KR20210064767A - Spiking neuron model and method for generating spike thereof - Google Patents

Abstract

A spiking neuron model of a spiking neural network (SNN) comprises: a preprocessing amplifier outputting different currents in accordance with the level of weights set in a plurality of synapses for spikes transferred from the synapses; a differentiator differentiating a spike through the preprocessing amplifier to generate a step-function including only a tail (+); an integrator integrating a spike differentiated by the differentiator; a comparator comparing an output value of the integrator and a preset threshold (V_th) to output a fire signal if the output value of the integrator exceeds the threshold (V_th); and a spike generator generating a spike when a fire signal is outputted by the comparator to adjust the weights of the synapses. The spiking neuron model further comprises a diode unit removing a tail (-) from a spike transferred from the plurality of synapses, and transferring only the tail (+) to the preprocessing amplifier. Therefore, integration is not carried out in an opposite direction, and integrated probabilities are made equal in the entire section of the tail (+) of a corresponding spike to stably change the weight of a synapse.

Description

SPIKING NEURON MODEL AND METHOD FOR GENERATING SPIKE THEREOF

The present invention relates to a spiking neuron model, and more particularly, the weight of a synapse in a spiking neural network using a spiking neuron model It relates to a spiking neuron model for stably regulating and a method for generating spikes thereof.

FIG. 1A is a simplified and schematic diagram of a general spiking neural network. Referring to FIG. 1A , in the spiking neural network, a synapse 10 exists between neurons 20 and 30. And, when the spike output from the pre-neuron 20 passes through the synapse 10 and is transmitted to the post-neuron 30 , the post-neuron (post-neuron) ) (30), the pre-neuron (pre-neuron) (20) through an appropriate processing process to create a spike of the same shape as the spike received from the synapse (feed-back) to the synapse (10).

At this time, the signal transmission state varies according to the weight of the synapse 10 . In general, as the weight increases, the signal of the pre-neuron 20 is transferred to the post-neuron 30 . It has good transmission characteristics. Therefore, in order to control the transmission state of the signal, it is necessary to adjust the weight of the synapse (10).

On the other hand, as mentioned in the description with reference to Figure 1a, the synapse 10 has a pre-neuron (pre-neuron) (20), although the output spike (spike) passes, post-neuron (post-neuron) ( 30), the spikes fed back are also passed However, when the two spikes overlap, the voltage at both ends of the synapse 10 increases as much as it can control the weight, and the weight of the synapse 10 increases or decreases by the time difference between the two spikes. Therefore, the weight of the synapse 10 can be adjusted by adjusting the time difference between the two spikes.

Spiking time dependent plasticity (STDP) may refer to a learning method measured in neurons and synapses of real animals and humans. Looking at A and B of FIG. 1B , it can be seen that V _mem (indicated in green) increases (leaky integration) when two pre spikes (indicated in red and blue) enter. At this time, if the pre spike passes through a synapse of small weight (shown in blue) and is delivered to a post neuron, it can be integrated small, and when it passes through a synapse of a large weight (shown in red) and is delivered to a post neuron, it can be highly integrated. During leaky integration in this way, when the size of _{V mem becomes larger than V th} , the post neuron can fire. Fire operation _{initializes the value of V mem} and may mean generating a spike. The spike generated in this way can be used as a feed-forward spike (or pre spike) for leaky integration of the next neuron, and at the same time can be used as a feed-back spike (post spike) for learning the weight change of the previous synapse.

As the pre and post spikes overlap at the synapse, the weight of the synapse may change. Referring to C of FIG. 1b , it can be seen that the amount of change in the weight of the synapse (ΔW) varies according to the time difference between the pre spike and the post spike. Δt may be defined as a value obtained by subtracting the time at which the pre spike occurs from the time at which the post spike occurs as shown in C of FIG. 1B . That is, if Δt is positive, post spike occurs after pre spike, and if Δt is negative, it may mean that pre spike occurs after post spike.

