WO2008072994A1

WO2008072994A1 - Neuron element

Info

Publication number: WO2008072994A1
Application number: PCT/RU2006/000668
Authority: WO
Inventors: Evgeny Nikolaevich Zakharov
Original assignee: Evgeny Nikolaevich Zakharov
Priority date: 2006-12-13
Filing date: 2006-12-13
Publication date: 2008-06-19

Abstract

The invention relates to cybernetics and can be used in the form of a cell for neuron networks used for solving the assessment problems of the operation of open complex systems, for assessing the optimality of different solutions, by building the hierarchical and recurrent model of a system to be investigated, taking into account the different initial and operational states of the elements thereof and by taking into account, during the modeling of the separate elements of a system by means of neuron elements, the level of the self-sufficiency thereof and the susceptibility to external signal action. The inventive neuron element makes it possible to take into account the setting error of the parameters thereof and of the parameters of input signals and to ensure the specified accuracy of the neuron network self-learning. The neuron networks based on the inventive neuron elements make it possible to perform different variants of the assessment of the operation of open complex systems and the optimality degree of different solutions and to develop different variants of adaptive information management systems.

Description

Neural element

Technical field

The invention relates to cybernetics and can be used as a cell of neural networks when creating devices for evaluating the functioning of various systems, including open complex systems, evaluating the degree of optimality of various solutions in adaptive information-control systems.

State of the art

Known neural-like element (NPE), containing a unit for summing the input signals, a unit for setting input weights, an adder of input signals, an input signal conversion unit, an output (S. Osovsky. Neural networks for processing information. -M.: Finance and statistics, 2002). The specified device allows you to create neural networks (HC) to solve certain problems associated with the generalization of the information obtained on the basis of training HC to determine the relationship between known input values and unknown outputs. To do this, the user must prepare a set of training data representing are examples of input actions and their corresponding outputs. As is known (Neural Networks. STATISTICA Natural Networks: Translated from English. -M .: Hotline-Telecom. 2001), the number of observations to be included in the learning process should be an order of magnitude greater than the number of connections in the network, while increasing the number of input signals the number of required observations is growing nonlinearly.

The disadvantage of this NPE is that for the class of problems of evaluating the functioning of open complex systems (OCC), evaluating the degree of optimality of various solutions, when the elements of the system identified as a result of the OCC formalization require (with the classical approach) separate modeling, including using HC, when the number of connections between elements is large, and there are no options for classical HC training, neural networks based on existing neuron circuits are unsuitable.

Disclosure of invention

The technical result of the invention is the possibility of creating a NPE, allowing to implement on its basis a neural network for solving a class of problems of evaluating the functioning of open complex systems (OCC) and assessing the degree of optimality of various solutions by providing the possibility of constructing a model of the studied system, both hierarchical and recurrent, taking into account various the initial and working state of its elements and their functioning options, taking into account when modeling neural-like elements of individual elements of the ka system to the level of their self-sufficiency, and exposure to external signals. At the same time, the proposed NPE makes it possible to take into account the errors in setting its parameters and the parameters of the input signals, as well as to ensure the specified accuracy of the self-learning of the neural network. The task is achieved in that the neural-like element contains an input unit including a first random number generator, the number of outputs of which is equal to the number of inputs of the input unit, the signal from each output of the first random number generator determines the change in the value of the corresponding input signal according to the law and within the limits determined by the type and the error of its assignment, the unit for setting and normalizing weighting coefficients, the number of weight elements of which is equal to the number of inputs of the input unit, the weight setting in the weight elements is carried out I, using the second random number generator according to the law and within the limits determined by the type and error of its assignment, normalize each weight coefficient by dividing each weight by the sum of all the weights of the weight elements, and each normalized weight coefficient sets the corresponding synaptic connection of the outputs of the input block and the inputs of the block calculation of signal parameters, the outputs of which are connected to the inputs of the adder of signal parameters, the output of which is connected to the first input of the calculation unit of the input part of the state of the NP E, the second input of which is connected to the limiter of the fraction of signals, a single-channel random number generator that determines the value of the signal of the limiter of the fraction of signals according to the law and within the limits determined by the type and error of setting this value, the unit for setting the internal state of the NPE, the second single-channel random number generator, which determines the value the signal of the unit for setting the internal state of the NPE according to the law and within the limits determined by the type and error of its task, the output of which is connected to the input of the unit for calculating the internal part of the thaw of the NPE, which multiplies the signal from the output of the unit for setting the internal state of the NPE by the difference between the unit and the signal value of the limiter of the fraction of signals, the output of the unit for calculating the internal part of the state of the NPE is connected to the first input of the unit for calculating the state of the NPE, combining the signal from the output of the unit for calculating the input part of the state of the NPE with the signal coming from the output of the unit for calculating the internal part of the state of the NPE, the output of the unit for calculating the state of the NPE is connected to the output unit, with the first input of the analyzer, the magnitude of the change in the state of the NPE directly, and with its second input through a memory unit that stores information about the previous value of calculating the state of the NPE, a unit for setting the accuracy of training the neural network, the output of which is connected to the third input of the analyzer of the magnitude of the change in the state of the NPE, to which compares the accuracy of self-learning with a value equal to the difference between the previous value of calculating the state of NPE and the current one, and if the specified condition is met, it sends a signal to the control unit to complete the process of self-training of the neural network, a control unit that controls the random number generators and the passage of pulses in the neural network .

