N>1 EB in each so that the outputs be one block are orthogonal or at an angle to the input end of another block, forming at least one grid with N nodes, use the new design multi-layer neural network.

For solving the problem of the outputs EB one block superimposed on the input EB of the other unit via the dielectric film with the semiconductor layer, each of these grid nodes when creating a difference sweaty the print materials, sufficient to dielectric breakdown, has the ability to form between the output of one EB and the input of another similar BAE resistive connection, conductivity which decreases with time, some arbitrary inputs be connected to the sensors and assigned to receive signals from external sources, outputs some arbitrary EB are connected with actuators controlling device and assigned to issue control signals, and a part of the above mentioned resistive ties laid in advance in the manufacture of the control device and the input signal forms at least one electrical circuit from input connected to the sensor, to the output connected to the actuating mechanism.

Another technical challenge is the creation of a neuron-like base element, suitable for creation of self-learning neural networks and physical implementation of the control devices for various destinations.

The technical problem is solved as follows.

As known analogues, neuron-like base element (BAE) adaptive control device includes a multi-point input for receiving and summing the electrical signals from similar basic elements (EB) or from external sources, work output and the signal processing unit, capable of generating electrical energy the economic signals in the excitation EB. In the base element according to the invention the signal processing unit contains the driver threshold voltage, connected to a source of adjustable reference voltage, a comparator, and two shaper normalized voltage (UNFPA), the first of which forms a positive rectangular pulses, the second generates the same magnitude, and duration of the negative rectangular pulses, a comparator connected to the multipoint entrance end and an exit of the driver threshold voltage, the input of the first UNFPA is connected to the output of the comparator, the output of the first UNFPA connected with working out EB and with the input of the shaper threshold voltage, the input of the second UNFPA is connected to the output of the comparator, the output of the second UNFPA connected with multipoint the entrance of the underlying element.

The design of the neural base of the element is new and constitutes a separate invention.

In one of the embodiments of the invention described neuron-like basic element grouped with many of the same basic elements, is made on a flat substrate, forming a modular unit basic elements, with two one-sided or two-sided pads with parallel inputs and outputs.

The third technical objective is the creation of the adaptive control device. On OPI the data following principles can work not only claimed the device, but a number of others, including a computer control system, and the claimed method is independent of the invention.

To solve this task in a well-known technique of operation of the control device, which includes receiving electrical signals from sensors, processing them by the neural network, including the establishment of time-varying electrical connections between a set of artificial neurons, each of which is able to receive and sum the electrical signals from nearby neurons or from external sensors, to generate a new signal and transmit it to other neurons or the actuator control device, use the following procedure. Part of these relationships set up to receive signals from sensors and use them for training the neural network, the electrical connections installed automatically during the work process control device, which received input to that neuron, the total signal with a magnitude that depends on the aggregate of all currently installed connections of the neural network is compared with the threshold potential of the excitation, the magnitude of which is set in dependence on the frequency of past excitation of the neuron, and if it exceeds excite the neuron, namely form at the input of the neuron normer the bathing negative pulse, at the same time or with small relative to the duration of the above-mentioned pulse delay form at the output of the neuron is identical positive pulse with establishing between neurons, excited state which at least partially overlaps in time mentioned electrical connection, and through the mentioned negative pulse partially neutralize the incoming electrical signals, providing competition of neurons for receiving the signal and preventing a situation in which the ratio of the reproduction signal exceeds unity, thanks to continuously modify a conductive structure of the neural network, including education between neurons are many feedback circuits, thus the output control signals of the control device or part of them are used for active influence on the parameters of the external environment and/or for correction of the conditions for receiving electrical signals from the sensors.

The described method can have a number of specific embodiments.

Established communication between neurons can wear chemical, capacitive, resistive nature. In the particular case of the complete method, it is preferable to perform the mentioned resistive connection that can be implemented, for example, breakdown of the dielectric separating the neurons, or even on the basis of a conventional relay e is the elements, if you let the size of the device. Thus, by analogy with biological synapses of neurons can implement aging resistive links, for example, the task of the special properties of the dielectric film with the semiconductor layer or by diffusion processes in the dielectric.

The delay of the front positive pulse at the output end toward the front of the negative pulse at the input EB is designed to increase the likelihood of the formation of relationships between events, reflecting the situation "for the event And is followed by event B"is more effective than education connection "before event B event A".

The training of the neural network can serve to develop a given response to a test configuration of external signals, for which the neural network pre-establish connections between neurons from the sensor to the control signal output.

To generate the desired response of the neural network on external training signals required some measure of behavior (minimizing energy expenditure, persistence dyslipidaemia thermal energy and the like), the deviation from which the correct training of the neural network. In the particular case of the method of this essential feature is to minimize the number of received external signals. Note that for the adaptive control device and methods for the and the organization of its work in the invention of any external signals are learning, as the learning process continues without a teacher in the process of operation of the device.

The learning process of the neural network controlling device can be greatly accelerated if a way to provide processes for modeling fatigue, recovery and self-excitation of the biological neuron. Fatigue provide a stepped increase of the threshold potential of the excitation in the case of frequent excitation of the neuron, when each time when the excitation potential receives additive pulse chain with a small resistance. This causes excitation of neighboring EB and passage of the signal for other circuits in the absence of a correct response after repeated passage of the signal through this EB. Restore the base element to provide a gradual decrease in threshold potential, until the reference, which provide high resistance circuit leakage.

The excitation implement due to the random fluctuation of the potential at the input EB. Such fluctuations, for example, induced EMF, are rare and small in size, therefore, the estimated reference voltage should be low enough.

