KR20210070589A - Apparatus for learning a spiking neural network - Google Patents

Abstract

The present invention provides an apparatus for training a spiking neural network, including: an up-change amount acquisition unit for updating an up-change amount whenever an input spike is applied, or whenever a plurality of time sections divided by a quantization spike-timing-dependent plasticity (STDP) function are reached after the input spike is applied, based on the quantization STDP function in which a change amount for adjusting a weight is divided by the time sections and set in advance based on a time difference between the input spike and an output spike, in order to train the spiking neural network for accumulatively adding a weight corresponding to the applied input spike among a plurality of pre-stored weights to spike the output spike when an acquired accumulative spike signal is greater than or equal to a predetermined spiking threshold; a down-change amount acquisition unit for updating a down-change amount whenever the input spike is applied, or whenever the time sections divided by the quantization STDP function are reached after the input spike is applied; a change amount acquisition unit for acquiring a weight change amount by using the up-change amount and the down-change amount; and a weight update unit for updating the weight by adding the weight change amount to the weight. Accordingly, the weight change amount for training the spiking neural network is easily acquired in real time.

Description

A learning device for a spiking neural network {APPARATUS FOR LEARNING A SPIKING NEURAL NETWORK}

The present invention relates to a learning apparatus for a spiking neural network, and to a learning apparatus for a spiking neural network that implements spike timing dependent plasticity in hardware.

Neuromorphic technology is a technology to mimic the human neural structure in hardware. It is proposed to overcome the limitations that the existing computing architecture is very low in efficiency and consumes a lot of power compared to humans in performing cognitive processing functions. became

A typical example of neuromorphic technology is a spiking neural network (SNN). SNN is a neural network designed by mimicking the characteristics that the human brain has a neuron-synapse structure, and a synapse between neurons transmits information as a spike-type electrical signal. These SNNs process information based on the time difference at which the spike signal is transmitted.

SNN is also a kind of artificial neural network, and learning to update weights must be performed, and a spike-timing-dependent plasticity (STDP) technique is generally used for SNN learning. Also, in the STDP technique, the all-to-all STDP technique, which calculates the weight change amount for updating the weight based on the occurrence time information of all input and output spikes that occurred within the training window period of a predetermined size, is well known. have.

However, the all-to-all STDP technique requires a large storage capacity and is implemented in hardware because the learning device for learning the SNN must store both the time information on the input and output spikes generated during the time period specified by the learning window. have difficulty doing

For this reason, the nearest STDP technique having a simple structure has been mainly used in the past. In the nearest STDP technique, when an input spike or an output spike occurs, the counter starts counting in response to a clock signal, and time information about the time when the input spike or output spike occurs is stored in the spike history. Then, when a spike that becomes a trigger is applied, the counting value of the counter is substituted into the lookup table prepared in advance, the weight change amount is obtained from the lookup table, and the weight is updated according to the obtained weight change amount. Here, the spike that is triggered for the input spike is the output spike, and the spike that is triggered for the output spike is the input spike. That is, the nearest STDP technique is a method of updating weights by referring only to the occurrence time information of other spikes generated at the closest time. However, since the weights are updated considering only the nearest spike, the nearest STDP technique has a limitation in that the learning performance is lower than that of the all-to-all STDP technique.

Korean Patent Publication No. 10-2018-0062934 (published on June 11, 2018)

SUMMARY OF THE INVENTION It is an object of the present invention to provide a learning apparatus for a spiking neural network that can easily acquire a weight change amount for learning a spiking neural network in real time.

Another object of the present invention is to provide a learning apparatus for a spiking neural network capable of learning a spiking neural network according to an all-to-all STDP technique by storing a spike history with simple hardware.

Another object of the present invention is to provide a learning apparatus for a spiking neural network capable of learning a spiking neural network with a more efficient hardware configuration by scaling a time interval of spike occurrence information stored in a spike history.

