RU2514931C1 - Method for intelligent information processing in neural network - Google Patents

Abstract

FIELD: information technology.

SUBSTANCE: composite signal, consisting of a useful signal, summed with a first part of pre-generated low-power noise and a similar second part of said noise, is transmitted to a multilayer recurrent neural network with control synapses as a signal which is pre-decomposed into components, with each component converted to a sequence of unit characters with repetition frequency as a predefined function of the amplitude of the component. The composite signal is presented in form of successive groups of unit characters. When transmitting groups of unit characters from layer to layer, unit characters relating to the useful signal are copied from the signal-noise groups into corresponding noise groups. Copies of the signal-noise groups of unit characters are formed from the noise groups and said copies are processed based on change in shape of cross-sections of diverging unit characters and rotations of said characters around directions of transmission thereof based on the current states of the layers.

EFFECT: high depth and intelligence of processing information.

7 dwg

Description

The invention relates to bionics, modeling of functional aspects of a person and can find application in computer technology for the construction of intelligent machines and systems.

Currently, many methods of intelligent information processing in a neural network are known, used for approximation, classification, pattern recognition, speech processing, forecasting, identification, estimation of production processes, associative control and for solving other creative tasks (Khaikin S. Neural networks: full course , 2nd edition .: Translated from English by M .: Publishing House "Williams", 2006. - 1104 p .; AI Galushkin, Theory of Neural Networks, Book 1: Textbook for universities / General ed. .I. Talushkina. - M .: IPRZhR, 2000 .-- 416 p.).

However, all of them are characterized by narrow capabilities for storing structurally complex time-varying signals, their recognition, association with other signals, retrieving from the network memory and reproducing in its original form. They do not allow, when processing information, to solve a wide range of problems with the same neural network.

The closest analogue of the invention is a method of intelligent information processing in a multilayer neural network with feedbacks closing the circuits with a delay time of single images less than the immunity time of network neurons after their excitation. According to it, the signal is fed into the network after decomposition into components in a basis consistent with the input layer of the network. In this case, each component is converted into a sequence of single images with a repetition rate as a predetermined function of the amplitude of the component before being fed into the network. The signal is represented on the network in the form of consecutive sets of single images in accordance with predefined rules for its recognition, taking into account the inverse recognition results. When transferring sets of single images from layer to layer, they are shifted along the layers, taking into account the current states of the latter. Recognition results are stored on network elements. As processing results, sequential sets of single images are used on the output layer of the network after the reverse conversion into the corresponding source signals (Osipov V.Yu. Associative Intelligent Machine / Information Technologies and Computing Systems. No. 2, 2010, P.59-67).

The disadvantage of this method is that, according to it, a neural network cannot perceive one thing, but give out as a result of processing a completely different one, weakly connected with input current signals, as a person can do. The results of processing information on the network over time are closely related to input signals. The network functions as an associative memory. In the known method, the input signals cause the stored signals associated with them from the long-term network memory and together represent the processing results. Long-term network memory is synapse memory. Network RAM is the memory on the neurons themselves. According to this method, continuous input signals suppress the network's ability to process inverse recognition results. This significantly limits the depth of information processing. Processing depth is understood as the time during which the network is able to process the separately allocated short signal in it with a continuous input stream.

In general, these disadvantages significantly limit the functionality of the method for intellectual information processing.

The objective of the invention is to expand the functionality of intelligent information processing in a neural network.

The technical result from the use of the invention is to increase the depth and intelligence of information processing in a neural network.

The problem is solved in that in the known method of intelligent processing of information in a neural network, which consists in supplying a multilayer recurrent network with feedbacks, closing circuits with a delay time of single images less than the immunity time of network neurons after their excitation, a signal decomposed into components in the basis matched with the input network layer, with each component converted into a sequence of single images with a repetition rate as a predefined function it from the amplitude of the component, representing the signal in the form of successive sets of single images in accordance with predefined recognition rules taking into account the inverse recognition results, shifts of sets of single images along the layers taking into account the current state of the layers, storing recognition results on network elements, using as results processing consecutive sets of single images on the output layer of the network after the inverse transformation to the corresponding to them, the initial signals, according to the invention, a preformed group signal consisting of a useful signal summed with the first part of the preformed noise and the second of the same part of this noise when transmitted from layer to layer of group sets of single images is fed into the network as a signal, each of which consists of a signal-noise group and a noise group of single images, copy from the signal-noise groups single images related to the useful signal into the corresponding pppam noise groups, form from noise groups copies of signal-noise groups of single images and process these copies taking into account changes in the cross-sectional shapes of diverging single images and rotations of these images around the directions of their transmission, taking into account the current state of the layers.