The proposed method can be fully implemented in software computing. Control device with a relatively small number EB mo is et to be simulated on a personal computer.

Thus, the adaptive control unit contains the required number of sensors, the neural network of the neural basic elements and the required number of actuators. Technologically it is made of many identical modular units, each of which with an area of 3-4 cm2and the thickness of 0.1 mm can be grouped in the order of 104the neural basic elements, construction and operation of which is disclosed in detail below.

The technical result is to simplify the design of the adaptive control device, extending its functionality and the complexity of the class of tasks.

The invention is illustrated by the schema.

Figure 1 shows the schematic diagram of a control device.

Figure 2 schematically shows the structure of the neural base element.

Figure 3 shows a diagram of the modular basic body elements.

Figure 4 shows the mounting of the control unit of the modular units.

Figure 5 shows the encoding process, the total input.

Figure 6 shows the correction circuits stimulus - response training the neural network.

Figure 7 shows the diagram of a simple physical model, equipped with a control device.

Used in the description of the terminology often borrowed the from of neuroanatomy and neurophysiology. In this regard, all analogies describe the neural elements and neural networks with their biological prototypes should be considered with a high degree of conditionality.

Description control device.

Control device (1) contains the basic neural elements 3 with inputs 1 and 2 outputs. The inputs of some arbitrary EB is connected to sensors (not shown) and receive signals from external sources, outputs some arbitrary EB are connected to actuators (not shown), and generates control signals. The basic elements form a multilayer structure and positioned between inputs and outputs nearby BAE, separated by a dielectric film on the semiconductor layer, may establish electrical connection a, b, C, d, etc. conduct Communication signal in one direction, from the EB to the entrance close to him another EB. The nature of these relationships is shown on the external element 1. Relationships are based on the known technical solutions, when the pulse voltage causes electrical breakdown of the barrier film layer, which leads to the formation of the conductive connecting channels between bit buses. When this bit between tires is included p-n diode (see, for example, electronics: Science, Technology, Business 6/2001. Walentina. Exotic memory; Pat. US 585396, Three-dimensional read-only memory, 1998).

According to the invention each output EB can have a relationship with many other inputs BAE, each input EB can have a relationship with multiple outputs other BAE, the maximum number of connections depends on the specific topology of a multilayer neural network. In addition, the EB can be formed numerous feedback indicated on figure 1 by the numeral 4. Feedback can be as short as between elements k2 and j1, or long, as between the elements k1 and i2. A diagram is given as an example only three groups of basic elements i, j, k n elements each, and three feedback, but such groups and relations according to the invention can be set.

According to the invention the electrical connections installed by the manufacturer, for example by soldering, and is used to perform device some of the basic reactions in the training of the neural network with the teacher. These relationships are required in the device and when unsupervised learning, i.e. during operation.

For example, the input i1 is connected to the photocell, and the output of the k2 to the motor that moves the photocell. Predefined links form an electrical circuit including the motor when the feed sensor photosignal.

The control device operates as follows.

For an arbitrary control inputs of the device is the CTV receives signals from the sensors, transforming heterogeneous external signals-stimuli (optical, acoustic, temperature, pressure, electromagnetic, radiation, etc.) into electrical signals standard amplitude and duration. The design of the neural EB is such that the signal strength is automatically coded by the number and frequency of pulses, as shown in figure 5. In the moment of crossing the line of the threshold voltage with the signal line basic element is excited and generates a normalized output pulse when the magnitude of the signal, a series of pulses, as shown in the bottom line of the diagram in figure 5. Each such pulse may contribute to the total potential on the many inputs of those basic elements, to which you have the appropriate electrical connection. At each excitation EB threshold voltage abruptly increases. It is seen that the initial part of the diagram EB skips even small signals, but at the end of the chart BAE does not transmit signals of much greater magnitude. This means that the signal gets the opportunity to get into the inputs of several basic ingredients, stir them up and lay in a neural network with new connections. In addition, choosing a possible path through communication, radvilas and decreasing in magnitude, the signal can return to the previous entry of the base element contributing posilive in the total potential of the other applicants on the input signals.

The combination of input signals from sensors (irritant) passes through the current in the SU to this point in time the tree a set of electrical circuits (see figa), creating all new relationships between the basic elements. The neural network not only processes the signals, the signals themselves continuously modify a conductive structure of the neural network. Close in location, and simultaneously excited EB left on the path signal label memory - electrical connection, thereby storing path similar configuration signals, and further neural network performs a predictive function. Ultimately, the signals at the outputs of the control unit is connected to the actuators, and become the control signals.

Described below is the topology of the neural network provides that as the depth of the connections in the control unit are formed connection between the outputs of the basic elements that are logically located farther from the input sensors to the inputs of the elements that are logically located closer to the entrance. Thereby implement a local feedback, which participates in the formation of the reaction. When large amounts of SU's local feedback circuit provides the possibility of long-term circulation of the internal configuration signals on structure the frames neural networks with sequential excitation of many neurons in the chain, incorporating feedback. So the number of series-excited EB in chains repetition signal is obviously much longer than the direct path from the sensor to the actuator. On the one hand, the existence of such domestic relations allows to obtain more adequate response to the configuration input signals. On the other hand, the circulating signal implements the prediction of the sequence of events, because in this circuit have already passed similar configuration signals and for coming to the input EB of new signals already there is some additive to the total potential, you can say, "stand", which is easier to exceed the threshold voltage.

Consider the workings of the neural base element.

