Autoencoding system, program, and method

Abstract

To improve energy efficiency of an information processing system.SOLUTION: A system 1 encodes input signals Si1,Si2 and then decodes the respective encoded results as output signals So1,So2. Each of a first input node Nd1, a second input node Nd2, a hub node Nd3, an auxiliary node Nd4, a first output node Nd5, and a second output node Nd6 outputs at least one of a positive signal indicating a state value of the corresponding node and a negative signal indicating a value obtained by inverting a sign of the state value. Each of the first input node Nd1, the second input node Nd2, the hub node Nd3, the auxiliary node Nd4, the first output node Nd5, and the second output node Nd6 changes the state value of the corresponding node so that a time derivative value of the state value of the corresponding node is the same as a value obtained by adding a noise signal received by the corresponding node to a value obtained by subtracting the state value of the corresponding node from an activation value corresponding to a summed value of signals input to the corresponding node.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a system, program, and method for performing self-encoding. [Background technology]

[0002] Conventionally, a configuration for performing autoencoding is known. For example, Non-Patent Document 1 discloses a variational autoencoder of an encoder-decoder model formed as a neural network. The variational autoencoder extracts latent variables of a lower dimension than the input data from the input data, and reproduces output data corresponding to the input data from the latent variables. According to the variational autoencoder, important features of the input data can be extracted as latent variables, so that the generalization performance of the inference model and the reduction in the amount of calculation can be achieved. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Diederik P. Kingma, Max Welling, "Auto-Encoding Variational Bayes", arXiv:1312.6114. [Non-Patent Document 2] Masayo Inoue1, Kunihiko Kaneko, "Cooperative Adaptive Responses in Gene Regulatory Networks with Many Degrees of Freedom", PLOS Computational Biology, April 4, 2013, 9(4): e1003001. Summary of the Invention [Problem to be solved by the invention]

[0004] Living systems (for example, neural networks or cell signaling networks) are formed as dynamic networks (dynamical systems with large degrees of freedom) that are composed of many elements and are constantly changing, and are known to carry out efficient information processing by utilizing noise such as thermal fluctuations.

[0005] On the other hand, in artificially designed information processing systems, noise is suppressed as a disturbance that interferes with information processing, and is often not actively used. In order to realize reliable information processing in an artificial information processing system, it is necessary to increase the signal-to-noise ratio by using a signal with energy higher than the noise to an extent that the effect of the noise can be ignored. As a result, the energy required for information processing by an information processing system is often higher than the energy required for the same level of information processing by a living system.

[0006] The present invention has been made to solve the above-mentioned problems, and has an object to improve the energy efficiency of an information processing system. [Means for solving the problem]

[0007] A system according to one aspect of the present invention encodes a first input signal and a second input signal and then decodes them as a first output signal and a second output signal, respectively. The system includes a first input node, a second input node, a hub node, an auxiliary node, a first output node, and a second output node. The first input node receives the first input signal. The second input node receives the second input signal. The first output node outputs the first output signal. The second output node outputs the second output signal. Each of the first input node, the second input node, the hub node, the auxiliary node, the first output node, and the second output node outputs at least one of a positive signal representing a state value of the node and a negative signal representing a value with the sign inverted of the state value. The first input node outputs the first input node negative signal to each of the first input node and the hub node. The second input node outputs the second input node negative signal to the hub node. The hub node outputs the hub node negative signal to each of the hub node, the first output node, and the second output node. The auxiliary node outputs a positive signal of the auxiliary node to each of the hub node and the auxiliary node. The first output node outputs the positive signal of the first output node to the first output node and outputs the positive signal of the first output node as a first output signal. The second output node outputs the positive signal of the second output node to each of the hub node and the second output node and outputs the positive signal of the second output node as a second output signal. Each of the first input node, the second input node, the hub node, the auxiliary node, the first output node, and the second output node changes the state value of the node so that a value obtained by subtracting the state value of the node from an activation value corresponding to the sum of the signals input to the node and adding a noise signal received by the node becomes equal to a time derivative of the state value of the node.

