Coupling strength update device, coupling strength update method, coupling strength update program, and spiking neural network system

Abstract

A coupling strength update device is configured to comprise: a coupling strength management unit (3) that manages the coupling strength between a first neuron, which is a certain neuron among a plurality of neurons included in a spiking neural network, and a second neuron, which is a neuron that acquires a spike signal outputted from the first neuron upon firing of the first neuron; an elapsed time management unit (1) that manages the elapsed time since the spike signal was outputted from the first neuron; and a firing rate management unit (2) that manages the firing rate, which is a rate at which the first neuron fires. In addition, when the second neuron fires, the coupling strength management unit (3) of the coupling strength update device updates the coupling strength between the first neuron and the second neuron on the basis of the elapsed time managed by the elapsed time management unit (1) and the firing rate managed by the firing rate management unit (2).

Description

Connection strength update device, connection strength update method, connection strength update program, and spiking neural network system

[0001] This disclosure relates to a coupling strength update device, a coupling strength update method, a coupling strength update program, and a spiking neural network system.

[0002] One example of a third-generation neural network is the spiking neural network. A spiking neural network has multiple neurons called spiking neurons. Each neuron has a value called membrane potential. The membrane potential of each neuron rises, for example, when a spike signal (described later) is applied from an upstream neuron. When the membrane potential of each neuron exceeds a threshold, the neuron fires and outputs a spike signal indicating that it has fired to, for example, a downstream neuron. There is a connection strength update device that updates the connection strength between a certain neuron among the multiple neurons (hereinafter referred to as the "first neuron") and a neuron (hereinafter referred to as the "second neuron") that acquires the spike signal output from the first neuron when the first neuron fires. As such a connection strength update device, for example, Non-Patent Document 1 discloses a connection strength update device that, when the second neuron fires, compares the firing frequency (the frequency at which the first neuron fires) with a certain threshold, and updates the connection strength between the first neuron and the second neuron based on the comparison result between the firing frequency and the threshold. Specifically, the connection strength update device updates the connection strength between the first and second neurons to increase if the firing frequency of the first neuron is above a certain threshold when the second neuron fires, and to decrease the connection strength between the first and second neurons if the firing frequency of the first neuron is below the threshold.

[0003] Diehl, Peter U., and Matthew Cook. "Unsupervised learning of digit recognition using spike-timing-dependent plasticity." Frontiers in computational neuroscience 9 (2015): 99.

[0004] The connection strength updating device disclosed in Non-Patent Document 1 updates the connection strength between the first neuron and the second neuron based solely on the comparison result between the firing frequency of the first neuron and a threshold. Therefore, the connection strength updating device disclosed in Non-Patent Document 1 has the problem that it may update the connection strength even when it is not desirable to update it. One example of a configuration where it is not desirable to update the connection strength is when the firing frequency of the first neuron is lower than the threshold, and the connection strength between the first neuron and the second neuron is repeatedly updated so that it weakens each time the second neuron fires. As a result, the connection strength between the first neuron and the second neuron may become zero, and the learned content may be erased.

[0005] This disclosure was made to solve the above-mentioned problems and aims to provide a bond strength updating device that can update bond strength only when it is necessary.

[0006] The connection strength update device according to this disclosure includes a connection strength management unit that manages the connection strength between a first neuron, which is a neuron among a plurality of neurons included in a spiking neural network, and a second neuron, which is a neuron that acquires a spike signal output from the first neuron when the first neuron fires; an elapsed time management unit that manages the elapsed time since the spike signal was output from the first neuron; and a firing frequency management unit that manages the firing frequency, which is the frequency at which the first neuron fires. When the second neuron fires, the connection strength management unit updates the connection strength between the first neuron and the second neuron based on the elapsed time managed by the elapsed time management unit and the firing frequency managed by the firing frequency management unit.

[0007] According to this disclosure, the bond strength can be updated only when it is necessary to update it.

[0008] This is a configuration diagram showing a connection strength update device according to Embodiment 1. This is a hardware configuration diagram showing the hardware of the connection strength update device according to Embodiment 1. This is a computer hardware configuration diagram when the connection strength update device is implemented by software or firmware, etc. Figure 4A is a flowchart showing the connection strength update method, which is the processing procedure of the connection strength update device, and Figure 4B is a flowchart showing an example of the connection strength w update process by the connection strength management unit 3. This is an explanatory diagram showing an example of a spiking neural network to which the connection strength update device shown in Figure 1 is applied. This is an explanatory diagram showing an example of a spiking neural network that solves a maze problem. This is a configuration diagram showing a connection strength update device according to Embodiment 2. This is a hardware configuration diagram showing the hardware of the connection strength update device according to Embodiment 2. This is a configuration diagram showing a connection strength update device according to Embodiment 3. This is a hardware configuration diagram showing the hardware of the connection strength update device according to Embodiment 3. This is an explanatory diagram showing an example of selection information given to the connection strength management unit 6. This is a configuration diagram showing a connection strength update device according to Embodiment 4. This is a hardware configuration diagram showing the hardware of the connection strength update device according to Embodiment 4. This is an explanatory diagram showing that the elapsed time management unit 1 and the firing frequency management unit 2 are each provided for the connection unit between two neurons. This is an explanatory diagram showing that the elapsed time management unit 1 and the firing frequency management unit 2 are each provided on a neuron-by-neuron basis.

[0009] To provide a more detailed explanation of this disclosure, the forms for implementing this disclosure will be described below with reference to the attached drawings.

[0010] Embodiment 1. Figure 1 is a configuration diagram showing a bond strength updating device according to Embodiment 1. Figure 2 is a hardware configuration diagram showing the hardware of the bond strength updating device according to Embodiment 1. The bond strength updating device shown in Figure 1 comprises an elapsed time management unit 1, an ignition frequency management unit 2, a bond strength management unit 3, and a bond strength maintenance unit 4.

[0011] The spiking neural network applied by the bonding strength updating device shown in FIG. 1 has a plurality of neurons called spiking neurons. Each neuron has a value called a membrane potential. The membrane potential of each neuron, for example, rises when a spike signal is given from an upstream neuron. When the membrane potential of each neuron becomes equal to or higher than a threshold value, each neuron fires and outputs a spike signal indicating that it has fired, for example, to a downstream neuron.

[0012] FIG. 5 is an explanatory diagram showing an example of the spiking neural network to which the bonding strength updating device shown in FIG. 1 is applied. In FIG. 5, ○ indicates a neuron. The symbol inside ○ is a symbol for identifying the neuron. In the example of FIG. 5, the number of neurons that the spiking neural network has is 11. However, the spiking neural network may have the necessary number of neurons according to the application, and the number of neurons is not limited to 11. In the example of FIG. 5, in the connection relationship from neuron N １ to neuron N ４ , in the connection relationship, N １ is the first neuron and N ４ is the second neuron. Also, in the connection relationship from neuron N ４ to neuron N ８ , in the connection relationship, N ４ is the first neuron and N ８ is the second neuron. The same applies to the connection relationships between other neurons. Therefore, in the example of FIG. 5, focusing on the connection from the first layer to the second layer, neurons N １ , N ２ , N ３ are the first neurons, and neurons N ４ , N ５ , N ６ , N ７ are the second neurons. Focusing on the connection from the second layer to the third layer, N ４ , N ５ , N ６ , N ７ are the first neurons, and neuron N ８, N ９ , N １０ However, this is the second neuron.

[0013] The elapsed time management unit 1 is implemented, for example, by the elapsed time management circuit 11 shown in Figure 2. The elapsed time management unit 1 receives a spike signal SO output from the first neuron when the first neuron fires. １ The elapsed time management unit 1 manages the elapsed time since the first neuron fired. The timing of the first neuron firing is indicated by the spike signal SO １ Since it is given by, the elapsed time management unit 1 receives a spike signal SO from the first neuron. １ The elapsed time since the output may be managed. More specifically, the elapsed time management unit 1 manages the value H corresponding to the above elapsed time. Ｂ (t) is managed. Value H Ｂ (t) will be discussed later.

[0014] The firing frequency management unit 2 is implemented, for example, by the firing frequency management circuit 12 shown in Figure 2. The firing frequency management unit 2 receives a spike signal SO output from the first neuron when the first neuron fires. １ The ignition frequency management unit 2 is given the spike signal SO per unit time. １ Based on the number of outputs, the firing frequency of the first neuron is managed. Here, the firing frequency can be, for example, the firing frequency in a predetermined time interval. The time interval is a value R corresponding to the firing frequency. Ｂ (t) etc. can be determined by the time constant in the formula for calculating the ignition frequency. More specifically, the ignition frequency management unit 2 uses a value R corresponding to the ignition frequency. Ｂ (t) is managed. Value R Ｂ (t) may be determined, for example, by the time constant of exponential decay, assuming exponential decay. Value R Ｂ (t) will be discussed later.