The scheme is one of the basic elements, the combination of which provides the adaptive control device, shown in figure 2. Basic element (EB) 3 contains the input 1, output 2 and unit conversion signal, comprising a comparator 6, a driver threshold voltage of 7 and two shaper normalized voltage 8 and UNFPA UNFPA 9. The input of the shaper threshold voltage 7 is connected to a source of reference voltage. The first input of the comparator 6 is connected to the input 1, the second input of the comparator 6 is connected to the output of the shaper threshold voltage 7, the output comparat the RA 6 is connected to the input of UNFPA 8 and to the input of UNFPA 9. For EB, you need external power supply: a source of adjustable reference voltage, a source of positive potential, a source of negative potential.

For the manufacture of a compact multi-layer neural network basic element can be structurally grouped with multiple N of the same basic elements, the combination of which forms a modular unit basic elements, is made on a flat substrate, as shown in figure 3. Inputs 1 and 2 outputs mentioned many basic elements are parallel to each other in one plane, forming two contact pads. As the inputs and outputs can be laid on both sides of the substrate.

Neuron-like base element (2) works as follows. On input 1 from neighboring EB or from external sensors can be served many signals Siwhich are summarized in accordance with currently numerous connections 5. The sum of the input signals is fed to the first input of the comparator 6, while the output of the shaper threshold voltage 7 to the second input of the comparator 6 is fed a certain threshold potential. The threshold potential is formed individually in each EB on the basis of the reference potential applied to the input of the shaper threshold voltage 7. At that moment, when the sum of the signals on the 1 exceeds the value of the threshold potential, BAE excited. Upon excitation of the base element, the comparator 6 supplies the command to the production of normalized stresses and UNFPA 8 generates output 2 normalized by the size and duration of positive potential, which determines the duration of the excitation EB is the signal output of the neural base element. Simultaneously UNFPA 9 generates input 1 negative potential, the magnitude and duration equal to the above-described positive potential.

A similar process occurs on the basic elements of the next segment (see figure 1). At some point, two fundamental elements in adjacent blocks at the same time be in the excited state, for example at the output 2 of item i1 generated voltage +U and the input 1 item j2 formed voltage-U. Since the output of the excited element i1 and the input of the excited element j2 is shared with the calculated dielectric breakdown potential of U<2U, breakdown occurs in the dielectric layer and between the base elements is formed resistor connection, indicated in the diagram by the point A. If only one excited EB, communication is impossible.

Output 2 EB 3 can establish electrical connection with a set of neighboring EB, as shown in the diagram of figure 2 as a set of signals Si.

The negative potential generated by the th at excitation EB through a low impedance resistor (not shown in figure 2), is fed to the input 1, dramatically reducing the total potential of the input signals Siduring initiation EB, thereby providing competition to the basic elements.

With frequent agitation EB on the cumulative capacitance of the input of the shaper threshold voltage 7 (not shown) from the output UNFPA 8 is supplied to a positive potential and the threshold voltage increases so that EB ceases to be excited, simulating fatigue synapse biological neuron. A negative pulse is no longer supplied to the input 1 EB and stops to help others total potential signals Si. This contributes to the excitation of other BAE, neighboring relations tired EB. During a long absence of excitations threshold potential flow through the high resistance resistor (not shown) and slowly decreases, tending to a reference potential U. So implementing a recovery EB.

According to the invention, the basic element of a neural network having a multi-point input, operates not by the signals of the external environment, and a superposition of signals received from different sources. Because the signals transmitted from the sensors to the actuators excited only through the basic elements of the signal received at the input may be sent in two cases: either it has a value greater than the threshold or entrance there were several si the channels and their total value is above the threshold, i.e. the condition of excitation of EB. Upon excitation generates a new signal that can be transmitted by the neural network. The signal is transmitted either through the newly established connection, or through communication, set earlier. If no ties, Horny EB does not pass the signal on. Since the signal is transmitted from layer to layer of the neural network is not in its original form, and is constantly generated, the signal is not attenuated.

The resistance of the newly formed connection R(t0)=Rmin increases over time R(t>t0)>R(t0). Figure 1 is a connection b and the input element j1 have different conductivity, so that two identical signal value from the outputs of the elements i1 and in make a different contribution to the total signal received at the input of the comparator element j1, eliminating synapses that calculates the weight of the signals in the device prototype. In the device according to the invention competition basic elements for arriving at their inputs the signals contribute to the current conduction ties one of the EB has the advantage and defines the path of further repetition signal as the most suitable for this configuration, the input signals.

Multilayer neural network is assembled from identical modular units MB by overlaying each other pads, orthogonal, as shown in figure 4, the sludge is at some angle to each other. In the latter case, the architectural options of the neural network can be expanded. The Assembly can be made in the form of spirals, wells, hexagonal honeycomb, combinations thereof and the like, the Number of layers is virtually unlimited and depends on the complexity of management tasks to be solved by the adaptive control unit. Plane all units remain parallel. Orthogonal planes blocks pave the power bus U, U+, U-, as shown in figure 3, supplied from the common to build the sources. Unlike traditional neural networks, in which layers of neurons follow each other, the invention provides contact outputs subsequent modular units to the inputs of the previous one, which ensures the formation of backward linkages. In the case of more complex tasks there is an opportunity to increase the number of layers and the number of modular units in each layer during operation of the system.

Electronic elements EB operate at low currents. If necessary, EB can always be equipped with a means of converting them into signals required power.