[0008] A program according to another aspect of the present invention is a system that receives a first input signal and a second input signal and outputs a first output signal and a second output signal, and encodes the first input signal and the second input signal and then decodes them as the first output signal and the second output signal, respectively. The system includes a first input node, a second input node, a hub node, an auxiliary node, a first output node, and a second output node. The first input node receives the first input signal. The second input node receives the second input signal. The first output node outputs the first output signal. The second output node outputs the second output signal. Each of the first input node, the second input node, the hub node, the auxiliary node, the first output node, and the second output node outputs at least one of a positive signal representing a state value of the node and a negative signal representing a value with the sign of the state value inverted. The first input node outputs the negative signal of the first input node to each of the first input node and the hub node. The second input node outputs the negative signal of the second input node to the hub node. The hub node outputs a negative signal of the hub node to each of the hub node, the first output node, and the second output node. The auxiliary node outputs a positive signal of the auxiliary node to each of the hub node and the auxiliary node. The first output node outputs a positive signal of the first output node to the first output node and outputs the positive signal of the first output node as a first output signal. The second output node outputs a positive signal of the second output node to each of the hub node and the second output node and outputs the positive signal of the second output node as a second output signal. The program is executed by a processor included in each of the first input node, the second input node, the hub node, the auxiliary node, the first output node, and the second output node to change the state value of the node so that a value obtained by subtracting the state value of the node from an activation value corresponding to the sum of signals input to the node and adding a noise signal received by the node becomes equal to a time derivative of the state value of the node.

[0009] A method according to another aspect of the present invention is a system that receives a first input signal and a second input signal and outputs a first output signal and a second output signal, the system encoding the first input signal and the second input signal and then decoding the first output signal and the second output signal, respectively. The system includes a first input node, a second input node, a hub node, a supplementary node, a first output node, and a second output node. The first input node receives the first input signal. The second input node receives the second input signal. The first output node outputs the first output signal. The second output node outputs the second output signal. Each of the first input node, the second input node, the hub node, the supplementary node, the first output node, and the second output node outputs at least one of a positive signal representing a state value of the node and a negative signal representing a value with the sign inverted of the state value. The first input node outputs the negative signal of the first input node to each of the first input node and the hub node. The second input node outputs the negative signal of the second input node to the hub node. The hub node outputs a negative hub node signal to each of the hub node, the first output node, and the second output node. The auxiliary node outputs a positive auxiliary node signal to each of the hub node and the auxiliary node. The first output node outputs a positive first output node signal to the first output node and outputs the positive first output node signal as a first output signal. The second output node outputs a positive second output node signal to each of the hub node and the second output node and outputs the positive second output node signal as a second output signal. The method includes a step of each of the first input node, the second input node, the hub node, the auxiliary node, the first output node, and the second output node calculating a sum of signals input to the node, and a step of changing a state value of the node such that a value obtained by subtracting a state value of the node from an activation value corresponding to the sum and adding a noise signal received by the node is equal to a time derivative of the state value of the node. Effect of the Invention

[0010] According to the system, program, and method of the present invention, the energy efficiency of the system can be improved by changing the state value of the first input node, the second input node, the hub node, the auxiliary node, the first output node, and the second output node so that the activation value corresponding to the sum of the signals input to the node minus the state value of the node plus the noise signal received by the node is equal to the time derivative of the state value of the node. [Brief description of the drawings]

[0011] [Figure 1] 1 is a block diagram showing a configuration of an information processing system according to an embodiment. [Diagram 2] This figure also shows 12 time charts each showing the time change in the state value of the hub node corresponding to four input signals and the time change in the state value of the output node when the S / N ratio of the information processing system is 10 dB. [Diagram 3] This figure shows four time charts that show the time changes in the state values ​​of the hub nodes corresponding to four input signals and the time changes in the state values ​​of the output nodes when the noise signal is 0 (when there is no noise signal). [Figure 4] FIG. 13 is a diagram showing a change in the potential energy curve of the first output node when the input signal switches from {0,0} to {1,0}. [Diagram 5] FIG. 13 is a diagram showing a change in the potential energy curve of the second output node when the input signal switches from {0,0} to {1,0}. [Figure 6] 2 is a block diagram showing an example of a hardware configuration of the node in FIG. 1. [Figure 7] 7 is a flowchart showing an example of a flow of processing performed by a processor that executes the auto-encoding program of FIG. 6. [Figure 8] FIG. 13 is a block diagram showing a configuration of an information processing system according to a modified example of the embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] Hereinafter, embodiments will be described in detail with reference to the drawings. In the drawings, the same or corresponding parts are designated by the same reference characters and, in principle, descriptions thereof will not be repeated.