If Δt is positive, it is judged that the pre spike affected the leaky integration and fire of the post spike, and the weight of the synapse connecting the neurons that generated each spike can be increased. At this time, as Δt is small, it is judged that it has a direct effect on fire, and the weight is increased. As Δt is large, it is helpful to fire, but it is judged that the influence is not large, so the weight can be increased a little.

If Δt is negative, since the post spike came in after the fire of the post neuron, it is judged that this pre spike is not a spike that helped fire, and the weight can be reduced. In this case, as Δt is smaller, it is determined that there is not much correlation with fire, and thus the weight can be greatly reduced. FIG. 2 shows a spike transmitted from a pre-neuron 20 (hereinafter, a pre-spike). The time at which the spike) occurred is 't _pre ', and the time at which the spike transmitted from the post-neuron 30 (hereinafter referred to as the post-spike) occurred is 't _post ', when the time difference between the two is 'Δt', the weight change is exemplified. In particular, FIG. 2 shows a case in which a 'tail (+)' of a pre-spike is a step function and a 'tail (-)' is a ramp function. may be an example of According to an embodiment, the a graph of FIG. 2 may be when Δt is both positive and small, the b graph may be when Δt is positive and Δt is large (however, when Δt is smaller than tail+), and the c graph shows that Δt is It is positive and may be when Δt is greater than tail+, and the d graph may be when Δt is negative. If only one of the pre and post spikes is applied to the synapse, the voltage across the synapse cannot exceed _{V t,i} , V _{t,d .} Therefore, the weight may not change. A, b, and c of FIG. 2 may be examples when Δt is a positive number (ie, the weight is increased). Biologically, since the magnitude of Δt can increase in the order of a, b, and c, ΔW can decrease in the order of a, b, and c.

Meanwhile, looking at the area indicated in yellow in FIG. 2 , the yellow area of b may include the yellow area of a. That is, the weight may vary more in the case of b than in the case of a. In case c, the maximum value of the voltage across the synapse may be lower than in cases a and b. Here, the weight change of the synapse deals with the case where the magnitude of the voltage applied to the synapse is more dominant than the duration (hatching area). In case of c, the amount of change in weight may be smaller than in the case of a and b. In case of d, Δt is negative, and in case of d, the pre spike and post spike are the same, so when Δt is positive, the magnitude of ΔW may be the same and only the sign may be opposite.

On the other hand, Figure 3 shows a weight (weight) variation (STDP: Spiking Time Dependent Plasticity) (ΔW) graph of the synapse 10 according to the time difference between the two spikes. In this case, the solid line represents the ideal (ie, biological) weight change amount (ΔW), and the dotted line indicates the actual weight change amount (ΔW). Referring to FIG. 3 , ideally, as the magnitude of the value (Δt) obtained by subtracting the time at which the pre-spike occurs from the time at which the post-spike occurs (t _{post ) increases, the weight (weight) increases.} ) the magnitude of the change is small. However, looking at the dotted line in FIG. 3 , it can be seen that the STDP graph is different from the ideal case on the basis of Δt=0 in the case of the actual change amount. That is, as in the case of FIG. 2 , the 'tail(+)' of the pre-spike is a step function, and the 'tail(-)' is a ramp function. ), it can be seen that the results differ from the ideal case during the 'tail (+)' time. Meanwhile, the regions indicated by a, b, c, and d in FIG. 3 may be positions corresponding to cases a, b, c, and d in FIG. 2 .

FIG. 4A is a diagram illustrating an example in which the shape of a spike is modified in order to make the STDP graph come out well even in the vicinity of Δt=0 in the diagram illustrated in FIG. 3 . That is, FIG. 4 is an example of a case in which both 'tail (+)' and 'tail (-)' of the pre-spike are ramp functions. The graph a of FIG. 4 may be when Δt is both positive and small, the graph b may be when Δt is positive and Δt is large (however, when Δt is smaller than tail+), the graph c is when Δt is positive and Δt is tail+ It may be greater than, and the d graph may be when Δt is negative.