The above described organization of the formation of the potentials at the output and at the input EB allows to solve two main tasks:

a) to provide for the establishment of electrical connection between two simultaneously excited EB by track breakdown on the electrician, with specially selected specific electric strength and separating the output of one EB from the entrance located next EB. Unit condition for establishing a new connection between BA and j:

|U i-U j|>|U| |U|>|U|, |U|

b) when any configuration of input signals to provide excitation only in EB, which due to the increased weights of the signals formed the largest value of the sum of the potentials at the input. This is implemented competition basic elements, which are connected with this configuration of input signals: the excitation of the first EB prevents the excitation of other BAE, the inputs of which receive the signals involved in the excitation of the first vozrodivshego EB.

In this arrangement the SU is continuously formed numerous new contacts, but not self-excitation of the SU. The average number of EB, vozrodivshiisya in the next stage, not more than the number of excitatory basic elements of the previous stage, so the excitation SU's impossible. In other words, in the control device, the multiplication factor of the signal does not exceed one. When applying to the input of the first cascade of m signals at the output of the N-th cascade on average will be m1=m·KNsignals and for K≤1 always m1≤m.

Thus, when implementing the method according to the invention to the input of the CS signal from the sensors, transforming heterogeneous external signals-stimuli (optical, acoustic, temperature, pressure, electromagnetic, radiation, etc.) into electrical signals standard amplitude and duration, the total input signal EB is compared with a threshold potential of the excitation and signal strength-stimulus encode the number and frequency of normalized pulses.

The combination of input signals from sensors, through current in the SU at this point, an extensive collection of electrical circuits of the neural elements, including long-and short-circuit feedback neural network is processed. Ultimately, the signals in the form of multi-dimensional response, i.e. the modified aggregate normalized signals are transferred to the outputs of the neural networks that are connected to actuators controlling device. According to the method according to the invention the control signals of the device or part of them are used for active influence on the parameters of the external environment, or for correction of the conditions for receiving signals from the sensors. In particular, the control signals provide Executive mechanisms aimed at reducing the input signals, i.e. provide feedback, not necessarily electrical, between the inlet and outlet of the control device is istwa through the external environment.

The action of the Executive mechanisms for reducing the input signals can be quite varied, from displacement CU in space to eliminate sources of signal, which leads to the reconfiguration of external signals into the inputs of the SU.

Simple (base) reactions at the elementary signals of the sensors set gasket a number of relationships between the basic elements in the manufacture of the device. Complex reactions form the initial training of the neural network, then the device adapts itself.

Configuration of input signals in the initial state is a function of the environmental conditions, the number and location of sensors. Signal passing through a conductive medium SU leads not only to the visible reaction of the SU on the input signals, manifested in the action of the actuators, but also to the continuous change of the internal conductive patterns SU under the influence of the transmitted signals. This change occurs due to the formation of new, the most important first electrical connections between pairs of simultaneously excited neighbouring EB.

According to the invention over time, there is a reduction of scales formed between BAE links that simulates forgetting, aging, relationships of biological neurons. In General, the aging of relations in various m is the pH of a neural network can have different speeds. The incoming signals have the opportunity to initiate other EB and pass through the SU on new circuits. Changing patterns of relations in the SU is manifested in the change in the reaction CU on the same configuration of input signals, which indicates the retraining of the neural network.

Note that the SU does not itself generate signals. It only transmits signals through their patterns with multiplication factor at each stage K≤1. This leads to the fact that in the absence of external signals SU does not give any commands to their actuating mechanisms, as it freezes up until will not have any disturbing signals at its input.

This behavior of the device is set to basic ties laid in its manufacture: the signal of the sensor device responds quite clear command to the Executive mechanism which reduces the magnitude of the input signal. It is clear that the envisaged objectives of the CS function of the actuator must be consistent with the nature of the signal of the external environment. At the same aspiration control device to reduce the magnitude of the input signal based, and training devices. If as a result of their previous responses to input signals SU gets into position with no input (excitatory, annoying) signals, it ceases to operate. Note that this condition is SU steadily, similarly, the minimum of the potential energy of the ball on an uneven surface, in contrast to the state upon receipt of the signals of the sensors, forcing the device to use the actuators, and is consistent with the principle of least action for mechanical systems.

When you desire to decrease to at least one signal device may collide with the growth of another signal. In the end, it proceeds to the process of minimizing the sum of the two signals, in the General case - the sum of all input signals. In other words, the SU strives to achieve their local goals - to get away from signals-stimuli regardless of the degree of adequacy of SU reactions to external signals, starting from the base of the reactions defined in the manufacture of the SU.

This bears some analogy with unconditional reflexes biological organisms. All unconditioned reflexes biological organism implement his desire to escape from danger, often due to a change in its location. So, the simplest biological body tries to reduce irritation to zero, the more complex the organism can carry out other, more complex causal response.

Using the principle of minimization of input (annoying) signals of the sensors for practical purposes can be illustrated on the action control device for an Elevator in Nagatino house. Call the Elevator passenger should have an advantage over command of the control device, sending the Elevator at the specified time on the specified floor. If the procedure call Elevator for floors subject to certain rules, for example in the morning in t1 - floor n, t2 - floor n+2, t3 - floor n-2 and so on, it is possible to optimize the energy consumption, training the neural network to minimize the time empty (without passengers) travel of the Elevator between floors, this time empty trips you want to assign the input (annoying) alerts.

Very rare cases of spontaneous excitation EB-generation signal is not initiated input signals, can only for a short time to bring the SU from the state of rest (equilibrium), so these spontaneously generated pulses can have a significant impact on the behavior of the SU. Although it can lead (and lead) the change in the structure of the SU by establishing new connections between samovozvodyaschiesya and other vozrodivshiisya EB, in particular to encourage the device to exit the absence of signals from the sensors and to repeat the whole cycle response to changes in the configuration of the input signals.