In order to achieve the above object, an apparatus for learning a spiking neural network according to an embodiment of the present invention includes a utterance threshold in which an accumulated spike signal obtained by accumulatively adding a weight corresponding to an input spike to be applied among a plurality of pre-stored weights is a predetermined utterance threshold. In order to train a spiking neural network that fires an output spike when it is greater than or equal to a value, the amount of change for adjusting the weight based on the time difference between the input spike and the output spike is divided into a plurality of time sections, and a preset quantization STDP function an up change amount obtaining unit that updates the up change amount each time the input spike is applied or a plurality of time intervals divided by the quantization STDP function after the input spike is applied, based on ? a down change amount obtaining unit that updates the down change amount whenever the input spike is applied or a plurality of time intervals divided by the quantization STDP function after the input spike is applied; a change amount obtaining unit for obtaining a weight change amount using the up change amount and the down change amount; and a weight update unit that updates the weight by adding the weight change amount to the weight.

a plurality of interval delay units connected in series to output the up change amount obtaining unit by delaying the input spike by a predetermined time interval when the input spike is applied; and when the input spike is applied, the pre-stored up-change amount is increased by the maximum value set by the quantization STDP function, and whenever the delayed input spike is output from each of the plurality of section delay units, a predetermined value set by the quantization STDP function is output. and a change amount calculation unit for decreasing the up change amount by a change unit.

a plurality of section delay units connected in series to the down change amount obtaining section to delay and output the output spikes by a predetermined time section, respectively, when the output spikes are applied; and when the output spike is applied, the pre-stored down change amount is increased by the maximum value set by the quantization STDP function, and whenever the delayed output spike is output from each of the plurality of section delay units, the group set by the quantization STDP function It may include a change amount calculation unit for decreasing the down change amount by a specified change unit.

Each of the plurality of section delay units is serially connected in a number corresponding to a corresponding time section among a plurality of time sections divided by the quantization STDP function, and receives the input spike in response to a clock of a predetermined period, It may include multiple flip-flops carrying previously applied input spikes.

Each of the plurality of section delay units includes a plurality of flip-flops for receiving the input spike or transferring the previously applied input spike in response to a scaling clock in which the cycle of the clock is increased according to a predetermined time step scaling ratio. , the plurality of flip-flops may be included in a reduced number by the time step scaling ratio in a corresponding time interval.

The change amount obtaining unit may obtain a weight change amount by subtracting the down change amount from the up change amount.

When the input spike or the output spike is applied, the weight update unit may update the weight by adding the weight change amount to a corresponding weight.

Therefore, the learning apparatus for a spiking neural network according to an embodiment of the present invention stores the spike history in a simple configuration using a plurality of flip-flops, and facilitates the amount of weight change for learning the spiking neural network based on the stored spike history in real time It can be acquired to train a spiking neural network. In addition, by scaling the time interval of the spike occurrence information stored in the spike history, it is possible to train the spiking neural network while maintaining the learning performance with a simpler hardware configuration.

1 shows a schematic structure of a spiking neural network according to an embodiment of the present invention.
2 and 3 are diagrams for explaining the operation of the weighted synapse and the output neuron of FIG. 1 .
4 shows an example of a quantized spike timing dependent plasticity function.
FIG. 5 shows a detailed configuration of the weight learning unit of FIG. 1 .
FIG. 6 shows a schematic structure of the change amount obtaining unit of FIG. 5 .
7 shows an example of a detailed configuration of the interval delay unit of FIG. 6 .
FIG. 8 shows another example of the detailed configuration of the interval delay unit of FIG. 6 .
9 shows the results of analyzing the learning accuracy according to the time step scaling ratio.
10 shows a result learned by applying time step scaling.

In order to fully understand the present invention, the operational advantages of the present invention, and the objects achieved by the practice of the present invention, reference should be made to the accompanying drawings illustrating preferred embodiments of the present invention and the contents described in the accompanying drawings.

Hereinafter, the present invention will be described in detail by describing preferred embodiments of the present invention with reference to the accompanying drawings. However, the present invention may be embodied in various different forms, and is not limited to the described embodiments. In addition, in order to clearly describe the present invention, parts irrelevant to the description are omitted, and the same reference numerals in the drawings indicate the same members.