According to FIG. 4A, in case c (ie, when Δt is greater than tail+), the pre and post spikes partially overlap, but the voltage applied to the synapse is not large enough, so the weight may not change. On the other hand, in the case of a and b, it can be seen that the weight increases, and in the case of a, Δt is smaller than in the case of b, so ΔW can be large. On the other hand, it is assumed that the magnitude of the voltage applied to the synapse is dominant over the duration (yellow region) for the change in the weight of the synapse in FIG. 4A . In FIG. 4a , the duration of b is larger than that of a, but the magnitude of the voltage applied to the synapse is greater in the case of a than in the case of b, so it can be confirmed that the biological STDP graph can be followed. As illustrated in FIG. 4A , when the pre-spike and the post-spike overlap, the voltage across the synapse decreases as the magnitude of Δt increases. Accordingly, as the magnitude of Δt increases, the change amount ΔW of the weight decreases. That is, in the case of using the spike as illustrated in FIG. 4A , as the size of Δt increases, the problem that the change amount ΔW of the weight also increases can be solved.

However, in this case, when integrating immediately after passing through a synapse, there may be a problem. That is, since the ramp function is integrated, the amount of integration increases as it goes back within the 'tail (+)' of one spike, and accordingly, the probability of fire increases as it goes back. . This increases the probability of a small change in the weight of a synapse when a certain spike fires a post-neuron through a synapse. Also, in this case, the ‘tail (-)’ is integrated and the proper integration is not performed. That is, if several spikes overlap when integrating the spikes, there is a problem that the integration may be reversed.

On the other hand, Figure 4b is a comparison of the ideal STDP graph and the STDP graph when using the spike shown in Figure 4a. Regions indicated by a, b, c, and d in FIG. 4B may be positions corresponding to cases a, b, c, and d in FIG. 4A.

Therefore, in order to solve the above problem, the present invention is such that, when integrating a spike that has passed through a synapse, only the tail (+) can be processed except for the tail (-) of the corresponding spike. Accordingly, an object of the present invention is to provide a spiking neuron model and a method for generating spikes thereof that prevent integration from being reversed.

In addition, in order to solve the above problem, the present invention stably changes the weight of the synapse by making the probability of integration in all sections of the tail (+) of the spike equal when integrating the spike that has passed through the synapse. An object of the present invention is to provide a spiking neuron model and a method for generating spikes thereof.

In addition, an object of the present invention is to provide a neuron model and a method for generating spikes thereof that make a graph of the amount of change in weight of a synapse close to an ideal one.

In order to achieve the above object, the spiking neuron model provided by the present invention is a plurality of pre-neurons and a plurality of post-neurons through a plurality of synapses are connected and , In the spiking neuron model of a spiking neural network (SNN) in which the synapse transmits a spike generated in a pre-neuron to a post-neuron, the plurality of a preprocessing amplifier for outputting different currents according to the magnitude of the weight set for each of the synapses for each of the spikes transmitted from the synapses; a differentiator for differentiating the spike through the pre-amplifier to generate a step-function including only a tail (+); an integrator for integrating the spike differentiated in the differentiator; When compared to the output value with a preset threshold value (V _th) of said integrator to an output value of the integrator exceeds the threshold value (V _th) a comparator for outputting a fire (Fire) signal; and a spike generator configured to adjust the weight of the synapse by generating a spike when a Fire signal is output from the comparator.

Preferably, the spiking neuron model further comprises a diode unit that removes a tail (-) from a spike transmitted from the plurality of synapses, and transmits only a tail (+) to the pre-amplifier. can

Preferably, the differentiator may determine the number of differentiators according to the order of the spikes transmitted from the plurality of synapses.