Consider the process of primary education adaptive control device (see figure 1).

Upon completion of the TOS is to install the basic relationships between some of the EB, so obersoultzbach from sensors to actuators, the inclusion of which leads to a change of configuration of the input signals, which input signals are reduced (care, moving from excitatory signals without changing their source; "shielding" from exposure etc). The more of these pre-installed links embedded in the SU, the greater the number of basic reactions provided for in its design. After performing the factory defaults, the SU is ready to function in the simplest situations. Pre-installation reference links at the stage of manufacturing the SU determines the response of the SU on the "danger" and bears some analogy with the formation of unconditioned reflexes in the biological organism.

Factory setting basic hardware connections is carried out by the method as follows:

a) on the sensor signal. Ask what EB (or modular unit) signal and output a EB, a signal will appear. Ask what block out the signal of the first stage and which blocks the second stage is suitable signal block of the first level. Between BAE can form a relationship. We do not lay it manually, and create conditions for its formation. These conditions are as follows:

b) slowly lower the adjustable reference voltage on the block of the second level in order to any EB in this block, the 2nd level was itself sbordone and the EB first level formed a relationship with BAE from the block of the second level during self-excitation. Note that before the formation of this basic communication all be from the block 2 level had no ties and could not be excited either through one of its inputs;

C) the fact of self-excitation of some EB 2nd level can be set to increase the current in the circuit block 2. The fact of communication between excited EB 1st level and one of samovozvodyaschihsya EB 2nd level can be set, for example, in radio emission from the contact region at the time of the breakdown. Expedient for the ends of the conductors of the inputs / outputs of the block to connect and be conclude to control, and after the establishment of the basic relations of installation conductors delete;

g) return the reference voltage block 2 level to normal and slowly lower the reference voltage for the unit 3 level, achieving some self-excitation EB in its composition, and, continuing to signal to the input b of the unit 1 level, conduct a procedure similar to points b) and C), for establishing basic connectivity for conducting a signal between the output excited by EB 2-level and entry of samovozvodyaschihsya EB 3-level;

d) repeat this procedure gasket base connections from the sensor to establish a connection with the actuator needs of the actuator, choosing the right audio signal path by selecting the block at each stage of the laying of the next link adjustable chain is. So upon completion of the connection strip and the return reference voltages for all blocks in the normal signal is applied to the input of the sensor is set to manual, the response at the output of the SU in the form of the action of the actuator, reducing the input;

(e) the procedure for establishing the basic relations from the originating sensor to the reaction, reducing its signals, repeated as many times as needed to create in SU's necessary set of initial reactions, providing the possibility of further training and functioning of the SU.

Pre-install the base relations a necessary but not sufficient for multifunctional SU.

Adaptable device must learn from their own experience, and at the initial stage through initial training, receiving input signals from the instructor. As soon as the device will create a model of their reactions, it will be able to recognize analogies, based on past experience, to forecast future events and to offer the solution of new tasks. There is a need to create high-speed learning control device.

After the formation of the SU acceptable number of basic reactions conduct initial training, an expansion set adequate (desired) reactions on the configuration of the input signals of the sensors.

Initially the training is based on the artificial suppression and natural forgetting the electrical connections, giving the wrong signals, and new relationships between neural elements, leading to more adequate reaction control device on the incoming sensory signals. The neural network is trained, replacing ties, giving an incorrect response, new relationships, giving a more correct response. Thus, in the control unit set the patterns for the development of good reactions, after which the device adapts to the changing external signals independently.

The training of the neural network is carried out with the aid of a feed device of corrective signals set by the teacher (see figure 5 and 6). The reaction of SU's on an external signal is to include the element that received the control signals after the passage of the signals of sensors installed in advance and developed new neural network circuits, from the sensor signal to the response, as shown in figa. Here BR - base reaction (similar to conditional reflex), HP - incorrect response, PR is the right reaction. If the result of the reaction of SU's signal sample will decrease, this is the correct response (PR). Otherwise, the observed incorrect response (HP). If the signal strength of the sensor increases, the increases and the response to the received output a series of pulses as due to the more frequent were the Denia signal on the same circuit, and due to the signal propagation in a parallel circuit that is incorporated in the design of the neural element.

Initial training is conducted with the teacher according to the following scheme.

The signal sample. If the reaction of SU's presented on the signal sample is correct, then in the memory of the SU remains just performed a chain of links, which will be played on subsequent presentations of this signal. If the same reaction on the SU sample wrong (HP), the teacher (instructor) includes the correction signal is a strong signal to the stimulus (bottom chart on figb), causing the passage of a large number of pulses BR on circuits basic reactions (thick line figa). As a result of gradual growth (branching) relations in the SU region appear with electrical circuits, which are as signals an incorrect response on the SU sample, and the signal response of the SU on the adjustment signal. This enables supply adjustment signal to cause competition EB and their ustawianie and to block the passage of the signal through the chain of an incorrect response.

In the result of Ustawienia and blocking of the wrong chain reaction formed a new chain reaction on the SU signal sample. If this new chain reactions are also not satisfied with the teacher, he continues to serve on the entrance SU strong corrective signal. When the reaction on the SU signal sample becomes correct (PR), i.e. the reaction of SU's lead to off or decrease of the signal sample, the teacher turns off the corrective signal. The training ends. Subsequent presentation of the same signal sample SU will act on the last option, because relationships that have implemented it have the greatest weight.