[0015] The coupling strength management unit 3 is implemented, for example, by the coupling strength management circuit 13 shown in Figure 2. The coupling strength management unit 3 receives a spike signal SO output from the second neuron when the second neuron fires. ２The connection strength management unit 3 manages the connection strength between the first neuron and the second neuron. When the second neuron fires, the connection strength management unit 3 updates the connection strength between the first neuron and the second neuron based on the elapsed time managed by the elapsed time management unit 1 and the firing frequency managed by the firing frequency management unit 2. In the connection strength update device shown in Figure 1, the connection strength management unit 3 updates the connection strength from the first neuron to the second neuron as the connection strength between the first neuron and the second neuron. The connection strength management unit 3 outputs the updated connection strength to the connection strength holding unit 4.

[0016] Specifically, the bond strength management unit 3 determines the update range of the bond strength based on the elapsed time managed by the elapsed time management unit 1. For example, the bond strength management unit 3 may determine the update range so that the longer the elapsed time, the smaller the update range of the bond strength, and so that the shorter the elapsed time, the larger the update range of the bond strength. As a result, when the elapsed time is sufficiently long, the update range of the bond strength becomes a value very close to zero, and when the elapsed time is sufficiently short, the update range of the bond strength becomes a value close to the maximum value. Here, the maximum value refers to the amount of change in bond strength determined based on an indicator other than elapsed time. In this example, the amount of change is the amount of change based on the firing frequency. In this way, when the elapsed time is sufficiently long, the bond strength is maintained regardless of the amount of change based on the firing frequency, which will be described later, and when the elapsed time is short, the bond strength change occurs in accordance with the amount of change based on the firing frequency. That is, the responsiveness of the change in bond strength to the amount of change in bond strength based on the firing frequency can be adjusted according to the elapsed time. Responsiveness can be described as the degree of response that indicates how much the final change in bond strength responds to the change determined based on the firing frequency. Adjusting the update range of bond strength based on elapsed time can be considered as adjusting the "coefficient" in the equation for change in bond strength based on firing frequency. Here, this coefficient can be considered as the reaction coefficient of firing frequency to the change in bond strength. For example, no matter how large the change determined based on firing frequency is, if the update range of bond strength based on elapsed time is zero, the reaction coefficient of firing frequency will also be zero, and the final change will also be zero. On the other hand, even if the change determined based on firing frequency is small, if the update range of bond strength based on elapsed time is 100%, the reaction coefficient of firing frequency will also be 1, and the final change will be equal to the change determined based on firing frequency. Determining such an update range based on elapsed time can also be considered an adjustment of the change in bond strength as a higher-level concept. As for the method of determining the update range by the bond strength management unit 3, for example, it is also possible to determine the update range using multiple different thresholds.In other words, the adjustment of the update width does not need to be done using a continuous value like the reaction coefficient described above. For example, a method can be used to adjust the update width in stages, such as setting the update width to 0% if the elapsed time is equal to or greater than a first threshold, and setting the update width to 10% if the elapsed time is less than the first threshold but equal to or greater than a second threshold. The number of thresholds is not particularly limited and may be one or more. Furthermore, as a method for determining the update width by the bond strength management unit 3, for example, it is possible to determine the update width corresponding to the elapsed time by referring to a table that stores the correspondence between elapsed time and update width.

[0017] The connection strength retention unit 4 is implemented, for example, by the connection strength retention circuit 14 shown in Figure 2. The connection strength retention unit 4 retains the connection strength before and after updating by the connection strength management unit 3. When configuring a spiking neural network system equipped with the connection strength update device shown in Figure 1, the spiking neural network system only needs to include a first neuron and a second neuron in addition to the connection strength update device shown in Figure 1.

[0018] In Figure 1, it is assumed that the components of the bond strength updating device—the elapsed time management unit 1, the ignition frequency management unit 2, the bond strength management unit 3, and the bond strength retention unit 4—are each implemented by dedicated hardware as shown in Figure 2. Specifically, the bond strength updating device is assumed to be implemented by an elapsed time management circuit 11, an ignition frequency management circuit 12, a bond strength management circuit 13, and a bond strength retention circuit 14. The coupling strength maintenance circuit 14 may include, for example, non-volatile or volatile semiconductor memory such as RAM (Random Access Memory), ROM (Read Only Memory), flash memory, EPROM (Erasable Programmable Read Only Memory), EEPROM (Electrically Erasable Programmable Read Only Memory), magnetic disks, flexible disks, optical disks, compact disks, minidiscs, or DVDs (Digital Versatile Discs). Each of the elapsed time management circuit 11, the firing frequency management circuit 12, and the coupling strength management circuit 13 can be, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a combination thereof.

[0019] The components of a bond strength update device are not limited to those implemented by dedicated hardware; the bond strength update device may also be implemented by software, firmware, or a combination of software and firmware. The software or firmware is stored as a program in the memory of a computer. A computer refers to the hardware that executes the program, and includes, for example, a CPU (Central Processing Unit), GPU (Graphics Processing Unit), central processing unit, processing unit, arithmetic unit, microprocessor, microcomputer, processor, or DSP (Digital Signal Processor).

[0020] Figure 3 is a hardware configuration diagram of a computer when the bond strength update device is implemented by software or firmware. When the bond strength update device is implemented by software or firmware, the bond strength maintenance unit 4 is configured on the computer's memory 21. A bond strength update program that causes the computer to execute the respective processing procedures in the elapsed time management unit 1, the ignition frequency management unit 2, and the bond strength management unit 3 is stored in the memory 21. The computer's processor 22 then executes the bond strength update program stored in the memory 21.

[0021] Furthermore, Figure 2 shows an example in which each component of the bond strength updating device is implemented by dedicated hardware, and Figure 3 shows an example in which the bond strength updating device is implemented by software or firmware. However, this is only one example, and some components of the bond strength updating device may be implemented by dedicated hardware, while the remaining components may be implemented by software or firmware.

[0022] Neural networks have a structure that mimics the neural circuits of living organisms and can achieve various tasks such as image recognition or voice recognition with high accuracy. On the other hand, training and running neural networks require a lot of power. Among neural networks, spiking neural networks, for example, have characteristics that are closer to those of neural circuits and operate with low power consumption when implemented in hardware. Compared to general neural networks, spiking neural networks can perform tasks with about 1 / 100 to 1 / 1000 of the power consumption.

[0023] In the neural circuits of living organisms, for example, place cells in the hippocampus are thought to represent location using a mechanism similar to one-hot coding, where only specific cells fire at specific locations. By using a group of place neurons that function similarly to these place cells and a group of behavioral neurons that determine the direction of movement, and by connecting the place neurons to the behavioral neurons, it is possible to construct a network that can remember appropriate actions at specific locations. For example, a network that can solve the maze task shown in Figure 6 can be constructed. Figure 6 is an explanatory diagram showing an example of a spiking neural network that solves a maze task.

[0024] For example, consider a scenario where the network is trained to sequentially learn the actions of an agent in order to have the agent perform the same action in multiple locations. In such a case, the connection strength update device disclosed in Non-Patent Document 1 may cause the connection strength between the other location neurons and the action neuron to weaken and the learned content to be erased because the firing frequency of other location neurons that provide spike signals to the action neuron is 0. In the maze shown in Figure 6, for example, in squares (1) and (3), the agent needs to move to the right, so the connection strength from neuron (1) to neuron (r) and the connection strength from neuron (3) to neuron (r) must be strengthened. However, if neuron (r) fires while the agent is in square (3), and the firing frequency of neuron (1) is 0, the connection strength from neuron (1) to neuron (r) may be weakened. As a result, the learned content may be erased, and the agent may not be able to perform the action of moving to the right while in square (3).

[0025] Next, the operation of the connection strength update device shown in Figure 1 will be explained. Figures 4A and 4B are flowcharts showing the connection strength update method, which is the processing procedure of the connection strength update device. The neural network to which the connection strength update device shown in Figure 1 is applied is, for example, as shown in Figure 5, a neural network with N neurons. １ ~N １１ It has. Here, for the sake of explanation, we will refer to the first neuron as neuron N. ２ Then, the second neuron, neuron N ５ Focusing on this, the connection strength renewal device shown in Figure 1 is used in neuron N ２ and neuron N ５ This section describes the process of updating the bond strength with the other party.

[0026] Figure 4A is a flowchart showing the update process of the bond strength w by the bond strength management unit 3. Steps ST1 and ST2 in Figure 4A are also called the update preparation process. In the update preparation process, the spike signal SO １The trigger updates the firing frequency and time elapsed since firing of the first neuron. In the update process, the spike signal SO ２ The connection strength is updated as a trigger. The elapsed time management unit 1 contains neuron N, which is the first neuron. ２ When it fires, neuron N ２ The spike signal SO output from １ The elapsed time management unit 1 controls the neuron N ２ From spike signal SO １ The elapsed time T since the output is managed (step ST1 in Figure 4A). Specifically, the elapsed time management unit 1 manages neuron N ２ As a value corresponding to the elapsed time since ignition, H is a value corresponding to the "update range" of the bond strength based on the elapsed time. Ｂ (t) is managed. In Embodiment 1, the elapsed time management unit 1 has a timer function, and the timer value CT of the elapsed time management unit 1 is a variable that holds a value of 0 or more and is initially set to 0. The timer value CT decreases as time passes, and the number of neurons N ２ From spike signal SO １ Each time an output is generated, a certain value is added. The magnitude of the timer value CT indicates how much time has passed since the first neuron fired. Neuron N ２ From spike signal SO １ The update interval, based on the elapsed time T since the output was generated, is, for example, H as shown in equation (1) below. Ｂ (t) is expressed by H Ｂ (t) is the spike signal SO １ The spike signal SO is large when not much time has passed since the output. １ The longer the time elapsed since the output, the smaller the value becomes, and if the decrease due to the timer becomes greater than the increase due to ignition, it will eventually become zero. The elapsed time management unit 1 uses a separately prepared timer value CT to H Ｂ (t) can be managed, or the timer value CT itself can be H Ｂ It may be managed as (t).