Throughout the specification, when a part "includes" a certain component, it means that other components may be further included, rather than excluding other components unless otherwise stated. In addition, terms such as "... unit", "... group", "module", and "block" described in the specification mean a unit that processes at least one function or operation, which is hardware, software, or hardware. and a combination of software.

1 shows a schematic structure of a spiking neural network according to an embodiment of the present invention, and FIGS. 2 and 3 are diagrams for explaining the operation of a weighted synapse and an output neuron of FIG. 1 . and FIG. 4 shows an example of a quantized spike timing dependent plasticity function.

Referring to FIG. 1 , a spiking neural network (SNN) includes an input neuron 100 , a weight synapse 200 , an output neuron 300 , and a weight learner 400 .

The input neuron 100 encodes input data according to a predetermined method, converts it into a plurality of input spikes, and transmits the converted plurality of input spikes to the weighted synapse 200 .

The input neuron 100 may be converted into a plurality of input spikes each generated with a time difference according to the value of the input data. For example, the input neuron 100 may receive the pixel values of each pixel corresponding to the image as input data, and may output an input spike generated with a time difference according to the applied pixel value.

Methods for the input neuron 100 to convert input data into input spikes have been studied in various ways, and for example, an input spike may be generated based on a probability. In this case, the input neuron of the input neuron 100 may generate and output a plurality of input spikes multiple times with a time difference based on the input data, rather than converting the input data into one input spike as described above. .

That is, as shown in FIG. 2 , a plurality of input spikes may be generated several times at different times for each input data.

Here, the input neuron 100 generates an input spike in the form of binary data. For example, it may be set to output a value of 1 when an input spike occurs, and output a value of 0 when no input spike occurs.

Also, the input neuron 100 may be configured to receive a plurality of input spikes in a binary data format. In this case, the input neuron 100 may be implemented as a buffer that buffers and outputs the plurality of input spikes.

In some cases, the input neuron 100 may receive an output spike output from the output neuron 300 of another neuron and output it as an input spike. In this case, if the matching relationship between the input neuron 100 and the output neuron 300 is different, a Synapse Routing Table (SRT) defining a correspondence between the input neuron 100 and the output neuron 300 may be used.

The weight synapse 200 is a configuration that mimics the synapse of the brain in the SNN, and may have a structure similar to a memory array storing a plurality of weights. For example, the weighted synapse 200 may be implemented as a static random access memory (SRAM) or the like.

The weight synapse 200 stores a plurality of weights, and when an input spike is applied from the input neuron 100 , a weight corresponding to the applied input spike is output to the output neuron 300 . Here, a plurality of weights stored in the weight synapse 200 are updated by the weight learning unit 400 .

The weight synapse 200 repeatedly outputs a corresponding weight to the output neuron 300 in response to each of a plurality of input spikes generated and applied with a time difference from the input neuron 100 . That is, whenever an input spike occurs in the input neuron 100 , a corresponding weight among the stored weights is output to the output neuron 300 .

The output neuron 300 generates a plurality of output spikes based on the cumulative sum of weights output from the weighted synapse 200 . The output neuron 300 may be configured based on a Leaky-Integrate and Fire (LIF) neuron model.

In SNN, a LIF neuron model that mimics the operation characteristics of a human neural network is mainly used. The LIF neuron model integrates weights according to input spikes, and when the accumulated weights are above a predetermined reference level, fire ) to generate a new input spike to be transmitted to the next neuron, and if the input spike is not transmitted, the accumulated intensity of the input spike leaks and becomes weaker as time goes by.

Accordingly, the output neuron 300 constructed based on the LIF neuron model also adds and accumulates the weights applied from the weighted synapses 200, and when the accumulated weight is greater than or equal to a predetermined firing threshold, an output spike is generated and output. And the output neuron 300 may be configured to gradually decrease the cumulative weight over time.

The output neuron 300 sums all _{the weights w i} applied from the weighted synapses 200 whenever at least one input spike is generated. 2, the weight synapses 200 is applied a plurality of input spike (x ₁ ~ x _N), the inputted input spike (x ₁ ~ x _N) weight corresponding to each (w ₁ ~ w _N ), and at least one output neuron receives a weighted input spike (x _i w _i ) with weight and adds them all.