Prolonged prove futile feed adjustment signal chain BR proliferate, raising a large number of nearby base elements. Finally they reach the region of passage of the chain on HP signal sample (the upper line of the signal chain on figa) and bring the result is a block circuit incorrect response.

So training is simple reactions using strong corrective signal. In this case, the efficiency, the effectiveness of the corrective signal is quite simple, however the configuration of the signal sample should be chosen such that the achievement of a correct response was fast enough.

Learning a complex reaction is carried out stepwise in several stages. At several stages on the basis of a strong correction signal is formed by the reaction of SU's on a much smaller adjustment signal. Multiple training is to include more and more strong corrective signals in the absence of a correct response on the SU presenting weak correctitude the signal. To do this, before applying a strong corrective signal is activated weaker corrective signal to create an electrical circuit forecast the possibility of applying a strong corrective signal, if the response does not change.

The operation of the adaptive control device in the simplest implementation is illustrated by example (7).

The truck is equipped with control device (CU), power supply, sound the buzzer. There is a power movement database, under which the carriage can move left and right between two walls with a spacing of 2 meters. There is a source of L strong light from above, depicting any additional external signal. A strong light source can be activated by the experimenter on and off automatically according to the audio signal. Using siren SU may file such a sound signal. Block movement of the trolley (DB) models of external influence, in the absence of which the cart is at rest (neural network falls asleep). At random points in time, the database includes the movement of the trolley at a random distance in the range 0,4÷0,6 m in a random direction (left or right).

On the trolley fixed sensor block BS with sensors:

the sensor S1 contact with the left wall;

the sensor S2 contact with the right wall;

the sensor C3 dangerous rapprochement with the left wall on the Sam. 0,2m (lamp LL);

the sensor C4 dangerous approach from the right wall on the Sam. 0.2 m (lamp LR);

sensor C5 strong light from above (lamp L).

Upon receipt of the signals from the sensors SU can apply for the DB truck three signals: order on motion to 0.3 m to the right, order on motion by 0.3 m to the left, the order to sound the alarm. Signals from a control device, have a higher priority than random signals on the move, supplied by the block motion database.

SU is mounted, for example, from 10 EB 20 inputs each to physically provide the possibility of formation of links between any EB. Before testing because SU's factory setting in the above algorithm, namely laid-chain base reactions.

The signal of the sensor S1 includes BAE contact with the left wall, which includes BAE enable movement to the right, which sends a signal to the database the order right. This chain of relationships ensures the functioning of the basic reaction in a collision with the left wall to enable the movement to the right".

The sensor signal S2 includes BAE contact with the right wall, which includes BAE enable movement to the left, which sends a signal to the database (the order "left"). This chain of relationships enables the operation of the mortgaged in advance the basic reaction in a collision on the right wall to enable movement to the left".

To enter is in the SU signals from sensors:

connect the sensor C3 with the corresponding EB;

connect the sensor C4 with the corresponding EB;

connect the sensor C5 with BAE sensor to strong light.

First, the basic elements are not involved in the chain reaction of SU's signals.

a) Adaptation (learning) neural network.

Disable the transmission of the signals for all sensors, except for sensor contact with the walls. The behavior of the SU will be very simple: in the course of random walks on random signals with DB in contact with one of the walls of the cart will move in the opposite side walls.

Now turn on signals for sensors proximity with the wall of 0.2 m In the case of the first response of the sensor closer to the right wall, but without contact with her in the SU nothing will happen. But if it works the sensor closer to the right wall and in the same maneuver to the right wall will be achieved, in addition to the installed during the Assembly of the order "left" will be simultaneously excited by two EB and SU installed electrical communication with the convergence on the right wall you should contact right. So next time when the sensor is activated rapprochement with the right wall through the established in the previous episode mentioned the connection with the output EB dangerous proximity to the input EB of contact with the right wall will receive an electrical signal and its output will display the command for the DB - "left", because at the same time, what about the excitement of convergence will be excited and be in contact with the wall. So later "dangerous" contact with the right wall will never be the be-comes the training of the SU. Similarly, automatically establishing a communication when interacting with the left wall and will have acquired a reaction to convergence.

b) Training a neural network.

Now if we materially change the terms of the interaction of CU with the external environment is to swap the sensors closer to the walls, when approaching from the right wall, when triggered, the former sensor closer to the left wall, existing for a rapprochement with the left wall should contact left the cart will include the movement to the right and face the right wall. But when reaching the right wall work installed at manufacturing base "absolute link" and the cart will get the order right. In the same episode will take place simultaneous excitation of EB who had signal on the approximation on the left wall and BAE contact with the right wall. Between them, a new connection, stronger than had outgrown the relationship between EB rapprochement with the left wall and BAE contact with the left wall. So subsequent excitations of the former BAE rapprochement with the left wall it will be faster to initiate EB contact with the right wall and turn signal-to-order "to the left". Horny BAE contact with the right wall will be immediately't impact napriazheniia output EB of the former closer to the left wall, so BAE contact with the left wall will not be actuated and will not give the order right. Exactly the same will happen retraining for SU BAE former closer to the right wall.

in the Search of solutions in an unknown situation.