[0013] 1 is a block diagram showing a configuration of an information processing system 1 according to an embodiment. The information processing system 1 may be formed, for example, as an analog integrated circuit included in a computer, or may be formed as an edge computing system including multiple edge nodes. The information processing system 1 may also be formed as a part of a larger-scale information processing system.

[0014] 1, the information processing system 1 encodes an input signal Si1 (first input signal) and an input signal Si2 (second input signal) and then decodes them into an output signal So1 (first output signal) and an output signal So2 (second output signal), respectively. The information processing system 1 includes an input node Nd1 (first input node), an input node Nd2 (second input node), a hub node Nd3, an auxiliary node Nd4, an output node Nd5 (first output node), and an output node Nd6 (second output node).

[0015] The input nodes Nd1 and Nd2 receive input signals Si1 and Si2, respectively. The output nodes Nd5 and Nd6 output output signals So1 and So2, respectively. In the information processing system 1, the two-dimensional input signals Si1 and Si2 are encoded as a one-dimensional state value x3 of the hub node Nd3, and the state value x3 is decoded into two-dimensional output signals So1 and So2.

[0016] Each of the input nodes Nd1, Nd2, the hub node Nd3, the auxiliary node Nd4, and the output nodes Nd5, Nd6 (nodes Nd i (i is a natural number between 1 and 6)) is a node Nd i The state value x i and the state value x iIn FIG. 1, the positive and negative signals are represented by solid and dotted arrows, respectively. Note that the state value x i The initial value of is 0.

[0017] The input node Nd1, the input node Nd2, the hub node Nd3, the auxiliary node Nd4, the output node Nd5, and the output node Nd6 form a network that can be expressed as a directed graph. j From Nd i The signal to be output is S j,i This is expressed as:

[0018] The input node Nd1 is connected to the negative signal S of the input node Nd1 and the hub node Nd3. 1,1 ,S 1,3 The input node Nd2 outputs a negative signal S 2,3 The hub node Nd3 outputs a negative signal S to the hub node Nd3, the output node Nd5, and the output node Nd6. 3,3 ,S 3,5 ,S 3,6 The auxiliary node Nd4 outputs a positive signal S to the hub node Nd3 and the auxiliary node Nd4. 4,3 ,S 4,3 The output node Nd5 outputs a positive signal S 5,5 The output node Nd6 outputs the positive signal S of the output node Nd6 to the hub node Nd3 and the output node Nd6. 6,3 ,S 6,6 and outputs the state value x6 as the output signal So2.

[0019] An adjacency matrix C of a directed graph formed by input node Nd1, input node Nd2, hub node Nd3, auxiliary node Nd4, output node Nd5, and output node Nd6 is expressed, for example, as in the following formula (1).

[0020]

number

[0021] The adjacency matrix C is a matrix of nodes Nd j The row number i represents the node Nd i Each element of the adjacency matrix C is 1 for a positive signal, -1 for a negative signal, and j From Nd i 0 indicates that no signal is being output to the signal S j,i is component C i,j and state value x j It is equal to the product of

[0022] For example, component C 3,4 When Nd4 is 1, a positive signal S 4,3 (=1·x4) is output. Component C 5,3 is -1, a negative signal S 3,5 (=(-1)·x3) is output. Component C 1,2 When the signal is 0, it means that no signal is output from the node Nd2 to the node Nd1 (signal S 2,1 is 0 (=0·x2).

[0023] Each of the node Nd1, the input node Nd2, the hub node Nd3, the auxiliary node Nd4, the output node Nd5, and the output node Nd6 has a sum y i In the information processing system 1, the natural number N is 6. The natural number N may be 7 or more.

[0024]

number

[0025] In equation (2), the signal S i is the node Nd iFor example, signals S1 and S2 correspond to input signals Si1 and Si2, respectively. Since there are no signals input to nodes Nd3 to Nd6 from components other than nodes Nd1 to Nd6, signals S3 and S 4, Each of S5 and S6 is 0.