[0027] In equation (1), τ Ｂ１ This is the time constant.

[0028] The firing frequency management unit 2 contains neuron N, which is the first neuron. ２ When it fires, neuron N ２ The spike signal SO output from １ The ignition frequency management unit 2 is given the spike signal SO per unit time. １ Based on the number of outputs, neuron N ２ The firing frequency, which is the frequency at which the neurons fire, is managed (step ST2 in Figure 4A). Specifically, the firing frequency management unit 2 manages the firing frequency of neurons N ２ The value R, which corresponds to the firing frequency, determines the "amount of change" in bond strength based on that firing frequency. Ｂ Manage (t). R Ｂ (t) decreases over time, and neuron N ２ From spike signal SO １ Each time it is output, a certain value is added. Ｂ (t) is neuron N ２ The higher the firing frequency, the larger the value becomes, and in a given time interval, neuron N ２ From spike signal SO １ If no output is produced at all, the value becomes 0. Neuron N ２ The value R corresponds to the firing frequency. Ｂ (t) can be expressed, for example, as shown in equation (2) below.

[0029] In equation (2), τ Ｂ２ This is the time constant.

[0030] The connection strength management unit 3 contains neuron N, which is the second neuron. ５ When it fires, neuron N ５ The spike signal SO output from ２ The connection strength management unit 3 obtains the connection strength w held by the connection strength holding unit 4. The connection strength w held by the connection strength holding unit 4 is the first neuron, neuron N. ２ and the second neuron, neuron N ５is the current coupling strength w with it. The coupling strength management unit 3 uses the spike signal SO ２ When it is given, that is, at the timing when the second neuron fires, the value H corresponding to the elapsed time T Ｂ (t) and the value R corresponding to the firing frequency Ｂ (t), the coupling strength w between the first neuron and the second neuron is updated (step ST3 in FIG. 4A). The coupling strength management unit 3 causes the coupling strength holding unit 4 to hold the updated coupling strength w.

[0031] Hereinafter, the update process of the coupling strength w by the coupling strength management unit 3 (step ST3 in FIG. 4A) will be specifically described. FIG. 4B is a flowchart showing an example of the update process of the coupling strength w by the coupling strength management unit 3. The coupling strength management unit 3 acquires the coupling strength w held by the coupling strength holding unit 4 (step ST11 in FIG. 4B).

[0032] The coupling strength management unit 3 determines the change amount Δw of the coupling strength w according to the value H corresponding to the elapsed time T Ｂ (t) and the value R corresponding to the firing frequency Ｂ (t) as shown in the following formula (3) (step ST12 in FIG. 4B). Then, the coupling strength management unit 3 updates the coupling strength w using the change amount Δw as shown in the following formula (4) (step ST13 in FIG. 4B). Δw = (A Ａ +R Ｂ (t))H Ｂ (t) (3) w ← w + Δw (4) In formula (3), A Ａ is the threshold for the value R Ｂ (t) based on the firing frequency, and a negative value is set. The threshold A Ａ can adjust the sign of the change amount Δw, that is, whether to strengthen or weaken the coupling strength w.

[0033] The longer the elapsed time T from the firing of the first neuron as the source of the connection, the smaller H Ｂ (t) becomes. Therefore, as shown in formula (3), the update width (the magnitude of the absolute value) of the change amount Δw is determined to become smaller. For example, if the elapsed time T is longer than a predetermined time, H ＢSince (t) is set to zero or a value close to zero, the change amount Δw is also determined to be zero or a value close to zero. A value close to zero means a value such that the change in the coupling strength w due to update is less than, for example, 1%. However, a value close to zero is not limited to a value such that the change in the coupling strength w due to update is less than 1%. The predetermined time is a time determined in relation to a value close to zero. The value H determined based on the elapsed time T Ｂ If (t) is zero or a value close to zero, the value R corresponding to the firing frequency Ｂ Even if (t) is large, the final change amount Δw will be zero or a value close to zero. As a result, the coupling strength w will be maintained.

[0034] Incidentally, the coupling strength management unit 3 is the neuron N ２ The value R corresponding to the firing frequency of Ｂ If (t) is Ａ higher than the absolute value of A, the change amount Δw will be a positive value, and the coupling strength w between the neuron N ２ and the neuron N ５ will be updated so that the coupling strength w becomes stronger. The coupling strength management unit 3 is the neuron N ２ The value R corresponding to the firing frequency of Ｂ If (t) is Ａ lower than the absolute value of A, the change amount Δw will be a negative value, and the coupling strength w between the neuron N ２ and the neuron N ５ will be updated so that the coupling strength w becomes weaker. However, as described above, depending on the value H Ｂ (t), the change amount Δw may become zero, and in such a case, the coupling strength w will be maintained regardless of the magnitude of R Ｂ (t).

[0035] In the above embodiment 1, the connection strength update device is configured to include a connection strength management unit 3 that manages the connection strength between a first neuron, which is one of the multiple neurons included in the spiking neural network, and a second neuron, which is a neuron that acquires the spike signal output from the first neuron when the first neuron fires; an elapsed time management unit 1 that manages the elapsed time since the spike signal was output from the first neuron; and a firing frequency management unit 2 that manages the firing frequency, which is the frequency at which the first neuron fires. Furthermore, when the second neuron fires, the connection strength management unit 3 of the connection strength update device updates the connection strength between the first neuron and the second neuron based on the elapsed time managed by the elapsed time management unit 1 and the firing frequency managed by the firing frequency management unit 2.Therefore, the connection strength update device can update the connection strength only when an update of the connection strength is necessary.In addition, the connection strength update device can prevent situations in which the connection strength is updated when an update of the connection strength is unnecessary. Furthermore, the bond strength updating device can prevent situations where the bond strength is updated based solely on a first indicator, such as the ignition frequency, even when updating the bond strength is necessary. Instead, it can appropriately update the bond strength based on multiple indicators, such as the passage of time and the ignition frequency.

[0036] In the connection strength updating device shown in Figure 1, the connection strength updating device is configured such that if the elapsed time managed by the elapsed time management unit 1 is longer than a predetermined time, the connection strength management unit 3 maintains the connection strength between the first neuron and the second neuron. Therefore, the connection strength updating device can suppress the decrease in connection strength between the first neuron and the second neuron when the firing frequency of the first neuron is low.

[0037] In the connection strength update device shown in Figure 1, the connection strength management unit 3 updates the connection strength between the first neuron and the second neuron based on the elapsed time managed by the elapsed time management unit 1 and the firing frequency managed by the firing frequency management unit 2 when the second neuron fires. However, this is only one example, and the connection strength management unit 3 may also update the connection strength between the first neuron and the second neuron based on the elapsed time managed by the elapsed time management unit 1 and an index that determines the amount of change in connection strength when the second neuron fires. The index that determines the amount of change in connection strength may be the firing frequency managed by the firing frequency management unit 2, or, for example, the probability of the first neuron firing, or the amount of change determined in the previous instance. The index that determines the amount of change may be any combination of, for example, the firing frequency, the probability of firing, or the amount of change determined in the previous instance. The same applies to other embodiments.

[0038] In the bond strength update device shown in Figure 1, the bond strength management unit 3 controls the value H corresponding to the elapsed time T, as shown in equation (3). Ｂ (t) and the value R corresponding to the ignition frequency Ｂ The change in bond strength w Δw is determined according to (t). However, this is just one example, and the bond strength management unit 3 determines the rate of change of the change in bond strength w Δw as the elapsed time T, based on the change in the first index Δw １ The reaction coefficient UF is calculated. In this case, the reaction coefficient UF may be, for example, a value inversely proportional to the elapsed time T, or a value H corresponding to the elapsed time T. Ｂ (t) is one example. The bond strength control unit 3 controls, for example, the change amount Δw based on the first index. １ The value R corresponds to the ignition frequency. Ｂ (t) and A Ａ We calculate the sum of the two, and then the calculated change Δw １ The final change amount Δw may be determined by multiplying the above reaction coefficient UF by the result. That is, the bond strength control unit 3 determines the change amount Δw calculated according to the elapsed time T. １ The change amount Δw is calculated according to the firing frequency using the rate of change (the update range mentioned above). １You may also adjust the following: In that case, the amount of change after adjustment Δw １ This is considered the final change Δw. It is also possible to modify equation (4) to treat the passage of time as the first indicator and the firing frequency as the second indicator. That is, the value H based on the passage of time. Ｂ A threshold is set for (t), and the change amount based on the passage of time, taking the threshold into account, is given a value R based on the firing frequency. Ｂ It is also possible to multiply by (t) as the reaction coefficient. As a change based on the passage of time with a threshold, for example, A Ｂ -H Ｂ (t) is one example.