The results of biological studies indicate that the processing of external signals by the human brain is carried out in the presence of internal noise. At the same time, internal noises as well as external useful signals stimulate the brain.

With this in mind, the inventor proposes to supply a pre-formed group signal to the network as a signal, consisting of a useful signal summed with the first part of the pre-generated noise and the second of the same part of this noise, which is a new significant feature of the invention.

Another new significant feature of the invention is the fact that when transferring from layer to layer group sets of single images, each of which consists of a signal-noise group and a noise group of single images, single images related to the useful signal are copied from signal-noise groups noise groups corresponding to these groups.

The third new significant feature of the invention is the fact that it is proposed to form copies of signal-noise groups of single images from noise groups and process these copies taking into account changes in the cross-sectional shapes of diverging single images and the rotations of these images around their transmission directions, taking into account the current state of the layers.

The invention is illustrated in figure 1 ÷ 4.

Figure 1 shows the structural diagram of a two-layer recurrent neural network with controlled synapses that implements the proposed method, where 1, 5 - the first, second layers of neurons; 2, 6 - the first, second blocks of unit delays; 3, 4 - the first, second blocks of dynamic synapses; 7 - control unit synapses.

Figure 2 a, b shows an example of the logical structure of this recurrent neural network with shifts of group sets of individual images along the layers. Fig. 2a is a plan view of a network with a spiral scheme for promoting populations of single images along layers, and Fig. 2b is a section of a network in two layers along the first line, where d, q are the values of unit shifts of populations, respectively, in X, Y coordinates; 1.1 and 1.2, 2.1 and 2.2, 3.1 and 3.2, ......., L.1 and L.2, ML.1 and ML.2 are the same subfields, respectively, of the first, second, third and other fields formed due to group shifts sets of single images along the layers during transmission from layer to layer. The arrows in FIGS. 2 a, b show the directions of advancement of sets of single images and their signal-noise and noise groups.

Figure 3 a, b shows an example of a group of individual images before transmission from one layer to another (Fig. 3a) and after transmission (Fig. 3b), explaining the formation of copies of signal-noise groups of single images.

Figure 4 shows examples of smooth elliptical cross-sections oriented at 90 ° in the plane of the receiving layer of divergent unit images, where 1 is a neuron of the transmission layer, forming a single image related to the signal-noise group; 2 - neuron of the transmission layer, forming a single image related to the noise or to a copy of the signal-noise group; 3, 4 - short synapses for which the attenuation function is equal to one; 5 - axis of the maximum extent of the distribution of power density in the cross section of a divergent single image transmitted from neuron 1 in the direction of neuron 7; 7, 9 - neurons of the receiving layer, in the direction of which, respectively, divergent single images are transmitted from neuron 1 and neuron 2; 8 - dash-dotted line dividing the fields of the receiving layer into subfields; 10 - direction of rotation of the axis of the maximum extent of the distribution of power density in the cross section of a diverging single image. The points around which the turns are carried out are neurons 7 and 9 in this case, in the direction of which, respectively, diverging single images are transmitted from neuron 1 and neuron 2; 11 is a line dividing the lines of the fields of the receiving network layer; d, q are the unit shifts of the sets, respectively, along the coordinates X, Y. Figure 4 shows two fields of the receiving network layer. The top margin refers to one row, and the bottom margin refers to another. It is believed that the promotion of group sets of individual images, taking into account figure 2, along the top row is carried out from left to right, and the bottom - from right to left.

Figure 5 shows the same example, but with the orientation of the elliptical shapes of the cross sections of the diverging individual images by 180 °. The orientation of these forms is understood as the orientation of the power density distribution (the angle between the axis of the laboratory coordinate system and the axis of the maximum length of the power density distribution). The designations shown in figure 5 are the same as in figure 4.

The method is as follows.

Consider it on the example of a neural network, a structural diagram of which is presented in figure 1.

As a signal, decomposed into components in a basis matched to the input layer of the network, with each component converted into a sequence of single images with a repetition rate as a predetermined function of the amplitude of the component, a preformed group signal consisting of a useful signal summed with the first part of the pre-formed noise, and from the second part of the same noise. As noise, low-power noise with respect to the useful signal is used in the frequency band of this signal. For example, such noise can be low-power noise with a uniform spectral density in this band. The first and second parts of this noise are formed by branching the noise from its generator.