Include bright light, resulting in excited BAE bright light. At this point, can work any EB movements, such as BAE convergence. Truck arisen under the influence of erroneous communication between BAE bright light and EB convergence will start to move, because she has no other reactions. As mentioned above, the steady state of the system is the lack of signals. In fact, the controller iterates through the possible reactions, trying to turn off the overhead light, i.e. to reduce the number of signals. For some time, to no avail, the light is off will not. The trolley reaches the wall (contact) and will roll in the opposite direction. And will ride some time, but BAE contact with the wall will start to get tired, their thresholds will be increased, so that will be a possibility of self-excitation of the base element turn on the siren. Turns on the siren, strong light will automatically turn off, what is the problem. So the result of random search SU will be found the exit almost impasse. The established relationship between BAE bright light and EB siren with the stored (within the aging communication). So in the next episode irritation bright light will trigger the siren and lamp L will be immediately turned off.

The described example illustrates the essential features that ensure fatigue of the base element, the base excitation element and the aging of its relationships. Indeed, each EB has a current threshold voltage of the comparator to trigger based on a preset reference voltage. Frequent triggering EB leads to an increase in internal threshold voltage and to the loss of the ability to excitation (case with multiple movement left-right), long stay BAE in the unexcited condition leads to excitation by reducing the threshold voltage below a certain limit (the case of siren), and the decrease in conductivity of old ties changes the path of movement of the signal (siren ahead of the traffic signal). Technical parameters of the base element are determined by calculation. Similarly, it is easy determination of the duration of the pulses generated by the BAE, the pulse duration of the movement, duration of response outputs of various sensors, which are calculated based on the physical conditions of the problem.

Sources of information

1. Artificial Neural Networks: Concepts and Theory, IEEE Computer Society Press, 1992; J.A.Anderson, An introduction to neural networks, Chapter 2, MIT, 1999.

2. Patents EN 2128857, 1999; EN 2295769, 2007.

3. Patents EN 6831, 2007; EN 7208, 2008; RU 75247, 2008; RU 77483, 2008.

4. Pat. US 5347613, 1994; US 6643627, 2003; US 7426501, 2008.

5. Pat. US 5220641, 1993 (prototype).

6. http://habrahabr.ru/blogs/hardware/14025/.

7. Walentina. Electronics: Science, Technology, Business, 6/2001; Pat. US 5835396, 1998.

1. Adaptive control device that contains many of the neural basic elements (be), each of which has input and output, able to receive and sum the electrical signals from sensors or from other EB and go to an excited state with the generation of an electrical signal, which can be transferred to many other EB or actuator devices grouped in M>1 flat modular units for N>1 EB in each so that the outputs be one block are orthogonal or at an angle to the input end of another block, forming at least one a grid with N2nodes, characterized in that the outputs be one block superimposed on the input EB of the other unit via the dielectric film with the semiconductor layer, each of these grid nodes when creating a potential difference sufficient to dielectric breakdown, has the ability to form between the output of one EB and the input of another similar BAE resistive connection, conductivity which decreases with time, some arbitrary inputs be connected to the sensors and assigned to receive signals from the NR is snih sources, outputs some arbitrary EB are connected with actuators controlling device and assigned to issue control signals, and a part of the above mentioned resistive ties laid in advance in the manufacture of the control device and the input signal forms at least one electrical circuit from input connected to the sensor, to the output connected to the actuating mechanism.

2. Neuron-like basic element of the adaptive control device containing a multi-point input for receiving and summing the electrical signals from similar basic elements (EB) or from external sources, work output and the signal processing unit, capable of generating electrical signals upon excitation EB, characterized in that the signal processing unit contains the driver threshold voltage, connected to a source of adjustable reference voltage, a comparator, and two shaper normalized voltage (UNFPA), the first of which forms a positive rectangular pulses, the second generates the same magnitude, and duration of the negative rectangular pulses, a comparator connected to the multipoint input EB and with the release of the driver threshold voltage, the input of the first UNFPA is connected to the output of the comparator, the output of the first UNFPA connected tolabor output EB and with the input of the shaper threshold voltage, the input of the second UNFPA is connected to the output of the comparator, the output of the second UNFPA connected with multipoint entrance of the underlying element.

3. Neuron-like base element according to claim 2, characterized in that it is grouped with many of the same basic elements, is made on a flat substrate, forming a modular unit basic elements, with two one-sided or two-sided pads with parallel inputs and outputs.

4. The technique of operation of the adaptive control device that includes receiving electrical signals from sensors, processing them by the neural network, including the establishment of time-varying electrical connections between a set of artificial neurons, each of which is able to receive and sum the electrical signals from nearby neurons or from external sensors, to generate a new signal and transmit it to other neurons or the actuator control device, characterized in that a part of these relationships set up to receive signals from sensors and use them for training the neural network, the electrical connections installed automatically during the work process control device, which arrived at the entrance of this neuron summed signal is compared with a threshold potential excitation, the value of which is tavat in the frequency dependence of the previous excitation of a neuron, and if it is exceeded excite the neuron, namely form at the input of the neuron normalized negative pulse, simultaneously or with small relative to the duration of the above-mentioned pulse delay form at the output of the neuron is identical positive pulse with establishing between neurons, excited state which at least partially overlaps in time mentioned electrical connection, and through the mentioned negative pulse partially neutralized supplied to the input of neuron electrical signals, providing competition of neurons for receiving the signal and preventing a situation in which the ratio of the reproduction signal exceeds unity, thanks to continuously modify a conductive structure of the neural network, including education between neurons many feedback circuits, thus the output control signals of the control device or part of them are used for active influence on the parameters of the external environment and/or for correction of the conditions for receiving electrical signals from the sensors.

5. The method according to claim 4, characterized in that the said electrical connection install predominantly resistive type.

6. The method according to claim 5, characterized in that it ensures reduction of the electric conductivity of us who set out resistive connections.

7. The method according to claim 4, characterized in that the said primary training of the neural network form the set of required output control signals for a given configuration of the signals of sensors.