Preferably, the differentiator may include n differentiators when the order of the spikes transmitted from the plurality of synapses is an nth-order function.

Preferably, the integrator may stop the integration in response to a Fire signal of the comparator, and maintain the integration stop for a time during which the spike generator generates a spike.

In addition, in order to achieve the above object, the method for generating a spike of a spiking neuron model provided by the present invention is a spiking neuron model that receives a spike through a plurality of synapses in a spiking neural network (SNN). Model) in the spike generation method, a differentiation step of differentiating the spikes transmitted from the plurality of synapses and changing it into a step-function; an integration step of integrating the changed spike; a comparison step of comparing the integration result with a preset threshold value (V _th ) and outputting a Fire signal when the integration result exceeds the threshold value (V _{th );} and generating a spike when the Fire signal is output to adjust the weight of the synapse.

Preferably, the method may further include a pre-processing step of removing a tail (-) from the spike transmitted from the plurality of synapses.

Preferably, the differentiation step may determine the number of differentiation according to the order of the spikes transmitted from the plurality of synapses.

Preferably, in the differentiation step, when the spike transmitted from the plurality of synapses is an nth-order function, the differentiation step may be repeated n times.

Preferably, in the integrating step, the integration may be stopped in response to the Fire signal of the comparing step, and a spike may be generated in the spike generating step.

The spiking neuron model and the method for generating a spike thereof provided in the present invention are such that, when integrating a spike that has passed through a synapse, only the tail (+) can be processed, excluding the tail (-) of the corresponding spike. By doing so, there is an advantage in that the integration is not reversed. In addition, the present invention has the advantage of stably changing the weight of the synapse by making the probability of integration in all the tail (+) sections of the corresponding spike the same when integrating the spike that has passed through the synapse. For this reason, the present invention provides a neuron model similar to a neuron in a real body by making the weight change graph of the synapse close to the ideal, and thereby, a spiking neural network that operates stably (SNN: Spiking Neural Network) can be provided.

1A is a simplified and schematic diagram of a general spiking neural network.
1B is a view for explaining the background of the present invention.
FIG. 2 is a schematic diagram showing a state of overlapping spikes according to a time difference between pre-spike and post-spike for a pre-spike form according to an exemplary embodiment of the related art; FIG. It is a drawing.
3 is a graph showing the amount of change in the weight of the synapse according to the generation time difference between the pre-spike and the post-spike.
4A is a schematic diagram illustrating a spike overlapping state according to a time difference between a pre-spike and a post-spike for a pre-spike form according to another exemplary embodiment; It is a drawing.
Figure 4b is a comparison of the ideal STDP graph and the STDP graph when using the spike shown in Figure 4a.
5 is a schematic block diagram of a spiking neuron model according to an embodiment of the present invention.
6A to 6D are diagrams illustrating an embodiment in which the spiking neuron model of the present invention is variously applied.
7 is a flowchart of a method for generating a spike in a spiking neuron model according to an embodiment of the present invention.
8A is a diagram illustrating the overlapping state of the spikes according to the time difference between the pre-spike and the post-spike for the pre-spike form according to an embodiment of the present invention. the drawing shown.
8B is a comparison of the ideal STDP graph and the STDP graph when the spike shown in FIG. 8A is used.

Hereinafter, an embodiment of the present invention will be described with reference to the accompanying drawings, but it will be described in detail so that a person of ordinary skill in the art to which the present invention pertains can easily practice the present invention. However, the present invention may be embodied in various different forms and is not limited to the embodiments described herein. On the other hand, in order to clearly explain the present invention in the drawings, parts irrelevant to the description are omitted, and similar reference numerals are attached to similar parts throughout the specification. In addition, even if the detailed description is omitted, descriptions of parts that can be easily understood by those skilled in the art are omitted.

Throughout the specification and claims, when a part includes a certain element, it means that other elements may be further included, rather than excluding other elements, unless otherwise stated.