Such a group signal is fed to the first input of the first layer of the neural network, which is divided into 1.1 and 1.2 inputs. Each sequence number is assigned in the general case, the frequency and spatial component corresponding to the signal. The details of this conversion are known (Osipov V.Yu. Direct and inverse signal conversion in associative intelligent machines / Mechatronics, automation, control. No. 7, 2010, P.27-32).

After applying such a group signal to the network on the first layer at the output of this layer, there are consecutive group sets of single images that carry all the information about the input signal and noise. Each of these group aggregates consists of a signal-noise group and a noise group of single images.

After a delay in the first unit of unit delays 2, successive group sets of unit images arrive at the first block of dynamic synapses 3. Each unit image from the current set is fed simultaneously in the first block of dynamic synapses 3 to the set of its dynamic synapses, which provide communication of each neuron that generated the unit image in in the general case with all the neurons of the second layer of neurons 5. Due to this branching, divergent single images are formed.

A feature of dynamic network synapses is as follows. Provided that the conductivity of the synapses is significantly less than the input conductivity of the neurons, the amplitude of the current of the single image at the output of each synapse is equal to the amplitude of the voltage of the input single image (pulse) multiplied by the weight (conductivity) of the w _ij (t) synapse. The weights w _ij (t) of the synapses are determined through the product of their weight coefficients k _ij (t) and attenuation functions β (r _ij (t), G _ij (x _1t , x _2t )),

w _ij (t) = k _ij (t) β (r _ij (t), G _ij (x _1t , x _2t )),

where r _ij (t) is the remoteness of neurons connected through synapses (the distance between them on the X, Y plane). It is believed that the distance between the interacting layers of the neural network tends to zero; G _ij (x _1t , x _2t ) is the directivity coefficient of the cross-sectional shape of the divergent single image in the direction from the i-th to the j-th neuron in the X, Y plane; x _1t , x _2t - current state, respectively, of the first and second layers at time t.

. При невозбужденном i-м нейроне передающего слоя и возбужденном j-м нейроне принимающего слоя $$b_{ij}(t)=D\cdot I_{ij}^{-}(t)<0$$

. В других случаях b_ij(t)=0. Здесь приняты следующие обозначения: A, D - константы; $$I_{ij}^{+}(t)$$

, $$I_{ij}^{-}(t)$$

- положительный и отрицательный ток, протекающий через синапс, соответственно, в прямом и обратном направлении.Weighting factors, k _ij (t), vary depending on the effects on the synapses of a single image and act as elements of a long-term network memory. When single images pass through synapses, they remove information about previous effects from them and leave information about their appearance through changes in weight coefficients. The values of the weight coefficients k _ij (t) can be calculated by the formulas: k _ij (t) = 1-exp (-γ · g _ij (t)), where γ is a constant positive coefficient; g _ij (t) is the number of unit images stored at the corresponding synapse, g _ij (t) = g _ij (t-1) + b _ij (t); g _ij (t-1) is the number of single images stored at the synapse at the previous time step; b _ij (t) is the effect of a single interaction of neurons at the t-th moment in time. In the particular case, if the ith excited neuron of the transmitting layer (together with other neurons) excites the jth neuron of the receiving layer, then $$b_{ij}(t)=A\cdot I_{ij}^{+}(t)>0$$

. With the unexcited i-th neuron of the transmitting layer and the excited j-th neuron of the receiving layer $$b_{ij}(t)=D\cdot I_{ij}^{-}(t)<0$$

. In other cases, b _ij (t) = 0. The following notation is accepted here: A, D - constants; $$I_{ij}^{+}(t)$$

, $$I_{ij}^{-}(t)$$

- positive and negative current flowing through the synapse, respectively, in the forward and reverse directions.

Each of the connections (synapses) has its own value of the attenuation function β (r _ij (t), G _ij (x _1t , x _2t )) of single images, depending on the distance r _ij (t) of the neurons connected through the synapses and the directivity of the cross-sectional shape divergent single image in the direction from the i-th to the j-th neuron in the X, Y plane.