[0039] As is clear from the above, the connection strength update device shown in Figure 1 can prevent situations in which connection strengths are updated when such updates are unnecessary. Specifically, it can prevent situations such as when the connection strength between the first neuron and the second neuron becomes zero, resulting in the erasure of learned data. As a result, the connection strength update device appropriately updates the connection strengths between multiple neurons included in the spiking neural network, thereby improving the accuracy of the tasks performed by the spiking neural network. Therefore, when a user uses a spiking neural network to which the connection strength update device shown in Figure 1 is applied, they can obtain more accurate processing results for tasks than when using a spiking neural network to which the connection strength update device disclosed in Non-Patent Literature 1 is applied.

[0040] If the task performed by a spiking neural network to which the connection strength update device shown in Figure 1 is applied is, for example, image recognition processing, and a system is configured such that the image recognition processing results from the spiking neural network are output to a security device (not shown), then the security device will be able to perform highly accurate detection of intrusions by animals, etc., into the monitored area. If the task performed by a spiking neural network to which the connection strength update device shown in Figure 1 is applied is, for example, voice recognition processing, and a system is configured such that the voice recognition processing results from the spiking neural network are output to a person identification device (not shown), then the person identification device will be able to perform highly accurate person identification. If the task performed by a spiking neural network to which the connection strength update device shown in Figure 1 is applied is, for example, pathfinding processing, and a system is configured such that the pathfinding processing results from the spiking neural network are output to a navigation device (not shown), then the navigation device will be able to display highly accurate pathfinding processing results on a map.

[0041] Embodiment 2. Embodiment 2 describes a bond strength updating device that, when updating the bond strength w, receives, for example, control information from an external source indicating that the elapsed time T managed by the elapsed time management unit 1 should be used, or control information indicating that the elapsed time T should not be used.

[0042] Figure 7 is a configuration diagram showing a bond strength updating device according to Embodiment 2. In Figure 7, the same reference numerals as in Figure 1 indicate the same or corresponding parts, so a detailed explanation is omitted. Figure 8 is a hardware configuration diagram showing the hardware of the bond strength updating device according to Embodiment 2. In Figure 8, the same reference numerals as in Figure 2 indicate the same or corresponding parts, so a detailed explanation is omitted. The bond strength updating device shown in Figure 7 comprises an elapsed time management unit 1, an ignition frequency management unit 2, a bond strength management unit 5, and a bond strength holding unit 4.

[0043] The connection strength management unit 5 is implemented, for example, by the connection strength management circuit 15 shown in Figure 8. When the connection strength management unit 5 is given control information indicating that the elapsed time T managed by the elapsed time management unit 1 should be used when updating the connection strength, it updates the connection strength w between the first neuron and the second neuron based on the elapsed time T and the firing frequency managed by the firing frequency management unit 2. When the connection strength management unit 5 is given control information indicating that the elapsed time T managed by the elapsed time management unit 1 should not be used when updating the connection strength, it updates the connection strength w between the first neuron and the second neuron based on the firing frequency managed by the firing frequency management unit 2. In the connection strength update device shown in Figure 7, the connection strength management unit 5 updates the connection strength from the first neuron to the second neuron as the connection strength between the first neuron and the second neuron. In Figure 7, the control signal is denoted as a "mode selection signal".

[0044] In Figure 7, it is assumed that the components of the coupling strength update device—the elapsed time management unit 1, the ignition frequency management unit 2, the coupling strength management unit 5, and the coupling strength retention unit 4—are each implemented by dedicated hardware as shown in Figure 8. Specifically, the coupling strength update device is assumed to be implemented by an elapsed time management circuit 11, an ignition frequency management circuit 12, a coupling strength management circuit 15, and a coupling strength retention circuit 14. The elapsed time management circuit 11, the ignition frequency management circuit 12, and the coupling strength management circuit 15 can each be, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC, an FPGA, or a combination thereof.

[0045] The components of the bond strength update device are not limited to those implemented by dedicated hardware; the bond strength update device may also be implemented by software, firmware, or a combination of software and firmware. When the bond strength update device is implemented by software or firmware, the bond strength holding unit 4 is configured on the memory 21 shown in Figure 3. A bond strength update program that causes the computer to execute the respective processing procedures in the elapsed time management unit 1, the ignition frequency management unit 2, and the bond strength management unit 5 is stored in the memory 21. Then, the processor 22 shown in Figure 3 executes the bond strength update program stored in the memory 21.

[0046] Furthermore, Figure 8 shows an example in which each component of the bond strength updating device is implemented by dedicated hardware, while Figure 3 shows an example in which the bond strength updating device is implemented by software or firmware. However, this is only one example, and some components of the bond strength updating device may be implemented by dedicated hardware, while the remaining components may be implemented by software or firmware.

[0047] Next, the operation of the bond strength updating device shown in Figure 7 will be explained. Except for the bond strength management unit 5, it is the same as the bond strength updating device shown in Figure 1. Therefore, only the operation of the bond strength management unit 5 will be explained here.

[0048] The coupling strength management unit 5 acquires the coupling strength w held in the coupling strength holding unit 4. When updating the coupling strength, the coupling strength management unit 5 receives a mode selection signal from, for example, an external source. When the mode selection signal is control information indicating that the elapsed time T should be used, the coupling strength management unit 5 determines the change amount Δw of the coupling strength w, as shown in equation (3), similar to the coupling strength management unit 3 shown in Figure 1. Then, the coupling strength management unit 5 updates the coupling strength w using the change amount Δw, as shown in equation (4).

[0049] The coupling strength management unit 5, if the mode selection signal is control information indicating that the elapsed time T is not used, then uses the H shown in equation (3). ＢAssuming (t) is 1, the change amount Δw of the bond strength w is determined. Then, the bond strength management unit 5 updates the bond strength w using the change amount Δw, as shown in equation (4). The bond strength management unit 5 has the bond strength holding unit 4 hold the updated bond strength w.

[0050] In the above embodiment 2, the connection strength update device shown in Figure 7 is configured such that when the connection strength management unit 5 is given control information indicating that it will use the elapsed time managed by the elapsed time management unit 1 when updating the connection strength, it updates the connection strength between the first neuron and the second neuron based on the elapsed time and the firing frequency managed by the firing frequency management unit 2. When control information indicating that it will not use the elapsed time managed by the elapsed time management unit 1 is given, it updates the connection strength between the first neuron and the second neuron based on the firing frequency managed by the firing frequency management unit 2. Therefore, when the connection strength update device shown in Figure 7 is given control information indicating that it will use the elapsed time, it can update the connection strength only when it is necessary to update the connection strength, similar to the connection strength update device shown in Figure 1. Furthermore, when the connection strength update device shown in Figure 7 is given control information indicating that it will not use the elapsed time, it can learn the connection strength while considering the relationship between pixels, for example, when it is desired to train a neural network on data in which adjacent pixel data is related, such as image data. Furthermore, the coupling strength update device can select a method for changing the coupling strength while taking external factors into consideration, allowing for more appropriate updates to the coupling strength and, as a result, maintaining the network relationship properly.

[0051] Embodiment 3. Embodiment 3 describes a bond strength updating device that, for example, receives selection information indicating information to be used when updating bond strength from an external source.

[0052] Figure 9 is a configuration diagram showing a bond strength updating device according to Embodiment 3. In Figure 9, the same reference numerals as in Figures 1 and 7 indicate the same or corresponding parts, so a detailed explanation is omitted. Figure 10 is a hardware configuration diagram showing the hardware of the bond strength updating device according to Embodiment 3. In Figure 10, the same reference numerals as in Figures 2 and 8 indicate the same or corresponding parts, so a detailed explanation is omitted. The bond strength updating device shown in Figure 9 comprises elapsed time management units 1a, 1b, ignition frequency management units 2a, 2b, bond strength management unit 6, and bond strength holding unit 4.

[0053] The elapsed time management unit 1a is implemented, for example, by the elapsed time management circuit 11a shown in Figure 10. The elapsed time management unit 1a receives a spike signal SO output from the first neuron when the first neuron fires. １ The elapsed time management unit 1a receives a spike signal SO from the first neuron. １ The elapsed time since the output is managed as the first elapsed time. The elapsed time management unit 1b is implemented, for example, by the elapsed time management circuit 11b shown in Figure 10. The elapsed time management unit 1b receives the spike signal SO output from the second neuron when the second neuron fires. ２ The elapsed time management unit 1b receives a spike signal SO from the second neuron. ２ The elapsed time since the output is managed as the second elapsed time. The coupling strength update device shown in Figure 9 is equipped with elapsed time management units 1a and 1b. However, this is merely an example, and the coupling strength update device may be equipped with one elapsed time management unit, and one elapsed time management unit may have the functions of both elapsed time management unit 1a and elapsed time management unit 1b. The elapsed time management unit 1a and the elapsed time management unit 1b differ only in the spike signals they receive, that is, only in the signals that trigger the processing, and the processing itself may be the same as that of the elapsed time management unit 1 in Embodiment 1. However, the time constants and the like related to the time intervals used in the processing are assumed to be independent of each other.