5 is a schematic block diagram of a spiking neuron model according to an embodiment of the present invention. Referring to FIG. 5 , the spiking neuron model 300 according to an embodiment of the present invention includes a preprocessing amplifier 310 , a differentiator/inverting amplifier 320 , a differentiator 330 , an integrator 340 , and a comparator 350 . and a spike generator 360 .

The preprocessing amplifier 310 collects and outputs the current of several spikes transferred from a plurality of synapses 100 . That is, the pre-processing amplifier 310 for each of the plurality of spikes transmitted from the plurality of synapses 100, the next stage (stage) different currents according to the size of the weight set in each of the synapses 100. ) can be printed. At this time, since each of the synapses 100 is modeled as a single resistor, the preprocessing amplifier 100 includes synapses 100 and one amplifier (op amp) 311 and an additional resistor ( 312) may be configured.

The differentiator/inverting amplifier 320 is selectively included according to the type of the input spike. For example, if the input spike is a quadratic function, the differentiator 322 may be selected, otherwise the inverting amplifier 321 may be selected.

Differentiator 330 differentiates the spike through preprocessor amplifier 310 . In particular, the differentiator 330 generates a step-function including only the tail (+) of the input spike. At this time, the differentiator 330 may determine the number of differentiators according to the order of the tail (+) of the spike transmitted from the plurality of synapses 100 . That is, the differentiator 330 may include n differentiators when the order of the tail (+) of the spike transmitted from the plurality of synapses 100 is an nth-order function. For example, when the order of the tail (+) of the spike transmitted from a plurality of synapses 100 is a quadratic function, it includes two differentiators, and the tail of the spike transmitted from a plurality of synapses 100 ( If the order of tail)(+) is a linear function, one differentiator may be included. To this end, the differentiator 330 may include an amplifier (op amp) 331 , a capacitor 332 , and a resistor 333 .

The integrator 340 integrates the spike differentiated by the differentiator 330 . In particular, the integrator 340 stops integrating in response to a Fire signal of the comparator 350 . Then, the integration stop is continuously maintained during the time the spike generator 360 generates the spike. To this end, the integrator 340 may include an amplifier (op amp) 341 , a capacitor 342 , and a resistor 343 . In addition, the integrator 340 further includes a switch 344 whose on/off is controlled by the Fire signal, and performs integration when the switch 344 is 'on', and when 'off' It can be controlled to stop the integration.

The comparator 350 stores the preset threshold (V _th), and compares the output value and the threshold value (V _th) of the integrator 340. In addition, when the output value of the integrator 340 exceeds the threshold V _th , a Fire signal is output.

The spike generator 360 generates a spike when a Fire signal is output from the comparator 350 . Then, the spike is fed back to adjust the weight of the synapses 100 .

Meanwhile, the spiking neuron model may further include a diode unit 200 , wherein the diode unit 200 removes a tail (-) from a spike transmitted from a plurality of synapses 100 , and Only (tail) (+) is delivered to the pre-aerator 310 . This is to reduce an error in which the integration may be reversed if several spikes overlap when integrating the spikes in the integrator 340 .

6A to 6D are diagrams illustrating an embodiment in which the spiking neuron model of the present invention is variously applied.

First, FIG. 6a shows a first pre-spike in which both 'tail (+)' and 'tail (-)' are ramp functions through the 3-terminal synapse 100a. -spike) (P1) is a diagram illustrating an example (hereinafter, referred to as a 'first embodiment'). Referring to FIG. 6A , in the first embodiment, the first pre-spike P1 passes through a synapse 100a and then passes through a diode 200a. tail) (-) disappears, and only tail (+) enters the neuron 300a. In addition, the pre-spike P1 with only the tail (+) remaining is transferred to the differentiator 330a through the pre-processing amplifier 310a and the inverting amplifier 320a. Then, the differentiator 330a differentiates it, and changes tail (+), which is a ramp function, into a step function (aka, square wave). The integrator 340a integrates the step function output through the differentiator 330a in this way. At this time, the integrator 340a integrates the signal transformed by the step function, so that the probability of being fired for all times within the tail (+) becomes the same.