[0026] Node Nd i is set to node Nd so as to satisfy the following formula (3). i The state value x i The η in equation (3) is the node Nd i The signal-to-noise ratio of the information processing system 1 is, for example, about 10 dB. The left side of the equation (3) represents the state value x i The first term on the right hand side is the time derivative of the sum y i is an output value (activation value) of the activation function for . In formula (3), a sigmoid function with a gain of β (>0) is used as an example of the activation function. The gain β is, for example, 10. Note that the activation function in formula (3) is not limited to the sigmoid function. The activation function in formula (3) may be, for example, a step function, a hyperbolic tangent function, or a ramp (ReLU (Rectified Linear Unit)) function.

[0027]

number

[0028] As shown in equation (3), the node Nd i is the total value y i from the activation value corresponding to the state value x i The value obtained by subtracting i The time derivative of the node Nd i The state value x i Change the value.

[0029] 2 is a diagram showing 12 time charts each showing a time change of the state value x3 of the hub node Nd3 corresponding to four input signals Si1 and Si2 and a time change of the state values ​​x5 and x6 of the output nodes Nd5 and Nd6 when the S / N ratio of the information processing system 1 is 10 dB. In FIG. 2, the first, second, third and fourth columns show the cases where the input signals Si1 and Si2 are {0,0}, {0,1}, {1,0} and {1,1}, respectively.

[0030] 2, when the input signals Si1 and Si2 are {0,0}, {0,1}, {1,0}, and {1,1}, the state values ​​x5 and x6 (output signals So1 and So2) reach an equilibrium state near the values ​​of the input signals Si1 and Si2 over time. That is, the information processing system 1 has an autoencoder function (auto-encoding function).

[0031] 3 is a diagram showing four time charts representing the time change of the state value x3 of the hub node Nd3 corresponding to four input signals Si1 and Si2 and the time change of the state values ​​x5 and x6 of the output nodes Nd5 and Nd6 when the noise signal η is 0 (when there is no noise signal). 3, The time changes of x5 and x6 are included in one time chart.

[0032] As shown in FIG. 3, when the input signals Si1 and Si2 are equal to each other, the state values ​​x5 and x6 (output signals So1 and So2) reach an equilibrium state in the vicinity of the values ​​of the input signals Si1 and Si2, respectively, over time. However, when the input signals Si1 and Si2 are different from each other, the state value x5 reaches an equilibrium state in the vicinity of the opposite value of the input signal Si1. In other words, when there is no noise signal, the information processing system 1 does not function as an autoencoder. This indicates that the noise signal η in the information processing system 1 is actively used as a means for realizing the autoencoding function, and is not a disturbance to be suppressed. The information processing system 1 does not need to increase the energy of the input signals Si1 and Si2 to a degree where the influence of the noise signal η can be ignored, and therefore can improve energy efficiency.

[0033] In the following, a mechanism by which the information processing system 1 functions as an autoencoder even when the input signals Si1 and Si2 are different from each other will be described with reference to FIG. 4 and FIG. 5. FIG. 4 is a diagram showing a change in the potential energy curve of the output node Nd5 when the input signals Si1 and Si2 switch from {0,0} to {1,0}. The potential energy curve is a curve showing the correspondence between the state value and the potential energy U of the corresponding node. A point on the potential energy curve with a slope of 0 is also referred to as a stable point of the state value corresponding to the potential energy curve below. Potential energy is an index showing the instability of the state value, and corresponds to a value obtained by integrating a force with respect to a position in a dynamic system. The state value changes in a direction in which the potential energy becomes lower. In FIG. 4, the state value x3 is shown by a dotted line. In addition, in order to make the change in the state value x5 easier to understand, the state value x5 is shown as a schematic circle rolling on the potential energy curve. The same is true in FIG. 5 described later.

[0034] 4, before the input signals Si1 and Si2 switch from {0,0} to {1,0}, the state value x3 is approximately 0.7, the potential energy curve of the output node Nd5 is PE11, and the state value x5 on the potential energy curve PE11 is stable at 0. In other words, the point on the potential energy curve PE11 corresponding to 0 is the stable point of the state value x5.