[0054] The firing frequency management unit 2a is implemented, for example, by the firing frequency management circuit 12a shown in Figure 10. The firing frequency management unit 2 receives a spike signal SO output from the first neuron when the first neuron fires. １ The ignition frequency management unit 2a is given the spike signal SO per unit time. １ Based on the number of outputs, the firing frequency, which is the frequency at which the first neuron fires, is managed as the first firing frequency. The firing frequency management unit 2b is realized, for example, by the firing frequency management circuit 12b shown in Figure 10. The firing frequency management unit 2b receives the spike signal SO output from the second neuron when the second neuron fires. ２ The ignition frequency management unit 2b is given the spike signal SO per unit time. ２ Based on the number of outputs, the firing frequency of the second neuron is managed as the second firing frequency. The connection strength update device shown in Figure 9 includes firing frequency management units 2a and 2b. However, this is merely an example, and the connection strength update device may have one firing frequency management unit, and this single firing frequency management unit may have the functions of firing frequency management unit 2a and firing frequency management unit 2b. The firing frequency management unit 2a and firing frequency management unit 2b differ only in the input spike signals, that is, only in the signals that trigger the processing, and the processing itself may be the same as that of the firing frequency management unit 2 in Embodiment 1. However, the time constants related to the time interval used in the processing are assumed to be independent for each.

[0055] The coupling strength management unit 6 is implemented, for example, by the coupling strength management circuit 16 shown in Figure 10. The coupling strength management unit 6 includes a first update processing unit 6a and a second update processing unit 6b. The coupling strength management unit 6 receives a spike signal SO output from the first neuron when the first neuron fires. １ And, when the second neuron fires, the spike signal SO is output from the second neuron. ２The connection strength management unit 6 selects either a first elapsed time, a second elapsed time, or no elapsed time, based on selection information indicating the information to be used when updating the connection strength, and also selects either a first firing frequency, a second firing frequency, or no firing frequency. The connection strength management unit 6 updates the connection strength between the first neuron and the second neuron based on the value based on the selected elapsed time and the value based on the selected firing frequency.

[0056] The first update processing unit 6a, when the second neuron, i.e., the neuron to which it is connected, fires, selects either a first elapsed time, a second elapsed time, or no elapsed time based on selection information indicating the information to be used when updating the connection strength, and also selects either a first firing frequency, a second firing frequency, or no firing frequency, and then updates the connection strength from the first neuron to the second neuron based on the value based on the selected elapsed time and the value based on the selected firing frequency.

[0057] The second update processing unit 6b, when the first neuron, i.e., the source neuron, fires, selects either a first elapsed time, a second elapsed time, or no elapsed time based on selection information indicating the information to be used when updating the connection strength, and also selects either a first firing frequency, a second firing frequency, or no firing frequency, and then updates the connection strength from the first neuron to the second neuron based on the value based on the selected elapsed time and the value based on the selected firing frequency.

[0058] In Figure 9, it is assumed that the components of the coupling strength update device—the elapsed time management units 1a, 1b, the ignition frequency management units 2a, 2b, the coupling strength management unit 6, and the coupling strength retention unit 4—are each implemented by dedicated hardware as shown in Figure 10. Specifically, the coupling strength update device is assumed to be implemented by elapsed time management circuits 11a, 11b, ignition frequency management circuits 12a, 12b, the coupling strength management circuit 16, and the coupling strength retention circuit 14. The elapsed time management circuits 11a, 11b, the ignition frequency management circuits 12a, 12b, and the coupling strength management circuit 16 can each be, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC, an FPGA, or a combination thereof.

[0059] The components of the bond strength update device are not limited to those implemented by dedicated hardware; the bond strength update device may also be implemented by software, firmware, or a combination of software and firmware. When the bond strength update device is implemented by software or firmware, the bond strength holding unit 4 is configured on the memory 21 shown in Figure 3. Bond strength update programs that cause the computer to execute the respective processing procedures in the elapsed time management units 1a, 1b, the ignition frequency management units 2a, 2b, and the bond strength management unit 6 are stored in the memory 21. Then, the processor 22 shown in Figure 3 executes the bond strength update programs stored in the memory 21.

[0060] Furthermore, Figure 10 shows an example in which each component of the bond strength updating device is implemented by dedicated hardware, and Figure 3 shows an example in which the bond strength updating device is implemented by software or firmware. However, this is only one example, and some components of the bond strength updating device may be implemented by dedicated hardware, while the remaining components may be implemented by software or firmware.

[0061] Next, the operation of the coupling strength update device shown in Figure 9 will be explained. The elapsed time management unit 1a receives a spike signal SO from the first neuron. １The elapsed time since the output is defined as the first elapsed time T. １ It is managed as follows. Specifically, the elapsed time management unit 1a has a timer function, and the timer value CT of the elapsed time management unit 1a １ It is initially set to 0. Timer value CT １ This decreases over time, and a spike signal SO is emitted from the first neuron. １ Each time an output is generated, a certain value is added. The spike signal SO is output from the first neuron. １ The first elapsed time T after the output. １ H is shown in the following equation (5). Ｂ１ (t) is expressed by H Ｂ１ (t) is the spike signal SO １ The spike signal SO is large when not much time has passed since the output. １ The longer the time elapsed since the output, the smaller the value becomes.

[0062] The elapsed time management unit 1b receives a spike signal SO from the second neuron. ２ The elapsed time since the output is defined as the second elapsed time T. ２ It is managed as follows. Specifically, the elapsed time management unit 1b has a timer function, and the timer value CT of the elapsed time management unit 1b ２ It is initially set to 0. Timer value CT ２ This decreases over time, and a spike signal SO is emitted from the second neuron. ２ Each time an output is generated, a certain value is added. A spike signal SO is sent from the second neuron. ２ The second elapsed time T after the output. ２ H is shown in the following equation (6). Ｂ２ (t) is expressed by H Ｂ２ (t) is the spike signal SO ２ The spike signal SO is large when not much time has passed since the output. ２ The longer the time elapsed since the output, the smaller the value becomes.

[0063] In equations (5) to (6), τ Ｂ１１ , τＢ２１ This is the time constant.

[0064] The ignition frequency management unit 2a controls the spike signal SO per unit time. １ Based on the number of outputs, R is a value corresponding to the first firing frequency, which is the frequency at which the first neuron fires. Ｂ１ Manage (t). R Ｂ１ (t) decreases over time, and a spike signal SO is emitted from the first neuron. １ Each time it is output, a certain value is added. Ｂ１ (t) increases as the firing frequency of the first neuron increases, and the spike signal SO from the first neuron １ If no output is produced at all, it becomes 0. R is the value corresponding to the first firing frequency, which is the firing frequency of the first neuron. Ｂ１ (t) can be expressed as shown in equation (7) below.

[0065] The ignition frequency management unit 2b controls the spike signal SO per unit time. ２ Based on the number of outputs, the value R corresponds to the second firing frequency, which is the frequency at which the second neuron fires. Ｂ２ Manage (t). R Ｂ２ (t) decreases over time, and a spike signal SO is emitted from the second neuron. ２ Each time it is output, a certain value is added. Ｂ２ (t) increases as the firing frequency of the second neuron increases, and the spike signal SO from the second neuron ２ If no output is produced at all, it becomes 0. R is the value corresponding to the second firing frequency, which is the firing frequency of the second neuron. Ｂ２ (t) can be expressed as shown in equation (8) below.

[0066] In equations (7) to (8), τ Ｂ１２ , τ Ｂ２２ This is the time constant.