As described above, the input spike (x ₁ to x _N ) is generated in the form of a binary value, and the weight synapse 200 transmits a weight corresponding to the _{input spike (x 1 to x N} ) to the output neuron 300 . Therefore, the sum (∑x _i w _i ) of the weighted input spikes (x _i w _i ) can be viewed as the sum of the weights transmitted to the output neuron 300 .

And the output neuron 300 further adds the previously acquired cumulative spike signal u(t-1) to the sum (∑x _i w _i ) of the weighted input spikes x _i w _{i .} At this time, the output neuron 300 _{does not add the previously acquired cumulative spike signal (u(t-1)) to the sum of input spikes (∑x i} w _i ) as it is, but the previously acquired cumulative spike signal u(t-1). )) to the leakage factor λ (here, a real number where 0 < λ < 1) is added to the leakage cumulative spike signal λu(t-1) to obtain the cumulative spike signal u(t).

Here, weighting the leakage factor (λ) to the previous cumulative spike signal (u(t-1)) reflects the characteristic that neurons in the brain gradually lose the information they remember over time, and the sum of the input spikes ( Adding the previous leakage accumulated spike signal (λu(t-1)) to ∑x _i w _{i ) reflects the characteristic of information being remembered through repetitive stimulation.} That is, it performs the functions of leakage and integration in the LIF neuron model.

The voltage level of the cumulative spike signal u(t) is generally referred to as a membrane potential, and as shown in FIG. 3 , when the membrane potential is _{lower than the firing threshold V th} , the output spike In general, when the voltage level of the cumulative spike signal u(t) _{is higher than the ignition threshold V th} , the voltage level of the accumulated spike signal u(t) is output by firing an output spike.

In some cases, the output neuron 300 may adjust the number of firings of the output spike by counting the number of times the output spike is fired _{and adjusting the firing threshold (V th ) based on the counted number of firings of the output spike.}

The weight learning unit 400 learns a plurality of weights of the weight synapses 200 according to a spike-timing-dependent plasticity (STDP) technique in a configuration corresponding to a learning device for learning the SNN. The weight learning unit 400 for learning the weights according to the STDP technique is based on the time information when the input spike is input from the input neuron 100 to the weight synapse 200 for a previously predetermined period based on the time when the output spike is generated. to increase the weight. On the other hand, when an input spike is input from the input neuron 100 to the weighted synapse 200 , the weight learning unit 400 reduces the weight based on the time information at which the output spike occurred during the previous predetermined period when the input spike was input. make it That is, the weight learning unit 400 increases or decreases the weight based on the generation time of the applied input spike and the output spike.

In this case, when the time difference between the time when the output spike is generated and the time when the previous input spike is input is short, the weight learning unit 400 increases the weight more so that the correlation between the input spike and the output spike increases. Conversely, if the time difference from the time when the input spike is input to the time when the previous output spike is generated is short, the weight is further reduced so that the correlation between the input spike and the output spike is lowered. This is because the output spike should be output in response to the previously inputted input spike, whereas the inputted input spike should be input regardless of the previously generated output spike as much as possible.

However, although the STDP of the brain has an amount of change that is analogically variable with time, it is very difficult to implement this in the weight learning unit 400 . Accordingly, as shown in FIG. 4 , the weight learning unit 400 in the SNN adjusts the weight change amount Δw of the weight w based on the quantization STDP function in the weight learning unit 400 implemented as a digital circuit.

In the quantized STDP function graph shown in FIG. 4 , the x-axis _{represents the time difference between the time point at which the output spike (OS) occurs (t OS} ) and the time point at which the input spike (IS) occurs (t _IS ), and the y-axis represents the weight ( w) represents the weight change amount (Δw). In the quantization STDP function graph of FIG. 4, since the x-axis _{represents the time difference between the time points (t IS} ) at which the input spike (IS) is generated, both the time point at which the input spike (IS) is applied and the time point at which the output spike (OS) is applied can be applied based on Here, the value of the x-axis is a unit time that can input and output spikes in SNN implemented as a digital circuit, and is a unit clock (CLK) (also called a cycle). In addition, in FIG. 4 , the value in the y-axis direction represents a predetermined change unit capable of adjusting the weight w having a digital value.