[0035] After that, when the input signal is switched from {0,0} to {1,0} and the state value x3 is reduced to 0.32, the potential energy curve changes from PE11 to PE12. The point corresponding to 0 on the potential energy curve PE12 is also the stable point of the state value x5, but the negative signal S output from the hub node Nd3 to the output node Nd5 3,5 The negative signal S increases to -0.32. 3,5 The suppression of the state value x5 by the noise signal η is alleviated. In such a state, the state value x5 may increase slightly due to the noise signal η. When the state value x5 increases, the slope of the potential energy curve PE12 becomes downward to the right, so the state value x5 moves downward to the right on the potential energy curve PE12 and reaches the vicinity of 1. Since the state value x3 is not affected by the state value x5, the potential energy curve does not change from PE12. Since the slope of the potential energy curve PE12 when the state value x5 is near 1 is 0 or negative (downward), the state value x5 converges to the vicinity of 1. In other words, the state value x5 reaches an equilibrium state near 1.

[0036] FIG. 5 is a diagram showing the change in the potential energy curve of the output node Nd6 when the input signals Si1 and Si2 switch from {0,0} to {1,0}. With reference to FIG. 5, it is assumed that before the input signals Si1 and Si2 switch from {0,0} to {1,0}, the state value x3 is near 0.7, the potential energy curve of the output node Nd6 is PE21, and the state value x6 is stable at 0 on the potential energy curve PE21. In other words, the point on the potential energy curve PE21 corresponding to 0 is the stable point of the state value x6. Note that the forces (negative signals S 3,5 ,S 3,6 ) is always the same, so the potential energy curve PE21 is equal to the potential energy curve PE11 in FIG.

[0037] After that, when the input signal is switched from {0,0} to {1,0}, the state value x3 decreases to 0.32, and the potential energy curve changes from PE21 to PE22. The point corresponding to 0 on the potential energy curve PE22 is also the stable point of the state value x6, but the negative signal output from the hub node Nd3 to the output node Nd6 increases to -0.32, and the negative signal S 3,6 In this state, the state value x6 may be slightly increased by the noise signal η. When the state value x6 increases, the slope of the potential energy curve PE22 becomes downward to the right, and the positive signal S output from the output node Nd6 to the hub node Nd3 decreases. 6,3 increases, so the state value x3 increases. As a result, for example, when the state value x3 increases from 0.32, the potential energy curve returns from PE22 to PE21 before the state value x6 reaches a value near 0.7. As a result, the state value x6 does not increase from near 0. That is, the state value x6 reaches an equilibrium state near 0.

[0038] When the input signals Si1 and Si2 switch from {0,0} to {0,1}, the state value x3 decreases immediately after the switch, causing both state values ​​x5 and x6 to increase to near 1. After that, the state value x5 changes according to a potential energy curve obtained by linearly inverting the potential energy curve PE12 shown in Fig. 4, and converges to near 0. The state value x6 changes according to a potential energy curve obtained by linearly inverting the potential energy curve PE22 shown in Fig. 5, and converges to near 1.

[0039] In the information processing system 1, a global optimum solution is searched for using a noise signal, as in the annealing method. In addition, in the information processing system 1, the input signals Si1 and Si2 are encoded into time-series data of a state value x3 (temporal coding or dynamic coding). Furthermore, in the information processing system 1, the state value x3 is dynamically changed using a noise signal η, thereby dynamically changing each of the potential energy curves of the output nodes Nd5 and Nd6.

[0040] For example, a genetic algorithm (see Non-Patent Document 2) can be used to search for the network of the information processing system 1 shown in FIG. 1. Specifically, a plurality of adjacency matrices C1 (for example, 100) are generated by randomly determining each element between -1, 0, and 1. An evaluation index of the auto-encoding function of the network corresponding to each of the plurality of adjacency matrices C1 is calculated, and the plurality of adjacency matrices C1 are ranked in descending order of the evaluation index. Among the plurality of adjacency matrices C1, a plurality of adjacency matrices C2 having a predetermined rank (for example, 25th place) or higher are identified. A predetermined number (for example, 3) of adjacency matrices C3 are generated from each of the plurality of adjacency matrices C2. The plurality of adjacency matrices C3 are adjacency matrices in which each element of the corresponding adjacency matrix C2 is replaced with a value different from the element among -1, 0, and 1 with a predetermined probability (for example, 1% to 10%). The processes subsequent to the calculation of the evaluation index of the auto-encoding function are repeated for the plurality of adjacency matrices C3 generated from the plurality of adjacency matrices C2.