The task is also achieved by the fact that a switch can be inserted between the output of the unit for calculating the input part of the NPE state and the input of the unit for calculating the state of NPE, the second output of the switch is connected to the input of the input block of deterministic dependencies, the output of which is connected to the first output of the switch, the signal is determined by the known functional dependence of the signal, which determines the input part of the state of the NPE, from the input signals.

Brief Description of the Drawings

The drawing shows a block diagram of the NPE.

The best embodiment of the invention NPE consists of an input block I ₅ containing the first random number generator 2, operating by command of the control unit 3, a unit for setting and normalizing weighting coefficients 4, also containing a second random number generator 5, operating by command of the control unit 3, and a block for calculating input signal parameters 6 , adder parameters of the input signals 7, the limiter of the proportion of signals 8 with its single-channel random number generator 9, the unit for calculating the input part of the state of NPE 10, the unit for setting the internal state of NPE H _{5 of the} second single-channel random generator 12, a unit for calculating the internal part of the state of the NPE 13, a unit for calculating the state of the NPE 14, a memory unit 15, an analyzer for changing the state of the NPE 16, a unit for setting the accuracy of self-learning of the neural network 17, a block of deterministic dependencies 18, a switch 19, an output block 20, and clock generator 21.

The essence of the invention lies in the fact that the neural-like element contains an input unit including a first random number generator, the number of outputs of which is equal to the number of inputs of the input unit, the signal from each output of the first random number generator determines the change in the value of the corresponding input signal according to the law and within the limits determined by the form and the error of its assignment, the unit for setting and normalizing weighting coefficients, the number of weight elements of which is equal to the number of inputs of the input unit, setting the weight in the weight elements is carried out are carried out by means of a second random number generator in accordance with the law and within the limits determined by the type and error of its assignment, each weighting coefficient is normalized by dividing each weight by the sum of all weights of the weight elements, and each normalized weighting factor sets the corresponding synaptic connection of the outputs of the input block and the inputs of the block calculation of signal parameters, the outputs of which are connected to the inputs of the adder parameters signals, the output of which is connected to the first input of the unit for calculating the input part of the NPE state, the second input of which is connected to the limiter of the signal fraction, a single-channel random number generator that determines the value of the signal limiter of the signal fraction according to the law and within the limits determined by the type and error of setting this value, block setting the internal state of the NPE, the second single-channel random number generator that determines the signal value of the unit for setting the internal state of the neural-like element according to the law and within the limits of consumed by the type and accuracy of its task, the output of which is connected to the input of the unit for calculating the internal part of the NPE state, multiplying the signal from the output of the unit for setting the internal state of the NPE by the difference between the unit and the signal value of the limiter of the fraction of signals, the output of the unit for calculating the internal part of the state of the NPE is connected to the first the input of the unit for calculating the state of the NPE, carrying out the addition of the signal from the output of the unit for calculating the input part of the state of the NPE with the signal coming from the output of the unit for calculating the inner part with the state of the NPE, the output of the unit for calculating the state of the NPE is connected to the output unit, directly with the first input of the analyzer of the magnitude of the change in the status of the NPE, and with its second input through the memory block, the neural network training accuracy setting unit, the output of which is connected to the third input of the analyzer of the magnitude of the change in the NPE, which compares the accuracy of self-learning with a value equal to the difference between the previous value of calculating the state of the NPE and the current one, a control unit that controls the generators s numbers and clock generator.