8. The method according to claim 4, characterized in that the said primary training of the neural network train control device to minimize the number of input signals.

9. The method according to claim 4, characterized in that the said threshold potential excitation set equal to some reference voltage, ensure the growth threshold potential when the number of excitations per unit of time and a gradual decrease up to the reference voltage while reducing the number of excitations per unit time, or in their absence.

10. The method according to claim 4, characterized in that provide excitation of the base element with a low level mentioned threshold potential due to the random fluctuation of the potential at its input, with the formation of new relationships between the basic elements in the absence of signals in the external environment.



 

Same patents:

Neuroprocessor // 2473126

FIELD: information technology.

SUBSTANCE: invention can be used in designing computing means for systems for controlling highly manoeuvrable aviation and space-rocket objects where there is need for fast computation of functions, for example, trigonometric functions, used in matrix transformations when solving tasks of forming an inertial coordinate system based on information from angular velocity sensors, and when solving tasks for maintaining operating capacity of computers during changes in parameters of LSI elements due to the effect of natural or artificial ionising radiations. The device has a unit for communication with an on-board computer or which controls a higher level subsystem, a control device, a buffer register, a memory device, multipliers, an adder and an output register.

EFFECT: fewer faults in microchips.

5 cl, 5 dwg

Neuroprocessor // 2473126

FIELD: information technology.

SUBSTANCE: invention can be used in designing computing means for systems for controlling highly manoeuvrable aviation and space-rocket objects where there is need for fast computation of functions, for example, trigonometric functions, used in matrix transformations when solving tasks of forming an inertial coordinate system based on information from angular velocity sensors, and when solving tasks for maintaining operating capacity of computers during changes in parameters of LSI elements due to the effect of natural or artificial ionising radiations. The device has a unit for communication with an on-board computer or which controls a higher level subsystem, a control device, a buffer register, a memory device, multipliers, an adder and an output register.

EFFECT: fewer faults in microchips.

5 cl, 5 dwg

FIELD: information technology.

SUBSTANCE: device includes a plurality of neuron-like base elements (BE) which are grouped into modular units such that outputs of like BE are orthogonal to inputs of other BE and are separated by a dielectric film with a semiconductor layer. The neuron-like BE has a multipoint input for reception and summation of electrical signals, a working output and a signal processing unit, having a threshold voltage former, a comparator and two normalised voltage formers.

EFFECT: simple design of the device, broader functional capabilities thereof and a more complex category of problems solved.

10 cl, 7 dwg

Neurocomputer // 2553098

FIELD: physics, computer engineering.

SUBSTANCE: invention relates to computer engineering and can be used in designing strapdown inertial reference systems which are part of automatic control systems for highly-manoeuvrable ships, aircraft, space rockets and spacecraft in particular, as well as mobile robotic systems which are characterised by operability in extreme conditions. The device includes a microprogramme control unit, two matrix neuroprocessor units, an operational device, matrix memory, a secondary power source, a communication unit, authorised access memory and an environment sensor.

EFFECT: faster matrix computations.

23 cl, 20 dwg

FIELD: medicine.

SUBSTANCE: individual's clinical and medical history data and laboratory findings are analysed including: sex, age, cancer diagnosis time, type of complication, degree of cancer spread according to TNM classification, stage of cancer, concomitant pathology, previous treatment, complete blood cell counts, clinical urine analysis results, biochemical blood analysis results, ultrasonic data, colonofibroscopy findings, type of surgery. The analysis is performed by means of an artificial neural network, which represents a feedforward network with three neuron layers: 4 neurons in the first layer, 2 neurons in the second layer, and one neuron in the third layer. The neurons in the layers are coupled each to each. Each neuron of the first layer has 28 synapses; neurons of the second layer have 5 synapses; the neuron of the output layer has 3 synapses. As an activation function, the neurons use a logistic function. The neural network divides an output signal space into two groups and relates the value below 0.5 to the unfavourable clinical course of the early postoperative period, and the value exceeding 0.5 - to the group with the favourable clinical course.

EFFECT: presented method is simple, cost-effective in comparison with most of approaches existing; it is non-invasive and based on analysing the data obtained as a part of routine preoperative examination and requires no additional diagnostic techniques that in turn is expected to increase reliability, objectification and correct the treatment as may be necessary.

2 ex, 2 dwg

Artificial neuron // 2579958

FIELD: information technology.

SUBSTANCE: invention can be used in the construction of information processing systems in the neural network basis, including pattern recognition (classification). Apparatus comprises an input nodes multiplying the input value by a weighting factor, adder block activation function, each node multiplying the input value by a weighting factor includes a multiplier input by a weighting coefficient, and forming unit weight, comprising a comparator input, and the threshold values of the cell nonvolatile memory with data written therein thresholds switch memory permanent storage device to record therein the weights.

EFFECT: technical result is the ability to ensure multi-parameter classification.

1 cl, 2 dwg

FIELD: computer engineering.

SUBSTANCE: invention relates to bionics and computer engineering and can be used as structural and functional element of artificial neural networks for simulating biological neural networks, as well as for designing of parallel neural computers and other computer systems for various applications, including tasks of image recognition, data classification, image processing, mathematical operations and creation of artificial intelligence. Device comprises information and modulating inputs, information outputs, computing core consisting of processor device, application non-volatile memory, random access memory and non-volatile data memory interacting through system bus.

EFFECT: technical result is providing most complete emulation of biological neuron functioning, as well as application versatility and multiplicity, which provides significant saving of hardware resources.

5 cl, 3 dwg

Up!