The function β (r _ij (t), G _ij (x _1t , x _2t ,)) can be defined as:

$$\beta (r_{ij}(t),G_{ij}(x_{\mathrm{one}t},x_{2t}))=\frac{\mathrm{one}}{\mathrm{one}+\vartheta \cdot \sqrt[h]{r_{ij}(t)}/G_{ij}({\varphi }_{ij}+\psi (x_{\mathrm{one}t},x_{2t}))},(\mathrm{one})$$

where h is the degree of the root, the higher it is, the wider the associative spatial interaction in the network; ϑ is a positive coefficient; φ _ij is the angle characterizing the direction of the receiving layer to the neuron from the neuron of this layer, in the direction of which a divergent single image is transmitted; ψ (x _1t , x _2t ) is the angle of rotation of the axis of the maximum extent of the power density distribution in the cross section of the diverging single image (angle of rotation of the diverging single image around the direction of its transmission). Types of predefined cross-sectional shapes of divergent unit images in a particular case can be assigned to the numbers of neurons that form these unit images. In this case, the value of the coefficient G _ij (φ _ij + ψ (x _1t , x _2t )) of the directivity of the cross-sectional shape of the diverging unit image also depends on the number of the neuron forming the unit image. In this case, we are talking about the maximum possible forms of cross sections of individual images, determined under the condition k _ij (t) = 1.

Included in (1), the value of r _ij in units of neurons, taking into account the possible spatial shifts of the aggregates of individual images along the network layers, can be expressed as:

; $$r_{ij}=\sqrt{{(\Delta x_{ij}+n_{ij}\cdot d)}^2+{(\Delta y_{ij}+m_{ij}\cdot q)}^2}$$

;

n _ij = ± 0, 1, ..., L-1; m _ij = ± 0, 1, ..., M-1;

Δx _ij , Δy _ij - the projection of the connection of the j-th neuron with the i-th on the X, Y axis without taking into account spatial shifts; d, q are the values of unit shifts, respectively, in the coordinates X, Y; L, M - the number, respectively, of columns and rows into which each layer of the neural network is divided due to shifts.

By changing Δx _ij , Δy _ij by the corresponding values of n _ij · d and m _ij · q, we can change r _ij and the flow direction of the sets of single images along the network layers.

Such shifts of the aggregates of single images along the layers are realized through the control of dynamic synapses from the synapse control unit 7.

To make a decision in the synapse control unit about the shift of the next set of single images, it first analyzes the state of the first and second layer. In the case when at the output of the first or second layer there is a collection of single images that cannot be transferred to another layer by short links due to the presence of the corresponding neurons of this layer in immunity states, this population is shifted. In the case where there are no obstacles to transmitting the totality of single images through short connections, it does not shift.

The angle φ _ij on which G _ij depends (φ _ij + ψ (x _1t , x _2t )), which characterizes the direction of the receiving layer to the neuron from the neuron of this layer, in the direction of which the divergent unit image is transmitted, is equal to φ _ij = arctg (Δx _ij / Δy _ij ), where Δx _ij , Δy _ij are the distances between the corresponding neurons in the X, Y plane of the receiving layer.

To determine G _ij (φ _ij + ψ (x _1t , x _2t )), one can preliminarily specify the types of cross-sectional shapes of divergent unit images. Circular, elliptical and other shapes are possible, without bias and with bias of energy maxima. For example, let one of them be an elliptical shape, initially oriented along the Y axis. For this form, each value of the angle χ from the Y axis corresponds to its own directivity coefficient G (χ). This data on G (χ) can serve as the source for determining the values of the coefficient G _ij (φ _ij + ψ (x _1t , x _2t )) in (1). The determination of ψ (x _1t , x _2t ) taking into account the current states of the network layers is feasible by searching for the angle at which at a given time interval the maximum number of unit images associatively called from the long-term network memory can be reached. Taking into account the inertia, this angle can be determined by varying the values of the rotation angle and estimating the number of unit images associatively called from the long-term memory. The rotation of the axis of the maximum extent of the power density distribution in the cross section of the diverging unit image by the angle ψ (x _1t , x _2t ) is equivalent to the rotation of this image by this angle around the transmission direction. In the particular case, these rotations can be carried out at predetermined angles, taking into account the current state of the layers, numbers of excited neurons.

In the general case, displaced group aggregates with altered cross-sectional shapes of diverging single images and with rotations of these images around the directions of their transmission from the output of the first block of dynamic synapses 3 enter the input of the second layer of neurons 5.

All single images arriving at the same neuron at different synapses are summarized. When this sum is exceeded a predetermined threshold of neuron excitation, it is excited and a single image is formed at its output. Then the available amount is reset, and the neuron itself then goes into the immunity state of the input signals. It contains the specified time, the same for all neurons of the network, which is greater than the total delay of single images in blocks 1, 2, 3, 5, 6, 4 included in the two-layer multi-beam circuit of the neural network.