[0067] The connection strength management unit 6 is provided with selection information that indicates the information to be used when updating the connection strength. Figure 11 is an explanatory diagram showing an example of the selection information provided to the connection strength management unit 6. Here, mode (0) is the spike signal S0 that indicates the firing of the target neuron.２ This mode is triggered by an update processing unit 6a, in which the first update processing unit 6a performs an update, and the first firing frequency (firing frequency of the source neuron) is selected as the firing frequency, and the elapsed time is set to unused. In mode (0), no update processing is performed by the second update processing unit 6b. In the example shown below, in an update processing unit that does not perform an update, all variables are set to zero so that the change amount Δw becomes zero, but the execution of the processing may be controlled by methods other than changing the values ​​of the variables. Also, in mode (1), the spike signal S0 indicating the firing of the target neuron is selected. ２ Mode (2) is a mode in which an update process is performed by the first update processing unit 6a triggered by a first firing frequency (firing frequency of the source neuron) and a first elapsed time (elapsed time since the source neuron fired) is selected as the firing frequency. ２ The update process by the first update processing unit 6a, triggered by the signal, and the spike signal SO indicating firing of the source neuron. １ Mode (3) is a mode in which an update process is performed by a second update processing unit 6b triggered by a spike signal S0, and in this mode, the first update processing unit 6a selects a first firing frequency (firing frequency of the source neuron) as the firing frequency and a first elapsed time (elapsed time since the firing of the source neuron) as the elapsed time, and the second update processing unit 6b selects a second firing frequency (firing frequency of the destination neuron) as the firing frequency and a second elapsed time (elapsed time since the firing of the destination neuron) as the elapsed time. ２ The update process by the first update processing unit 6a, triggered by the signal, and the spike signal SO indicating firing of the source neuron. １Mode (4) is a mode in which an update process is performed by a second update processing unit 6b triggered by a spike signal S0, and in the first update processing unit 6a, a first firing frequency (firing frequency of the source neuron) and a first elapsed time (elapsed time since the source neuron fired) are selected as the firing frequency, and in the second update processing unit 6b, no firing frequency is selected as the firing frequency, and a second elapsed time (elapsed time since the destination neuron fired) is selected as the elapsed time. ２ The update process by the first update processing unit 6a, triggered by the signal, and the spike signal SO indicating firing of the source neuron. １ Mode (5) is a mode in which an update process is performed by a second update processing unit 6b triggered by a spike signal S0 ２ The update process by the first update processing unit 6a, triggered by the signal, and the spike signal SO indicating firing of the source neuron. １ This mode involves performing an update process by a second update processing unit 6b triggered by a certain event, wherein in the first update processing unit 6a, the firing frequency is set to "fire frequency unused" and the elapsed time is set to "first elapsed time (elapsed time since firing of the source neuron)", and in the second update processing unit 6b, the firing frequency is set to "fire frequency unused" and the elapsed time is set to "second elapsed time (elapsed time since firing of the destination neuron)".

[0068] In the example in Figure 11, if the selected information is mode (0), the first update processing unit 6a uses H Ｂ (t) is set to 1, and the first update processing unit 6a uses R Ｂ (t) as R Ｂ１ (t) is selected. Also, H used by the second update processing unit 6b Ｂ(t) is set to 0, and the second update processing unit 6b uses R Ｂ (t) is set to 0. If the selected information is mode (1), the first update processing unit 6a uses H Ｂ (t) as H Ｂ１ (t) is selected, and the first update processing unit 6a uses R Ｂ (t) as R Ｂ１ (t) is selected. Also, H used by the second update processing unit 6b Ｂ (t) is set to 0, and the second update processing unit 6b uses R Ｂ (t) is set to 0. If the selected information is mode (2), the first update processing unit 6a uses H Ｂ (t) as H Ｂ１ (t) is selected, and the first update processing unit 6a uses R Ｂ (t) as R Ｂ１ (t) is selected. Also, H used by the second update processing unit 6b Ｂ (t) as H Ｂ２ (t) is selected, and the second update processing unit 6b uses R Ｂ (t) as R Ｂ２ (t) is selected.

[0069] If the selected information is mode (3), the first update processing unit 6a uses H Ｂ (t) as H Ｂ１ (t) is selected, and the first update processing unit 6a uses R Ｂ (t) as R Ｂ１ (t) is selected. Also, H used by the second update processing unit 6b Ｂ (t) as H Ｂ２ (t) is selected, and the second update processing unit 6b uses R Ｂ (t) is set to 0. If the selected information is mode (4), the first update processing unit 6a uses H Ｂ (t) as H Ｂ１ (t) is selected, and the first update processing unit 6a uses R Ｂ (t) as R Ｂ２ (t) is selected. Also, H used by the second update processing unit 6b Ｂ (t) as H Ｂ２ (t) is selected, and the second update processing unit 6b uses RＢ (t) as R Ｂ１ (t) is selected. If the selected information is mode (5), the first update processing unit 6a uses H Ｂ (t) as H Ｂ１ (t) is selected, and the first update processing unit 6a uses R Ｂ (t) is set to 0. Also, H used by the second update processing unit 6b Ｂ (t) as H Ｂ２ (t) is selected, and the second update processing unit 6b uses R Ｂ (t) is set to 0.

[0070] The first update processing unit 6a receives a spike signal SO from the second neuron. ２ When the output is generated, the change amount Δw of the connection strength w is determined as shown in equation (9) below. The first update processing unit 6a updates the connection strength from the first neuron to the second neuron using the change amount Δw as shown in equation (4). The first update processing unit 6a has the connection strength holding unit 4 hold the updated connection strength w.

[0071] Δw = (A Ａ +R Ｂ (t) H Ｂ (t) (9) Mode (0) → H Ｂ (t) = 1, R Ｂ (t) = R Ｂ１ (t) Mode (1) → H Ｂ (t) = H Ｂ１ (t), R Ｂ (t) = R Ｂ１ (t) Mode (2) → H Ｂ (t) = H Ｂ１ (t), R Ｂ (t) = R Ｂ１ (t) Mode (3) → H Ｂ (t) = H Ｂ１ (t), R Ｂ (t) = R Ｂ１ (t) Mode (4) → H Ｂ (t) = H Ｂ１ (t), R Ｂ (t) = R Ｂ２ (t) Mode (5) → H Ｂ (t) = H Ｂ１ (t), RＢ (t) = 0

[0072] The second update processing unit 6b receives a spike signal SO from the first neuron. １ When the output is obtained, the change in bond strength w Δw is determined as shown in equation (10) below. In equation (10), A Ｂ This is a negative value. The second update processing unit 6b updates the connection strength from the first neuron to the second neuron using the change amount Δw, as shown in equation (4). The second update processing unit 6b has the connection strength retention unit 4 retain the updated connection strength w.

[0073] Δw = -(A Ｂ +R Ｂ (t) H Ｂ (t) (10) Mode (0) → H Ｂ (t) = 0, R Ｂ (t) = 0 Mode (1) → H Ｂ (t) = 0, R Ｂ (t) = 0 Mode (2) → H Ｂ (t) = H Ｂ２ (t), R Ｂ (t) = R Ｂ２ (t) Mode (3) → H Ｂ (t) = H Ｂ２ (t), R Ｂ (t) = 0 Mode (4) → H Ｂ (t) = H Ｂ２ (t), R Ｂ (t) = R Ｂ１ (t) Mode (5) → H Ｂ (t) = H Ｂ２ (t), R Ｂ (t) = 0

[0074] In the above embodiment 3, the elapsed time management unit 1a manages the elapsed time since the first neuron outputs a spike signal as the first elapsed time, and the elapsed time management unit 1b manages the elapsed time since the second neuron outputs a spike signal as the second elapsed time. The firing frequency management unit 2a manages the firing frequency of the first neuron as the first firing frequency, and the firing frequency management unit 2b manages the firing frequency of the second neuron as the second firing frequency, as shown in Figure 9, thus configuring the connection strength updating device. Furthermore, the connection strength management unit 6 of the connection strength update device shown in Figure 9 selects either a first elapsed time, a second elapsed time, or no elapsed time based on selection information indicating the information to be used when updating the connection strength when the second neuron or the first neuron fires, and also selects either a first firing frequency, a second firing frequency, or no firing frequency, and updates the connection strength between the first neuron and the second neuron based on the value based on the selected elapsed time and the value based on the selected firing frequency. Therefore, the connection strength update device shown in Figure 9, like the connection strength update device shown in Figure 1, can update the connection strength only when it is necessary to update it. In addition, the connection strength update device shown in Figure 9 can apply the learning rule of STDP (Spike-Timing-Dependent Plasticity) or the learning rule of T-STDP (Triple STDP). Since STDP and T-STDP are known technologies, a detailed explanation is omitted. Furthermore, the learning rule for STDP corresponds to the learning rule in mode (5), and the learning rule for T-STDP corresponds to the learning rule in mode (4).

[0075] Embodiment 4. Embodiment 4 describes a connection strength updating device in which the connection strength management unit 7 updates the connection strength from the first neuron to the second neuron based on a first elapsed time and a first firing frequency when the second neuron fires, and updates the connection strength from the first neuron to the second neuron based on a second elapsed time and a second firing frequency when the first neuron fires.

[0076] Figure 12 is a configuration diagram showing a bond strength updating device according to Embodiment 4. In Figure 12, the same reference numerals as in Figures 1, 7, and 9 indicate the same or corresponding parts, so a detailed explanation is omitted. Figure 13 is a hardware configuration diagram showing the hardware of the bond strength updating device according to Embodiment 4. In Figure 13, the same reference numerals as in Figures 2, 8, and 10 indicate the same or corresponding parts, so a detailed explanation is omitted. The bond strength updating device shown in Figure 13 comprises elapsed time management units 1a, 1b, ignition frequency management units 2a, 2b, bond strength management unit 7, and bond strength holding unit 4.