[0041] The evaluation index of the autoencoding function is a value obtained by adding the sum of the absolute values ​​of the differences between the output signals So1, So2 and the input signals Si1, Si2 when the input signals Si1, Si2 are returned to {0,0} to the sum of the absolute values ​​of the differences between the output signals So1, So2 and {0,0} when the input signals Si1, Si2 are returned to {0,0} when multiple input signals Si1, Si2 ({0,0}, {0,1}, {1,0}, {1,1}) are input to a network corresponding to each of multiple adjacency matrices C1. Note that, in order to reflect the vibration state of the output signals So1, So2, etc. in the evaluation index, the average value of the output signals So1, So2 from when half a predetermined waiting time has elapsed since the input signals Si1, Si2 are input to the network until the waiting time has elapsed may be used as the output signals So1, So2.

[0042] In the following, a case where the information processing system 1 is formed as an edge computing system will be described. In this case, the node Nd i forms an edge node. Figure 6 shows the node Nd i 6 is a block diagram showing an example of a hardware configuration of a node Nd i includes a storage 101 , a processor 102 , a memory 103 , a network interface 104 , and an input / output unit 105 .

[0043] The storage 101 is a non-volatile storage device, and includes, for example, a hard disk or an external storage medium. The storage 101 stores, for example, an operating system program (not shown) and a self-encoding program Pg1.

[0044] The processor 102 executes various programs stored in the storage 101 to realize the functions of the information processing system 1. The processor 102 includes, for example, a central processing unit (CPU) or a graphics processing unit (GPU). iThe processor 102 executes the auto-encoding program Pg1, thereby realizing the auto-encoding function of the information processing system 1. The processor 102 transmits data to an external device connected to the Internet via the network interface 104, and receives data from the external device.

[0045] The memory 103 is a volatile storage device, and includes, for example, a random access memory (RAM). The memory 103 functions as a working memory, and temporarily stores various data necessary for executing a program.

[0046] The input / output unit 105 receives operations from a user and outputs to the user the processing results of a program corresponding to the operations. The input / output unit 105 includes, for example, a touch panel, a display, a lamp, and a speaker.

[0047] 7 is a flowchart showing an example of the flow of processing performed by the processor 102 that executes the self-encoding program Pg1 of FIG. i The step is called, for example, at each sampling time by a main routine (not shown) which controls the above steps in an integrated manner. In the following, a step will be simply abbreviated as S.

[0048] As shown in FIG. 7, the processor 102 selects a node Nd i The sum of the signals input to y i Then, the process proceeds to step S102. i from the activation value corresponding to the state value x i The value obtained by subtracting i The state value x is set to be equal to the time derivative of i and returns the process to the main routine.

[0049] 8 is a block diagram showing a configuration of an information processing system 1A according to a modified example of the embodiment. The configuration of the information processing system 1A is such that an output signal So3 representing a state value x3 is output from a hub node Nd3 to the outside of the information processing system 1A in the information processing system 1 of FIG. 1. The other configuration of the information processing system 1 is similar to that of the information processing system 1, and therefore the description of the similar configuration will not be repeated. The state value x3 is a feature of the input signals Si1 and Si2 obtained by reducing the dimensions of the input signals Si1 and Si2. According to the information processing system 1A, the feature of the input signals Si1 and Si2 can be acquired outside the information processing system 1.

[0050] As described above, according to the information processing system, the program, and the method according to the embodiment and the modified example, it is possible to improve the energy efficiency of the information processing system.

[0051] The embodiments disclosed herein should be considered to be illustrative and not restrictive in all respects. The scope of the present invention is defined by the claims, not the above description, and is intended to include all modifications within the meaning and scope of the claims. [Explanation of symbols]

[0052] 1,1A Information processing system, 101 Storage, 102 Processor, 103 Memory, 104 Network interface, 105 Input / output unit, Nd1 to Nd6, Nd i ,Nd j Node, Pg1 self-encoding program, Si1,Si2 input signals, So1~So3 output signals.