In addition, a switch is installed, installed between the output of the unit for calculating the input part of the NPE state and the input of the unit for calculating the state of NPE, the second output of the switch is connected to the input of the input unit of deterministic dependencies, the output of which is connected to the first output a switch, the magnitude of the signal generated by it being determined by the known functional dependence of the signal determining the input part of the NPE state on the input signals.

The first random number generator (RNG1) 2, operating on command of the control unit (BU) 3 and issuing the values Kgcclj, multiplying by which the amplitudes of the input analog signals allows you to vary their amplitudes according to a given law (for example, normal, uniform, etc.) and specified boundaries in the range from 0 to 1.

The distribution law and the boundaries of the amplitudes, defined by random number generators, are determined by the type and errors of the values. For example, the law of changing the value and / or weight of an input signal, i.e. the type of signal assignment and / or its weight is determined. Or, for example, when obtaining a numerical determinate estimate of certain quantities (input signals, their weights, share limiter, internal state), this is determined by the measurement error, calculation, etc., and the range can be determined by the possible minimum and maximum values of these quantities . Upon receipt of a numerical expert assessment, the distribution law can be determined, for example, by the law of distribution of opinions of experts, etc., and the amplitude - by the minimum and maximum estimates obtained during the examination.

The second random number generator (GCCH2) 5 also works on command BU 3 and outputs the multiplication value Kgcch2 _s for which values of balance weight elements can vary their amplitude by a given law and within certain limits in the range from 0 to 1. The calculation of each normalized weight W ^* ; , is carried out by dividing the weight of a given signal by the sum of the weights of all weight elements.

The block calculating the parameters of the input signals 6 (PDU BC) carries out the multiplication of each signal from the output of input block 1 to the corresponding normalized weight W _{ : Pvx.j = Xj- W ^* ; .

After the PDU BC 6, the signals are fed to the adder parameters of the input signals (SP BC) 7, where the total parameter of the input signals is calculated:

_.

The limiter of the fraction of signals in the state of NPE (OD) 8 with its single-channel GSChZ 9, operating similarly to GSCh1 and GSCh2 and issuing the value of Kshchz, multiplying by which the value of K _ode specified in the range from 0 to 1, allows you to vary the amplitude of K _ode according to a given law and within the given boundaries also in the range from 0 to 1.

The unit for calculating the input part of the state of NPE 10 multiplies the signal that carries information about the value of the total parameter of the signals by the proportion of input signals: Cvx. = Pvx. _∑ ^• K _od .

The unit for setting the internal state of the NPE And, with its single-channel RNF4 12, determines the value corresponding to the internal state of the NPE, C _{∞ b} .. At the same time, RNP4 12 gives the value of Kgg4, multiplying by which the amplitude values of the magnitude of the eigenstate of the NPE allows you to vary the amplitude C _cob. according to a given law and within given boundaries in the range from 0 to 1.

The unit for calculating the internal part of the state of NPE 13 multiplies the value of evaluating the intrinsic state of NPE by its share in the general state of NPE, i.e. the difference between the unit and the signal value of the limiter of the share of signals (1 -K _ode ): Cvn. = C _COb. * (lK _Od ).

IPPs state calculation unit 14 (RSB) performs addition of this state-dependent directly or through a deterministic function of the input signals Cvx. (Or Cvx. ^Ds) and the inner part IPPs state EHV .: Cnpe = Cvx. (Or Cvx. ^dz ) + Svn.

The unit for setting the accuracy of self-study of the neural network (BUT) 17 implements setting the accuracy of the self-learning of the neural network ε, which allows, after starting at least one of the NPEs of the neural network, where they operate, transients caused by the intrinsic state of the NPE or changes in the parameters of the input signals, other parameters of the NPE (K _o , BDZ), as well as a combination of such changes , complete at a certain predetermined ε stage the passage of pulses in the neural network, thereby completing the HC self-learning process in the microcycle and fix the current state of the NPE at the output.

The block of determinate dependencies (BDZ) 18, the input of which receives the signal Cx., Generates a signal Cbx at the output. ^Dz calculated by means of the determinate function of the input: Cx. ^dz = f (Cvx.), given by the control unit. The signal from the output of the database 3 is fed to the unit for calculating the state of the NPE (BRS) 14.