[0077] The coupling strength management unit 7 is implemented, for example, by the coupling strength management circuit 17 shown in Figure 13. The coupling strength management unit 7 includes a first update processing unit 7a and a second update processing unit 7b. The coupling strength management unit 7 receives a spike signal SO output from the first neuron when the first neuron fires. １ And, when the second neuron fires, the spike signal SO is output from the second neuron. ２ The connection strength management unit 7 updates the connection strength from the first neuron to the second neuron based on the first elapsed time and the first firing frequency when the second neuron fires. The connection strength management unit 7 updates the connection strength from the first neuron to the second neuron based on the second elapsed time and the second firing frequency when the first neuron fires.

[0078] The first update processing unit 7a updates the connection strength from the first neuron to the second neuron based on the first elapsed time and the first firing frequency when the second neuron fires. The second update processing unit 7b updates the connection strength from the first neuron to the second neuron based on the second elapsed time and the second firing frequency when the first neuron fires.

[0079] In Figure 12, it is assumed that the components of the coupling strength update device—the elapsed time management units 1a, 1b, the ignition frequency management units 2a, 2b, the coupling strength management unit 7, and the coupling strength retention unit 4—are each implemented by dedicated hardware as shown in Figure 13. Specifically, the coupling strength update device is assumed to be implemented by elapsed time management circuits 11a, 11b, ignition frequency management circuits 12a, 12b, the coupling strength management circuit 17, and the coupling strength retention circuit 14. The elapsed time management circuits 11a, 11b, the ignition frequency management circuits 12a, 12b, and the coupling strength management circuit 17 can each be, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC, an FPGA, or a combination thereof.

[0080] The components of the bond strength update device are not limited to those implemented by dedicated hardware; the bond strength update device may also be implemented by software, firmware, or a combination of software and firmware. When the bond strength update device is implemented by software or firmware, the bond strength holding unit 4 is configured on the memory 21 shown in Figure 3. A bond strength update program that causes the computer to execute the respective processing procedures in the elapsed time management unit 1a, 1b, the ignition frequency management unit 2a, 2b, and the bond strength management unit 7 is stored in the memory 21. Then, the processor 22 shown in Figure 3 executes the bond strength update program stored in the memory 21.

[0081] Furthermore, Figure 13 shows an example in which each component of the bond strength updating device is implemented by dedicated hardware, while Figure 3 shows an example in which the bond strength updating device is implemented by software or firmware. However, this is merely one example, and some components of the bond strength updating device may be implemented by dedicated hardware, while the remaining components may be implemented by software or firmware.

[0082] Next, the operation of the coupling strength update device shown in Figure 12 will be explained. Except for the coupling strength management unit 7, it is the same as the coupling strength update device shown in Figure 9. Therefore, only the operation of the coupling strength management unit 7 will be explained here. The coupling strength update device shown in Figure 12 can handle multiple modes even without selection information being provided from the outside.

[0083] In the bond strength updating device shown in Figure 12, A is shown in the following formula (11) Ａ and A shown in the following equation (12) Ｂ And shall be 0 or a positive value. Also, in the bond strength updating device shown in Figure 12, R Ｂ１ (t) and R Ｂ２ (t) is assumed to be 0 or a positive value.

[0084] The first update processing unit 7a receives a spike signal SO from the second neuron. ２ When this is output, the first elapsed time T is as shown in equation (11) below. １ The corresponding value H Ｂ１ (t) and the value R corresponding to the first ignition frequency. Ｂ１ Based on (t), the change in bond strength w, Δw, is determined. Δw = (A Ａ +R Ｂ１ (t) H Ｂ１ (t) (11) The first update processing unit 7a updates the connection strength from the first neuron to the second neuron using the change amount Δw, as shown in equation (4). The first update processing unit 7a has the connection strength w after the update stored in the connection strength holding unit 4.

[0085] The second update processing unit 7b receives a spike signal SO from the first neuron. １ When this is output, the elapsed time T of 2 is as shown in equation (12) below. ２ The corresponding value H Ｂ２ (t) and the value R corresponding to the second firing frequency. Ｂ２ Based on (t), the change in bond strength w, Δw, is determined. Δw = -(A Ｂ +R Ｂ２ (t) H Ｂ２(t) (12) The second update processing unit 7b updates the connection strength from the first neuron to the second neuron using the change amount Δw, as shown in equation (4). The second update processing unit 7b has the connection strength w after the update stored in the connection strength holding unit 4.

[0086] First elapsed time T １ If it is longer than the predetermined time, H Ｂ１ Since (t) is almost 0, the change Δw is almost 0. Therefore, the first elapsed time T １ If the elapsed time is longer than a predetermined time, the connection strength w from the first neuron to the second neuron is maintained. ２ If it is longer than the predetermined time, H Ｂ２ Since (t) is almost 0, the change Δw is almost 0. Therefore, the second elapsed time T ２ If this duration is longer than a predetermined time, the connection strength w from the first neuron to the second neuron is maintained.

[0087] In the above embodiment 4, the connection strength management unit 7 is configured to update the connection strength from the first neuron to the second neuron based on a first elapsed time and a first firing frequency when the second neuron fires, and to update the connection strength from the first neuron to the second neuron based on a second elapsed time and a second firing frequency when the first neuron fires, as shown in the connection strength update device in Figure 12. Therefore, the connection strength update device shown in Figure 12 can update the connection strength only when it is necessary, similar to the connection strength update device shown in Figure 1. Furthermore, the connection strength update device shown in Figure 12 can be configured more simply than the connection strength update device shown in Figure 9.

[0088] Embodiment 5. Embodiment 5 describes a bond strength updating device in which both a lower limit and an upper limit for the bond strength w are set.

[0089] The configuration of the connection strength updating device according to Embodiment 5 is the same as the configuration of the connection strength updating device according to any of Embodiments 1 to 4. Therefore, the configuration diagram showing the connection strength updating device according to Embodiment 5 is Figure 1, Figure 7, Figure 9, or Figure 12. In the connection strength management unit 3 (or 5, 6, 7), the lower limit value w of the connection strength w between the first neuron and the second neuron is set. ｍｉｎ And the upper limit w of the connection strength w between the first neuron and the second neuron. ｍａｘ And is set. Lower limit w ｍｉｎ is a value greater than or equal to 0, and is the upper limit w ｍａｘ The lower limit is lol ｍｉｎ This value is greater than [value]. Hereinafter, the bond strength control unit 3 (or 5, 6, 7) will be referred to as the bond strength control unit 3, etc.

[0090] The bond strength control unit 3 calculates the bond strength w according to equation (4), and when the bond strength w is below the lower limit w ｍｉｎ If it is less than the lower limit of the bond strength w, then the lower limit of w ｍｉｎ Set to the upper limit value w. Also, the bond strength control unit 3 etc. will set the bond strength w to the upper limit value w. ｍａｘ If it is greater than, the bond strength w is set to the upper limit w ｍａｘ This is set to the lower limit value w, which is the updated bond strength w by the bond strength management unit 3, etc. ｍｉｎ And the upper limit lol ｍａｘ It falls between these two points.

[0091] Here, the bond strength control unit 3, etc., controls the bond strength w when it reaches the lower limit w. ｍｉｎ If it is less than the lower limit of the bond strength w, then the lower limit of w ｍｉｎ Set to the upper limit value w for bond strength w ｍａｘ If it is greater than, the bond strength w is set to the upper limit w ｍａｘ This is the setting. However, this is just one example, and the bond strength management unit 3 may set the lower and upper limits of the bond strength w by multiplying the update width Δw by a value that depends on the current bond strength w, and updating the current bond strength w based on the result of that multiplication. The value that depends on the current bond strength w is, for example, (w - w ｍｉｎ ) (w ｍａｘ -w) can be used.

[0092] In the above embodiment 5, both a lower limit and an upper limit for bond strength are set, and the bond strength update device is configured such that the bond strength management unit 3 updates the bond strength so that it falls between the lower limit and the upper limit. Therefore, the bond strength update device can keep the updated bond strength within a certain range.

[0093] Embodiment 6. In the connection strength updating device according to Embodiments 1 to 5, it is assumed that the elapsed time management unit 1 and the firing frequency management unit 2 are each provided at the connection unit between two neurons, as shown in Figure 14. Figure 14 is an explanatory diagram showing that the elapsed time management unit 1 and the firing frequency management unit 2 are each provided at the connection unit between two neurons. The elapsed time management unit 1 and the firing frequency management unit 2 may each be provided at the neuron unit of the neural network, as shown in Figure 15. Figure 15 is an explanatory diagram showing that the elapsed time management unit 1 and the firing frequency management unit 2 are each provided at the neuron unit. By managing the elapsed time and firing frequency at the neuron level, it is possible to prevent overlapping management of result time and firing frequency in a network structure where, for example, one neuron has multiple connection destinations. For example, neuron N Ｓ from neuron N ２ Neuron N obtained when updating the connection strength to Ｓ Information on the elapsed time and frequency based on the firing of neurons N Ｓ from neuron N ３ Since this is the same information obtained when updating the connection strength to neuron N, these are shared and used to create a common neuron N Ｓ It can be managed in conjunction with other items.