On the other hand, the differentiator in the integrated circuit has a capacitor at the input side of the amplifier (amp), and with this structure, a spike at a synapse cannot be directly received. This is because a different magnitude of current must be received according to the weight of the synapse. Therefore, by attaching an additional circuit to the front-end of the post-neuron, the size of the current can be received differently depending on the weight of the synapse, regardless of the time when multiple spikes are received. Should be. This method cannot be used in a 2-terminal synapse device because a diode must be used to ignore the tail (-) when integrating, and the 3-terminal synapse (3) -terminal synapse) Available only on device.

6B to 6D show a second pre-spike (P2) in which 'tail (+)' is a quadratic function and 'tail (-)' is a ramp function. ) is a diagram showing examples of a case in which the second pre-spike (P2) is transmitted, and the types of synapses or configurations of peripheral devices are different.

6b is an example of a case in which the second pre-spike (P2) passing through the 3-terminal synapse 100b is transmitted to the pre-amplifier 310b through the diode 200b (hereinafter, ' Referring to the second embodiment '), FIG. 6c shows that the second pre-spike (P2) passing through the synapse 100c, which is a three-terminal, does not pass through a diode, and the preprocessing amplifier 310c ) shows an example (hereinafter, referred to as the 'third embodiment') in the case of being transferred, and Figure 6d is the second pre-spike (pre-spike) (P2) that has passed through the 2-terminal synapse 100d. ) is transmitted to the preprocessing amplifier 310d through a separate device (eg, a switch, etc.) 400d (hereinafter, referred to as a 'fourth embodiment') is shown.

At this time, the second pre-spike P2 is another spike shape that can make the STDP graph come out well even in the vicinity of Δt=0. That is, the second pre-spike (P2), like the first pre-spike (P1), when the spikes overlap, as the size of Δt increases, at both ends of the synapse The applied voltage becomes smaller, and accordingly, the amount of change in the weight becomes smaller. The difference from the first pre-spike P1 is that the tail (+) of the spike is not a ramp function but a quadratic function.

Therefore, if the second pre-spike (P2) is used, there is no need to use a diode after synapse. In particular, as in the example of FIGS. 6B and 6C , a diode is not necessarily required in a 3-terminal synapse. For example, in the example of Figure 6b, including a diode 200b, the second pre-spike (P2) that has passed through the synapse 100b is passed through the diode 200b to the preprocessing amplifier 310b. Although the case of transmission is illustrated, the second pre-spike (P2) that has passed through the synapse 100c as in FIG. 6c does not pass through the diode and is directly transmitted to the preprocessing amplifier 310c. Instead, the differentiation is performed twice instead of once after passing through the preprocessing amplifier 310b or 310c. That is, in the second embodiment illustrated in FIG. 6B , the 'tail (-)' of the second pre-spike P2 disappears while passing through the diode 200b, so 2 While passing through the differentiators 320b and 330b, a step function including only a tail (+) is output, and in the third embodiment illustrated in FIG. 6C , the second pre-spike (pre) is output. -spike)(P2) goes through the first differentiator 320c, the quadratic function 'tail(+)' becomes a ramp function, and the ramp function 'tail(-)' becomes a step become a function In addition, while passing through the second differentiator 330c, the ramp function 'tail(+)' becomes a step function, and the step function 'tail(-)' becomes '0'. As a result, the second pre-spike P2 is transformed into a step function including only a 'tail (+)' while passing through the first and second differentiators 320c and 330c.

If integration is performed after the differentiator is performed twice as described above, only the tail (+) is properly integrated in the integrators 340b and 340c of FIGS. 6B and 6C.

At this time, if a diode is used, the tail (-) can be easily eliminated, but in order to equalize the size integrated within the tail (+), even when there is a diode, there are two differentiators. need.