Note that the power of single images at the input of neurons is higher, the less its loss at the synapses. The power dissipated at each synapse is determined by the square of the current flowing through it, multiplied by the resistance of the synapse, equal to the reciprocal of its conductivity - the weight of the synapse.

Successive sets of single images from the output of the second layer 5, after a delay in block 6, go to the second block of dynamic synapses 4. In block 4 they are processed in the same way as in block 3 and, in general, with altered forms of cross sections diverging single images and with the rotations of these images around the directions of their transmission arrive at the second input of the first layer of neurons 1.

With this in mind, the direct and inverse sets of unit images arriving at the first layer of 1 neurons in it are correctly connected, recognized, and generate at its output new sets of unit images that carry information about both the current and previously memorized signals associated with the first . At the same time, due to the corresponding shifts of the sets of single images along the layers, the overlapping of inverse recognition results on direct sets is excluded. By changing the shapes of the cross-sections of divergent single images and the rotations of these images around the directions of their transmission, it is possible to quickly change communication levels between individual single images, their groups and groups in the network.

Due to the priority of short connections in a neural network, an unambiguous correspondence between the components of the input and output signals is easily established between the input and output of the network.

Given this correspondence, the frequency and spatial characteristics of the components of the original signal are determined by the numbers of neurons that form the sequence of single images at the output of the device. According to the repetition frequencies and relative delays of individual images, respectively, the amplitudes and phases of the components of the original signal are set. Then reproduce the components of the original signals and by adding them restore the original, for example, speech and other signals. To determine the amplitudes of the components of the source signal, the current number of unit images falling into a predetermined time interval is determined. To reproduce the components of the source signals, we apply, for example, their digital synthesis according to known parameters.

Due to the cyclical exchange of information of the first layer with the second, the entire space of the layers (Fig.2 a, b) is divided into ML fields. In addition, each field is further subdivided into two identical subfields. According to the invention, a useful signal, summed with the first part of the pre-formed noise, after decomposition into components and their transformation into a sequence of single images, is fed to the first subfield 1.1 of the first field of the first layer (Fig.2a). To the second subfield 1.2 of the first field of the first layer, after decomposition into components and their transformation into a sequence of single images, the second part of the previously generated noise is identical to the first part. Sets of single images move along the network in the indicated directions and are associated with each other. In this case, the input sequential sets of single images are converted to parallel, and removed from the network after associations again as sequential. The main output of the network according to the proposed method, in accordance with figure 2 a stands output ML.2. The output of ML.1 is auxiliary.

According to figure 2 a, b, there is a scheme for promoting group sets of single images in the form of a spiral with their counter advancement.

Due to the fact that, in the presence and absence of a useful signal, two identical parts of pre-formed noise are fed in parallel and continuously, in the form of consecutive group sets of unit images, continuously strong associative components are established and supported between the same signal-noise and noise components in the network communication through the corresponding synapses. For these synapses, the values of the weight coefficients k _ij (t) tend to unity with time. Provided that the values of the attenuation functions β (r _ij (t), G _ij (x _1t , x _2t )) for these synapses are not close to zero, through these synapses, when transmitting from the layer to the layer of group sets of single images, copying from the signal - noise groups of single images related to the useful signal into the noise groups corresponding to these groups. In the case when, due to changes in the cross-sectional shapes of diverging unit images and (or) rotations of these images around the directions of their transmission, taking into account the current state of the layers, the attenuation functions β (r _ij (t), G _ij (x _1t , x _2t )) tend to zero, such copying becomes impossible. When it has already happened, copies of the signal-noise groups with the corresponding β (r _ij (t), G _ij (x _1t , x _2t )) close to zero are “detached” from the originals and processed independently of them, taking into account the inverse results recognition of these copies and network impacts of the second part of low-power noise. In this case, low-power noise only stimulates the functioning of the network and creates new conditions for processing these copies. The suppression of the reverse results of recognition of copies by the stream of input signals is excluded. The depth of information processing in the neural network increases, as the number of cycles of processing the inverse recognition results increases.

Figure 3 a, b shows an example explaining the copying of signal-noise groups into the noise group of single images. Given that the power of the processed noise is significantly less than the power of the useful signal, it does not significantly affect the useful signal.