[0094] When the elapsed time management unit 1 and the firing frequency management unit 2 are provided at the connection unit between two neurons, the configuration becomes redundant compared to when they are provided at the neuron level, but the exchange of managed values ​​can be accelerated. When the elapsed time management unit 1 and the firing frequency management unit 2 are provided at the neuron level, even if the number of connections a neuron has increases, the elapsed time and firing frequency are each stored in only one area, thus reducing the area required to store each of the elapsed time and firing frequency. Depending on the structure of the spiking neural network, the processing time, memory area size, circuit area size, etc. can be adjusted according to the purpose by using either the connection unit between two neurons or the neuron level. It is also possible to have a mix of configurations where the elapsed time management unit 1 and the firing frequency management unit 2 are provided at the connection unit between two neurons and configurations where they are provided at the neuron level.

[0095] The spiking neural network system to which the connection strength update device according to Embodiments 1 to 6 is applied can handle, for example, image data or audio data. As an example, when handling image data, data can be input to the spiking neural network by preparing a neuron corresponding to each pixel of the image and adding or continuously adding a value corresponding to the pixel value to that neuron, or by generating a spike sequence according to the pixel value and inputting that spike sequence to the corresponding neuron. The specific method of applying data to the spiking neural network in Embodiments 1 to 6 is not particularly limited.

[0096] Furthermore, this disclosure allows for free combination of each embodiment, modification of any component in each embodiment, or omission of any component in each embodiment.

[0097] This disclosure is applicable to a coupling strength update device, a coupling strength update method, a coupling strength update program, and a spiking neural network system.

[0098] 1, 1a, 1b Elapsed time management unit, 2, 2a, 2b Ignition frequency management unit, 3 Bond strength management unit, 4 Bond strength retention unit, 5, 6, 7 Bond strength management unit, 6a, 7a First update processing unit, 6b, 7b Second update processing unit, 11, 11a, 11b Elapsed time management circuit, 12, 12a, 12b Ignition frequency management circuit, 13 Bond strength management circuit, 14 Bond strength retention circuit, 15, 16, 17 Bond strength management circuit, 21 Memory, 22 Processor.

Claims

A connection strength management unit manages the connection strength between a first neuron, which is one of several neurons included in a spiking neural network, and a second neuron, which is a neuron that acquires the spike signal output from the first neuron when the first neuron fires. A time management unit manages the elapsed time since the spike signal was output from the first neuron, The system includes a firing frequency management unit that manages the firing frequency, which is the frequency at which the first neuron fires. The aforementioned bond strength control unit is When the second neuron fires, the connection strength between the first neuron and the second neuron is updated based on the elapsed time managed by the elapsed time management unit and the firing frequency managed by the firing frequency management unit. A bonding strength updating device characterized by the following features. 　 The aforementioned bond strength control unit is The bond strength updating device according to claim 1, characterized in that it determines the update range of the bond strength according to the elapsed time managed by the elapsed time management unit, and updates the bond strength based on the update range. 　 The aforementioned bond strength control unit is The bond strength updating device according to claim 2, characterized in that the update width is determined such that the longer the elapsed time managed by the elapsed time management unit, the smaller the update width of the bond strength. 　 The aforementioned bond strength control unit is The bond strength updating device according to claim 2, characterized in that if the elapsed time managed by the elapsed time management unit is longer than a predetermined time, the bond strength update range is determined to be zero or a value close to zero. 　 The aforementioned bond strength control unit is The connection strength updating device according to claim 1, characterized in that if the elapsed time managed by the elapsed time management unit is longer than a predetermined time, the connection strength between the first neuron and the second neuron is maintained. 　 The aforementioned bond strength control unit is The bond strength updating device according to claim 1, characterized in that it determines a first change in bond strength based on the ignition frequency managed by the ignition frequency management unit, determines the update width which is a reaction coefficient to be applied to the first change based on the elapsed time managed by the elapsed time management unit, and updates the bond strength based on the first change and the update width. 　 The aforementioned bond strength control unit is The bond strength updating device according to claim 6, characterized in that the final amount of change in bond strength is determined by adjusting the first amount of change according to the update width. 　 The aforementioned bond strength control unit is When updating the connection strength, if control information is provided indicating that the elapsed time managed by the elapsed time management unit should be used, the connection strength between the first neuron and the second neuron is updated based on the elapsed time and the firing frequency managed by the firing frequency management unit. When control information indicating that the elapsed time managed by the aforementioned elapsed time management unit will not be used is provided, the connection strength between the first neuron and the second neuron is updated based on the firing frequency managed by the aforementioned firing frequency management unit. The bond strength updating device according to claim 1, characterized by its features. 　 The aforementioned elapsed time management unit, The elapsed time since the spike signal was output from the first neuron is managed as the first elapsed time, and the elapsed time since the spike signal was output from the second neuron is managed as the second elapsed time. The aforementioned ignition frequency management unit is: The firing frequency of the first neuron is managed as the first firing frequency, and the firing frequency of the second neuron is managed as the second firing frequency. The aforementioned bond strength control unit is When the second neuron fires, or when the first neuron fires, based on selection information indicating the information used to update the connection strength, one of the first elapsed time, the second elapsed time, or no elapsed time is selected, and one of the first firing frequency, the second firing frequency, or no firing frequency is selected. The connection strength between the first neuron and the second neuron is updated based on the selected elapsed time and the selected firing frequency. The bond strength updating device according to claim 1, characterized by its features. 　 The aforementioned elapsed time management unit, The elapsed time since the spike signal was output from the first neuron is managed as the first elapsed time, and the elapsed time since the spike signal was output from the second neuron is managed as the second elapsed time. The aforementioned ignition frequency management unit is: The firing frequency of the first neuron is managed as the first firing frequency, and the firing frequency of the second neuron is managed as the second firing frequency. The aforementioned bond strength control unit is When the second neuron fires, the connection strength from the first neuron to the second neuron is updated based on the first elapsed time and the first firing frequency. When the first neuron fires, the connection strength from the second neuron to the first neuron is updated based on the second elapsed time and the second firing frequency. The bond strength updating device according to claim 1, characterized by its features. A lower limit and an upper limit for the bond strength are set. The aforementioned bond strength control unit is The bond strength is updated so that it falls between the lower limit and the upper limit. A bonding strength updating device according to any one of claims 1 to 10. 　 The elapsed time management unit and the ignition frequency management unit, respectively, A connection strength updating device according to any one of claims 1 to 11, characterized in that it is provided at the connection unit between two neurons. 　 The elapsed time management unit and the ignition frequency management unit, respectively, The connection strength updating device according to any one of claims 1 to 11, characterized in that it is provided for each neuron in the spiking neural network. 　 A connection strength management unit manages the connection strength between a first neuron, which is one of several neurons included in a spiking neural network, and a second neuron, which is a neuron that acquires the spike signal output from the first neuron when the first neuron fires. The system includes an elapsed time management unit that manages the elapsed time since the spike signal was output from the first neuron, The aforementioned bond strength control unit is When the second neuron fires, the update range, which is a reaction coefficient applied to the first change in the connection strength determined based on a first indicator other than the elapsed time, is determined based on the elapsed time managed by the elapsed time management unit, and the connection strength between the first neuron and the second neuron is updated based on the first change and the update range. A bonding strength updating device characterized by the following features. 　 The connection strength management unit manages the connection strength between a first neuron, which is one of several neurons included in the spiking neural network, and a second neuron, which is a neuron that acquires the spike signal output from the first neuron when the first neuron fires. The elapsed time management unit manages the elapsed time since the spike signal was output from the first neuron. The firing frequency management unit manages the firing frequency, which is the frequency at which the first neuron fires. The aforementioned bond strength control unit is When the second neuron fires, the connection strength between the first neuron and the second neuron is updated based on the firing frequency managed by the firing frequency management unit and the elapsed time managed by the elapsed time management unit. A method for updating bond strength, characterized by the following features. 　 The connection strength management unit performs a processing procedure to manage the connection strength between a first neuron, which is one of several neurons included in the spiking neural network, and a second neuron, which is a neuron that acquires the spike signal output from the first neuron when the first neuron fires. The elapsed time management unit performs a processing procedure to manage the elapsed time since the spike signal was output from the first neuron, A program that causes a computer to execute a processing procedure for managing the firing frequency, which is the frequency at which the first neuron fires, by having a firing frequency management unit. The aforementioned bond strength control unit is When the second neuron fires, the connection strength between the first neuron and the second neuron is updated based on the firing frequency managed by the firing frequency management unit and the elapsed time managed by the elapsed time management unit. A coupling strength update program characterized by the following features. 　 The first neuron included in the spiking neural network, A neuron included in the spiking neural network, the second neuron acquires a spike signal output from the first neuron when the first neuron fires, The system includes a connection strength management unit that manages the connection strength between the first neuron and the second neuron, The aforementioned bond strength control unit is The connection strength between the first neuron and the second neuron is updated based on the elapsed time since the spike signal was output from the first neuron and the firing frequency of the first neuron. A spiking neural network system characterized by the following:

Images