Even in the case of Figure 6d, including the 2-terminal synapse (100d), the second pre-spike (P2) is the first and the second differentiator while passing through the differentiator twice. While passing through (320d and 330d), it is transformed into a step function including only a 'tail (+)'.

7 is a flowchart of a method for generating a spike in a spiking neuron model according to an embodiment of the present invention. 5 and 7 , in the method for generating a spike of a spiking neuron model according to an embodiment of the present invention, a spiking neuron model that receives a spike through a plurality of synapses in a spiking neural network (SNN) In the spike generation method of (Spiking Neuron Model), in step S110, the diode 200 removes a tail (-) from the spike transmitted from the plurality of synapses. In step S120, the input spike passed through step S110 is changed to a square wave. That is, in step S120, the differentiator 330 differentiates the spike from which the tail (-) has been removed and changes it to a step-function (aka, spherical stroke). At this time, in step S120, it is preferable to determine the number of differentiations according to the order of the tail (+) of the spike transmitted from the plurality of synapses. That is, in step S120, when the tail (+) of the spike transmitted from the plurality of synapses is an nth-order function, the differentiation step may be repeated n times. For example, when the tail (+) of the spike transmitted from the plurality of synapses is a quadratic function, the differentiation step is repeated twice.

In step S130 , the integrator 340 integrates the changed spike, and in step S140 , the comparator 350 compares the integration result with a preset threshold V _th . When it is determined in step S140 that the integration result exceeds the threshold V _th , in step S150 , the comparator 350 outputs a Fire signal, and in step S160 , the spike generator 360 . adjusts the weight of the synapse by generating a spike.

On the other hand, if a Fire signal is output in step S150, the integrator 340 stops integration in response to the Fire signal, and it is preferable to maintain the integration stop during the time generated in step S160. Do.

8A is a diagram illustrating the overlapping state of the spikes according to the time difference between the pre-spike and the post-spike for the pre-spike form according to an embodiment of the present invention. the drawing shown. That is, FIG. 8A is an example of a case in which 'tail (+)' of a pre-spike is a quadratic function and 'tail (-)' is a ramp function. . According to an embodiment, graph a of FIG. 8A may be when Δt is positive and small, graph b may be when Δt is positive and Δt is large (however, when Δt is smaller than tail+), and graph c shows that Δt is It may be positive and Δt may be greater than tail+, and the d graph may be when Δt is negative. According to FIG. 8a, in case c (that is, when Δt is greater than tail+), the pre spike and the post spike partially overlap, but the voltage applied to the synapse is not large enough, so the weight may not change. On the other hand, in the case of a and b, it can be seen that the weight increases, and since Δt is smaller in case of a than in case of b, ΔW may be large. On the other hand, it is assumed that the magnitude of the voltage applied to the synapse is dominant over the duration (yellow region) for the change in the weight of the synapse in FIG. 8A . In FIG. 8a, the duration of b is larger than that of a, but the magnitude of the voltage applied to the synapse is greater in case a than in case b, so it can be confirmed that the biological STDP graph can be followed.

Referring to FIG. 8A , when the pre-spike and the post-spike overlap, the voltage across the synapse decreases as the magnitude of Δt increases. Accordingly, as the magnitude of Δt increases, the change amount ΔW of the weight decreases. That is, when the spike as illustrated in FIG. 8 is used, it is possible to solve the problem that as the size of Δt increases, the change amount ΔW of the weight also increases. In addition, in this case, as mentioned in the description with reference to FIGS. 6A to 6D , by integrating the transformed signal with a step function in the integrator, fire is fired for all times within the tail (+). ) is equally probable. Therefore, when the present invention is used, there is an advantage that the weight of the synapse can be changed stably. That is, when following the spike shown in FIG. 8A, the spike can be processed (leaky integration & fire) based on the differentiation of the CMOS neuron model without the need to make an additional diode at the synapse.