It can be seen from Figs. 4 and 5 that by changing the shapes of the cross sections of diverging single images and by turning these images around the directions of their transmission, taking into account the current state of the layers, it is possible to significantly affect the relationship of the useful signals processed in the network and their copies. In accordance with figure 4, the cross sections of divergent single images related to signal-noise groups, cover the neurons responsible for the formation of copies of the signal-noise groups. Under these conditions, such copies can be successfully formed. To “detach” copies of signal-noise groups from the originals, it is enough to carry out an additional 90 ° clockwise rotation of the axes of the maximum length of the power density distribution in the cross sections of diverging single images. The result of this rotation is shown in FIG. Figure 5 shows that, unlike figure 4, the cross section of the divergent single image formed by neuron 1 does not cover neuron 9. In this case, neuron 1 cannot excite neuron 9. But the cross sections of divergent single images well cover the neurons of others subfields of the second line, responsible for the formation of the same types of groups of single images. As a result, independent full-fledged processing in the network of originals and copies of signal-noise groups is ensured.

If the elliptical shape of the cross sections of divergent single images is transformed for a given time, for example, into a circular one, then the situation regarding the relationship of copies and originals will also change significantly.

The formation of copies of signal-noise groups and their processing in the network, taking into account the noted conditions, can significantly expand the functionality for the intellectual processing of information in a neural network. Increases the depth and intelligence of information processing. The opportunity is provided to temporarily tear off the "thoughts" of the network from the perceived stream of input signals, as well as generate new conscious signals, taking into account the influence of low-power noise.

To prove the advantages of the proposed method in comparison with the known solutions, a software model was developed that implements this method of a two-layer recurrent neural network with controlled synapses with the structure shown in Fig. 1. The number of neurons in each layer was 2100 units. Due to shifts of group sets of single images along the layers (during transmission from layer to layer), the layers were divided into two rows of 25 fields of 6 × 7 neurons each. Moreover, each field was divided into two subfields of size 3 × 7. The processing of such a network of group signals according to the proposed method was simulated. It has been established that due to pre-generated noise applied in parallel to the first and second subfields of the first field of the first network layer, it is possible to provide an unambiguous correspondence between the same signal-noise and noise components and successfully generate copies of signal-noise groups when processing the group signals under consideration. By changing the shapes of the cross sections of divergent single images and the rotations of these images around the directions of their transmission, you can significantly affect the relationship of copies of signal-noise groups and their originals. Due to the formation of signal-noise copies and “separation” of them from the originals, it is possible to eliminate the suppression of the inverse recognition results of these copies by the input signal stream. This increases the depth and extends the functionality of intelligent information processing in a neural network.

A method of intelligent information processing in a neural network can be implemented using a well-known element base. As the neurons of the layers and elements of unit delays in units of unit delays of a neural network that implements the method, standby multivibrators are applicable. In this case, the waiting multivibrators in the unit of single delays should be started not on the leading, but on the trailing edge of the input pulse. Dynamic synapse blocks can be implemented using memristors (memory resistors) and conventional controlled attenuators. The synapses control unit is implemented by specialized processors, programmable integrated circuits, operating in accordance with the rules discussed above. The method can also be implemented by emulating a two-layer recurrent neural network with controlled synapses on modern computers.

Claims (1)

A method of intelligent processing of information in a neural network, which consists in feeding a multilayer recurrent network with feedbacks closing circuits with a delay time of single images less than the immunity time of the network neurons after their excitation, a signal decomposed into components in a basis matched to the input network layer, with each component converted into a sequence of single images with a repetition rate as a predetermined function of the amplitude of the component, representing the signal and in the form of consecutive sets of single images in accordance with predefined rules for its recognition, taking into account the inverse recognition results, shifts of sets of single images along the layers taking into account the current state of the layers, storing recognition results on network elements, using sequential sets of single images as processing results the output layer of the network after the reverse conversion into the corresponding source signals, characterized in that the network in as a signal, a pre-formed group signal is provided, consisting of a useful signal summed with the first part of the pre-generated noise and the second of the same part of this noise, when transmitting from layer to layer group sets of single images, each of which consists of a signal-noise group and noise groups of single images, copy from signal-noise groups single images related to the useful signal into noise groups corresponding to these groups, form from noise groups of cop and signal-noise groups of single images and process these copies taking into account changes in the cross-sectional shapes of diverging single images and rotations of these images around the directions of their transmission, taking into account the current state of the layers.

Images