The analyzer of the magnitude of the change in the state of the NPE (AC) 16, which receives a signal from the memory unit 15 about the previous value of the state of the NPE, SNPE ⁰ ^, and the signal from the BRS 14 about the current state of the NPE, SNPE ^k , analyzes the absolute value of the difference between these signals and compares with the value of the accuracy of self-training of the neural network ε arriving at AC 16 with BUT 17: I ΔСнпэ | <ε. When this condition is fulfilled, a signal is supplied to the control unit 3, which gives a command to the clock generator 21 by this signal and completes the passage of pulses in the neural network, thereby completing the HC self-learning process in the microcycle. The current value of the NPE state is fixed at the output.

NPE works as follows.

In the initial state, a numerical deterministic or expert estimate is set:

- the internal state of NPE (a value in the range from 0 to 1); Yu

- weights of the input signals Wj (values in the range from 0 to 1);

- the proportion of input signals in the state of NPE, Ki (value in the range from O to l);

- the value of the accuracy of self-study HC, ε (the value is usually 10 ^"1 10 ^{" 5} )

- operating procedure (distribution law, range of variation of the issued numbers) required for the operation of the NPE random number generators.

Connection of at least one RNG is made upon command from the control unit. At the same time, the included RNGs issue and record one random signal (in accordance with the established operating procedure). In non-functioning, i.e. in the off state, a signal equal to 1 is output at the RNG output. Next, the input signals are fed to the input unit, the signals (signal) pass through the NPE blocks and, in accordance with the established algorithm (option) of the NPE operation, they are converted with the output, ultimately , to the output of the output signal Y, which is either an input signal for other NPEs, or the final signal of the neural network.

It is accepted that in the neural network NPEs that do not have synaptic connections with other NPEs in input are input NPEs that have synaptic connections with other NPEs, both in input and output, are intermediate NPEs, while intermediate NPEs with their own state of NPEs are intermediate -input NPE, and NPE, which has synaptic connections with other NPEs only at the input, is the output NPE.

The work of the NPE is possible in twelve main options.

The first option corresponds to the operation of the NPE in the mode of weighing, normalization and simple summation of input signals (effects). The value of K _o = 1. In this case, control unit 3 sends commands to multichannel random number generators 2 and 5 and to single-channel random generators numbers 9 and 12 and sets the signal value at their outputs to 1, i.e. the initial state of the NPE is set equal to 0. At the output of the unit for calculating the state of the NPE (BRS):

Cnpe = Cvx. or Cnpe = Cbx. ^ds where BZD 18 is connected.

In this embodiment, the state of the NPE is defined as the sum of the parameters of the input signals (effects), and the signal Y corresponding to the state of the NPE is fed to the input of the output block 20:

Y = Cnpe = Pvx. _Σ = ∑ Pvx.i = ∑ (xi -W ₁ ^* ).

The option corresponds to the case of the operation of the NPE as an intermediate or output NPE in the neural network (HC), in the structure of which it consists.

The second version of the NPE operation differs from the first one by the operation of the RNG1 2 and / or RNG2 5. Accordingly, from the output of the BV 1, the signal Xi ^RNl = X _f Kgschlj is received, the signals of the scales Wi ^GCch2 - Wj ^• Kgcch2 are generated in the IAC _; and signals of normalized weights

Wi ^HSCh2 / Σ (Wi ^HSCh2i ).

This option, as well as the first option, corresponds to the operation of the NPE in the mode of weighing and simple summation of the input signals (effects), but taking into account possible scatter according to the given distribution laws and in the given range of the value and / or weight of at least one input signal:

Y = Σ (Xi ^GCchli - Wi ^{* gcch2i} ) _.

The option corresponds to the case of the operation of the NPE as an intermediate or output NPE in the neural network.

The third version of NPE operation involves setting the initial intrinsic state of NPE in the absence of input signals. The value of K _od = 0, i.e. the proportion of input signals in the state of the NPE is 0. The signal corresponding to the value of the original intrinsic state of the NPE, C _COb , is input output block 20.

In this embodiment, the NPE does not act as an adder, but as a direct source of signals, i.e. in the role of the input NPE of the neural network:

The fourth option differs from the first and third options in that it involves their combination. In this case, the control unit 3 sends commands to the random number generators 2 and 5, and to the single-channel random number generators 9 and 12 and sets the signal value at their outputs to 1.

Y

-WGUK _OD + C _co. - (1-K _od ).

The option corresponds to the case of the operation of the NPE as an intermediate input or output NPE in the neural network.