Claims

1. 1. A system for encoding first and second input signals and then decoding them into first and second output signals, respectively, comprising: a first input node for receiving the first input signal; a second input node for receiving the second input signal; a hub node; an auxiliary node; a first output node that outputs the first output signal; a second output node that outputs the second output signal; each of the first input node, the second input node, the hub node, the auxiliary node, the first output node, and the second output node outputs at least one of a positive signal representing a state value of the node and a negative signal representing a value obtained by inverting the sign of the state value; the first input node outputs a negative signal of the first input node to each of the first input node and the hub node; the second input node outputs a negative signal of the second input node to the hub node; the hub node outputs a negative signal of the hub node to each of the hub node, the first output node, and the second output node; the auxiliary node outputs a positive signal of the auxiliary node to each of the hub node and the auxiliary node; the first output node outputs a positive signal of the first output node to the second output node and outputs the positive signal of the first output node as the first output signal; the second output node outputs a positive signal of the second output node to each of the hub node and the second output node, and outputs the positive signal of the second output node as the second output signal; a system in which each of the first input node, the second input node, the hub node, the auxiliary node, the first output node, and the second output node changes the state value of the node so that a value obtained by subtracting the state value of the node from an activation value corresponding to the sum of signals input to the node and adding a noise signal received by the node is equal to a time derivative of the state value of the node.

2. The system of claim 1 , wherein the hub node outputs the state value of the hub node external to the system.

3. each of the first input node, the second input node, the hub node, the auxiliary node, the first output node, and the second output node is formed as an edge node; The system according to claim 1 or 2, wherein the edge node includes a processor, a memory, and a communication unit.

4. A system that receives a first input signal and a second input signal and outputs a first output signal and a second output signal, comprising: a program that encodes the first input signal and the second input signal and then decodes the encoded first input signal and the second input signal into the first output signal and the second output signal, The system comprises: a first input node for receiving the first input signal; a second input node for receiving the second input signal; a hub node; an auxiliary node; a first output node that outputs the first output signal; a second output node that outputs the second output signal; each of the first input node, the second input node, the hub node, the auxiliary node, the first output node, and the second output node outputs at least one of a positive signal representing a state value of the node and a negative signal representing a value obtained by inverting the sign of the state value; the first input node outputs a negative signal of the first input node to each of the first input node and the hub node; the second input node outputs a negative signal of the second input node to the hub node; the hub node outputs a negative signal of the hub node to each of the hub node, the first output node, and the second output node; the auxiliary node outputs a positive signal of the auxiliary node to each of the hub node and the auxiliary node; the first output node outputs a positive signal of the first output node to the second output node and outputs the positive signal of the first output node as the first output signal; the second output node outputs a positive signal of the second output node to each of the hub node and the second output node, and outputs the positive signal of the second output node as the second output signal; The program, when executed by a processor included in each of the first input node, the second input node, the hub node, the auxiliary node, the first output node, and the second output node, changes the state value of the node so that the value obtained by subtracting the state value of the node from the activation value corresponding to the total value of the signals input to the node and adding the noise signal received by the node becomes equal to the time derivative of the state value of the node.

5. 1. A method for a system that receives first and second input signals and outputs first and second output signals, comprising encoding the first and second input signals and then decoding the first and second output signals, respectively, the method comprising: The system comprises: a first input node for receiving the first input signal; a second input node for receiving the second input signal; a hub node; an auxiliary node; a first output node that outputs the first output signal; a second output node that outputs the second output signal; each of the first input node, the second input node, the hub node, the auxiliary node, the first output node, and the second output node outputs at least one of a positive signal representing a state value of the node and a negative signal representing a value obtained by inverting the sign of the state value; the first input node outputs a negative signal of the first input node to each of the first input node and the hub node; the second input node outputs a negative signal of the second input node to the hub node; the hub node outputs a negative signal of the hub node to each of the hub node, the first output node, and the second output node; the auxiliary node outputs a positive signal of the auxiliary node to each of the hub node and the auxiliary node; the first output node outputs a positive signal of the first output node to the second output node and outputs the positive signal of the first output node as the first output signal; the second output node outputs a positive signal of the second output node to each of the hub node and the second output node, and outputs the positive signal of the second output node as the second output signal; The method comprises: each of the first input node, the second input node, the hub node, the auxiliary node, the first output node, and the second output node; calculating a sum of signals input to the node; and changing the state value of the node so that the value obtained by subtracting the state value of the node from the activation value corresponding to the sum plus a noise signal received by the node is equal to the time derivative of the state value of the node.

Images