8B is a comparison of the ideal STDP graph and the STDP graph when the spike shown in FIG. 8A is used. Regions indicated by a, b, c, and d in FIG. 8B may be positions corresponding to cases a, b, c, and d in FIG. 8A. Meanwhile, since tail+ is a quadratic function in the graph of FIG. 8B , the slope according to the position a and the position b in FIG. 8B may be greater than the slope according to the position a and the position b in FIG.

In the exemplary system described above, the methods are described on the basis of a flowchart as a series of steps or blocks, but the present invention is not limited to the order of steps, and some steps may occur in a different order or concurrently with other steps as described above. can

In addition, those skilled in the art will understand that the steps shown in the flowchart are not exhaustive and that other steps may be included or that one or more steps in the flowchart may be deleted without affecting the scope of the present invention.

Claims (10)

A plurality of pre-neurons and a plurality of post-neurons are connected through a plurality of synapses, and the synapse is a spike (pre-neuron) generated in In the spiking neuron model of the spiking neural network (SNN) that delivers the spike) to the post-neuron,
a preprocessing amplifier for outputting different currents according to the magnitude of the weight set for each of the synapses with respect to each of the spikes transmitted from the plurality of synapses;
a differentiator for differentiating the spike through the pre-amplifier to generate a step-function including only a tail (+);
an integrator for integrating the spike differentiated in the differentiator;
When compared to the output value with a preset threshold value (V _th) of said integrator to an output value of the integrator exceeds the threshold value (V _th) a comparator for outputting a fire (Fire) signal; and
The spiking neuron model, characterized in that it comprises a spike generator that adjusts the weight of the synapse by generating a spike when a Fire signal is output from the comparator. According to claim 1,
The spiking neuron model, characterized in that it further comprises a diode unit that removes a tail (-) from the spike transmitted from the plurality of synapses, and transmits only the tail (+) to the pre-amplifier. The method of claim 1, wherein the differentiator
A spiking neuron model, characterized in that the number of differentiators is determined according to the order of the tail (+) of the spike transmitted from the plurality of synapses. The method of claim 3, wherein the differentiator
When the order of the tail (+) of the spike transmitted from the plurality of synapses is an n-order function,
A spiking neuron model comprising n differentiators. The method according to any one of claims 1 to 3, wherein the integrator is
In response to the Fire signal of the comparator, the integration is stopped,
The spiking neuron model according to claim 1, wherein the integration interruption is maintained for a period of time during which the spike is generated in the spike generator. In the spike generation method of a spiking neuron model receiving a spike through a plurality of synapses in a spiking neural network (SNN),
Differentiation step of differentiating the spike transmitted from the plurality of synapses and changing it into a step-function;
an integration step of integrating the changed spike;
a comparison step of comparing the integration result with a preset threshold value (V _th ) and outputting a Fire signal when the integration result exceeds the threshold value (V _{th );} and
and a spike generating step of generating a spike when the Fire signal is output and adjusting the weight of the synapse. 7. The method of claim 6,
The method for generating spikes in a spiking neuron model, characterized in that it further comprises a pre-processing step of removing a tail (-) from the spikes transmitted from the plurality of synapses. 7. The method of claim 6, wherein the differentiating step
A method of generating a spike in a spiking neuron model, characterized in that the number of differentiations is determined according to the order of the tail (+) of the spike transmitted from the plurality of synapses. The method of claim 8, wherein the differentiating step
When the tail (+) of the spike transmitted from the plurality of synapses is an nth-order function,
A method for generating spikes in a spiking neuron model, characterized in that repeating the differentiation step n times. The method according to any one of claims 6 to 8, wherein the integrating step is
In response to the Fire signal of the comparison step, the integration is stopped,
The method for generating a spike in a spiking neuron model, characterized in that in the step of generating a spike, the integration stop is maintained for a time for generating a spike.

Images