The fifth NPE operation variant differs from the third option in that BU 3 gives commands to a single-channel random number generator 12 and changes, according to a given distribution law and within certain limits, the value of the NPE's own state at its output: C _COb . ^hch3 - C _COb. * Kgch ₃ . This option corresponds to the case when the NPE, as well as in the third embodiment, is the input NPE in HC, i.e. acts as a direct source of signals, but taking into account various scatter according to a given distribution law and within specified limits in assessing the value of the initial intrinsic state of NPE:

Y _ ^ i

The sixth version of the NPE operation differs from the third option in that it assumes the presence of input signals and the operation of the RNG in the input unit and (or) in the IAC, as well as the presence of a fraction of the state of the NPE, K _ode , depending on input signals, and the share of the state of NPE, (1-K _OD ), independent of the input signals. The value of the share limiter K _ОД =] ОД [.

This option corresponds to the case when the NPE acts as an adder of input signals with possible changes according to a given distribution law and within certain limits of the value and / or weight of at least one of them, and not depending on the input signals, the share of the NPE state:

Y = Cnpe + Cvn = Pvx. _Σ ^gcch _0D -K + IOS. = ∑ (Xi ^ГСчli - W ₁ ^{* гсч2i} ) -K _o + C _co. - (1-K _od ).

The seventh version of the NPE operation differs from the sixth option in that only the RNG4 12 in the BC 11 are functioning from the RNG, which allows changing the value of the NPE intrinsic state according to a given distribution law and within certain limits: C _CoG ^GCH3 - ⁼ C _Cob. 'Kghh _. . K _ode -] 0.1 [. The signal at the output of NPE:

Y = Snpe + St. = Pvx. _∑ -K _od + Svn. = ∑ (x _g W ₁ ^* ) ^• K _od + Ccob. ^gcc4 - (1-K _od ).

The eighth option for operation of the NPE differs from the sixth option in that the GSPZ 9 of the limiter of the fraction (OD) 8 is additionally included, which allows changing, according to a given distribution law and within certain limits, the proportion of the state of the NPE, K _ode , which depends on the input signals. K _OD =] 0.1 [. The signal at the output of NPE:

Y = Cnpe + Cvn. = Pvx. _∑ ^rcch- Code ^hdd3 + Cbn = ∑ (Xi ^{GSChli -Wi} ^{* hch2i} ) ^• Code ^hdd3 + C _006- - (I-Code ^hdd3 ).

The option corresponds to the case of the operation of the NPE as an intermediate input or output NPE in the neural network. The ninth variant of the NPE operation differs from the eighth variant in that the RNGs do not function in the input unit and (or) in the unit for setting and normalizing weight coefficients. K _ode =] 0, 1 [. The signal at the output of NPE:

Y = SNPE + Svn = Pvx.∑-Code ^gsc3 + Svn. = E (X ₁ -W ₁ ^* ) ^• Code ^hdd3 + Comb .- (1-Code ^hdd3 ). The option corresponds to the case of the operation of the NPE as an intermediate input or output NPE in the neural network.

The tenth version of the NPE operation differs from the sixth option in that the GESZ 9 has a share limiter (OD) 8, which allows changing, according to a given distribution law and within certain limits, the proportion of the state of the NPE, K _ode , which depends on the input signals. Code =] 0D [.

This option corresponds to the case when the NPE acts as an adder of input signals with possible changes according to a given distribution law and within certain limits of the value and / or weight of at least one of them and the share of the NPE state, depending on the input signals:

Y = Cnpe + Cvn. = Pvx. _Σ ^ghc- Code ^ghch3 + Cvn. = ∑ (x _i ^ghcli -W _i ^{* ghch2i} ) -Kod ^ghch3 + Cob .- (l- Code ^ghch3 ).

The eleventh option of NPE operation differs from the tenth option in that it assumes the operation of GSC4 12 in the unit for assessing the internal state of NPE (BO VS) 11. Code =] 0, 1 [. In this option, all RNGs work.

The signal at the output of NPE:

Y = Cnpe + Cvn. ^hh = pvx. _∑ ^hdd- code ^hdd3 + cvn. ^gcc = ∑ (x _i ^gccli -W _i ^{* gcc2i} ) -Code ^gcc3 + Cbs. ^hdd4 - (1- ^{code hdd3} ). The option corresponds to the case of the operation of the NPE as an intermediate input or output NPE in the neural network. This is the most common version of the NPE.

The twelfth variant of NPE operation can be for all variants where K _O d ≠ 0. It assumes the calculation of the input part of the NPE state by means of deterministic dependencies in the block of deterministic dependencies (BDZ) 18, the signal to which the key 19 sends according to the command BU 3. The value of the generated BDZ signal is determined by the known functional dependence of the signal determining the input part of the NPE state from the input signals. The signal corresponding to the calculated value of the input part of the NPE state: C ¹⁴¹⁴ Bx. ¹³ = f (Ex.), Is supplied to the key output and to the BRS I.

The NPE (AI) 16 state change analyzer contained in the NPE operates with all NPE operation options, comparing the absolute value of the difference between the values of the current, SNPE ^k and previous, SNPE ^, NPE states with the neural network self-learning accuracy established in BUT 17, ε. Under the condition | ΔСнпе <ε with AI, a signal is supplied to the control unit 3 to stop clock pulses by the clock generator 21 in the pulse train and to fix the last NPE state value, SNPE on the NPE output block 20.

Work (random numbers) of any RNG or any channel of the RNG NPE, if their work is provided, occurs only at the first clock pulse from sending clock pulses. During the passage of the remaining clock pulses of the pulse sending, the set value is fixed at the RNG outputs. Industrial applicability

The present invention can be used to create NPE, allowing to implement on its basis a neural network for solving a class of problems of evaluating the functioning of open complex systems (OCC) and assessing the degree of optimality of various solutions by providing the possibility of constructing a model of the studied system, both hierarchical and recurrent, taking into account different initial and working conditions of its elements and their functioning options, taking into account when modeling neural-like elements of individual elements of the system as a level l their self-sufficiency, and exposure to external signals. At the same time, the proposed NPE makes it possible to take into account the errors in setting its parameters and the parameters of the input signals, as well as to ensure the specified accuracy of the self-learning of the neural network.

Claims

Claim

A neural-like element (NPE) containing an input unit including a first random number generator, the number of outputs of which is equal to the number of inputs of an input unit, the signal from each output of the first random number generator determines the change in the value of the corresponding input signal according to the law and within the limits determined by its type and error tasks, unit for setting and normalizing weighting coefficients, the number of weight elements of which is equal to the number of inputs of the input unit, the weight setting in the weight elements is carried out by means of a second generator random numbers according to the law and within the limits determined by the type and error of its assignment, each weighting coefficient is normalized by dividing each weight by the sum of all the weight elements of the weight, and each normalized weighting factor sets the corresponding synaptic connection of the outputs of the input block and the inputs of the block for calculating signal parameters, the outputs of which are connected to the inputs of the adder of signal parameters, the output of which is connected to the first input of the unit for calculating the input part of the NPE state, the second input of which is connected to a limiter of the share of signals, a single-channel random number generator that determines the signal value of the limiter of the share of signals according to the law and within the limits determined by the type and error of setting this value, a unit for setting the internal state of the NPE, a second single-channel random generator that determines the value of the signal of the unit for setting the internal state of the NPE the law and within the limits determined by the type and error of its assignment, the output of which is connected to the input of the unit for calculating the internal part of the state of the NPE performing knife signal from the output of the unit for setting the internal state of the NPE to the difference between the unit and the signal value of the limiter of the fraction of signals, the output of the unit for calculating the internal part of the state of the NPE is connected to the first input of the unit for calculating the state of the NPE combining the signal from the output of the unit for calculating the input part of the state of the NPE with the signal coming from the output of the unit for calculating the internal part of the state of the NPE, the output of the unit for calculating the state of the NPE is connected to the output unit, with the first input of the analyzer, the magnitude of the change in the state of the NPE directly, and with its second input through a memory unit that stores information about the previous value of calculating the state of the NPE, a unit for setting the accuracy of training the neural network, the output of which is connected to the third input of the analyzer of the magnitude of the change in the state of the NPE, to which compares the accuracy of self-learning with a value equal to the difference between the previous value of calculating the state of NPE and the current one, and if the specified condition is met, it sends a signal to the control unit to complete the process of self-learning of the neural network, a control unit that controls the random number generators, clock generator and the passage in the neural network of pulses.

2. A neural-like element according to claim 1, characterized in that a switch is inserted between the output of the unit for calculating the input part of the NPE state and the input of the unit for calculating the NPE state, the second output of the switch is connected to the input of the input block of deterministic dependencies, the output of which is connected to the first output of the switch moreover, the magnitude of the signal generated by it is determined by the known functional dependence of the signal determining the input part of the state of the NPE from the input signals.