If the position at (0, 0) is the time when the output spike (OS) occurs, the region with a positive value on the x-axis (x > 0) is the input spike (IS) before the output spike (OS) occurs. Represents the weight change amount (Δw) according to the time of occurrence, and the region having a negative value (x < 0) is the weight change amount (Δw) according to the time when the input spike (IS) is generated after the output spike (OS) is generated indicates

On the other hand, if the position at (0, 0) is the point at which the input spike (IS) occurs, the region with a positive value (x > 0) on the x-axis is the output spike (OS) after the input spike (IS) occurs. It represents the weight change amount (Δw) according to the time of occurrence, and the region having a negative value (x < 0) is the weight change amount (Δw) according to the time when the output spike (OS) is generated before the input spike (IS) is generated indicates

For example, if one input spike (IS) is generated within the previous 20 clocks from the time when the output spike (OS) is generated, and three input spikes (IS) are generated within 60 clocks from the previous 20 clock, the weight change amount ( Δw) is the number of input spikes (IS) generated within 20 clocks * 2 + the number of input spikes (IS) generated between 20 clocks and 60 clocks * 1, calculated as 1 * 2 + 1 * 3 = 5. Accordingly, the weight w is increased by 5 by the weight change amount Δw.

On the other hand, if two input spikes (IS) are generated within 20 clocks after the time when the output spike (OS) is generated, and then two input spikes (IS) are generated within 60 clocks from 20 clocks, the weight change amount ( Δw) is the number of input spikes (IS) generated within 20 clocks * -2 + the number of input spikes (IS) generated between 20 clocks and 60 clocks * -1 as 2 * -2 + 2 * -1 = - counted as 6. Accordingly, the weight w is reduced by 6 by the weight change amount Δw.

As a result, the weight learning unit 400 calculates the weight change amount (Δw) by giving a higher weight to the spike generated at the closer time among the spikes generated for a predetermined period from the standard spike according to the quantization STDP function, and , the weight is updated by adding the calculated weight change amount Δw to the existing weight w.

In FIG. 4, as an example of the quantization STDP function, it is shown that the weight change amount (Δw) is varied by 1 unit based on 20 (or -20) clocks and 60 (or -60) clocks. The number of clocks and the variable unit can be variously adjusted. In addition, the quantization period may be further subdivided.

FIG. 5 shows a detailed configuration of the weight learning unit of FIG. 1 .

Referring to FIG. 5 , the weight learning unit 400 may include an up variation acquiring unit 410 , a down variation acquiring unit 420 , a weight variation acquiring unit 430 , and a weight updating unit 440 .

_{The up change amount obtaining unit 410 obtains an up change amount Δw UP} for increasing the weight change amount Δw according to the input spike IS applied to the weighted synapse 200 based on the quantized STDP function. When the input spike IS is applied, the up variation obtaining unit 410 increases the _{up variation Δw UP to a predetermined first level according to the applied input spike IS.} Here, the first level is the maximum value of the quantization STDP function, and is set to 2 in FIG. 4 . Accordingly, when the input spike IS is applied, the up change amount obtaining unit 410 _{increases the up change amount Δw UP} by 2 . And after the input spike IS is applied, when a predetermined first time interval (20 clocks in FIG. 4) passes, the increased up change amount Δw _UP is reduced by 1, which is a predetermined change unit, for 2 hours After the interval (60 clocks in FIG. 4 ) passes, the increased up change amount Δw _UP is decreased by 1, which is a predetermined change unit.

That is, the up change amount obtaining unit 410 increases the up change amount _{Δw UP} by a first level corresponding to the maximum value of the quantized STDP function whenever the input spike IS is applied, and thereafter, the predetermined first and Each time the second time period passes, it is decreased by a predetermined change unit, and after the second time period, the input spike IS is applied and the up change amount Δw _UP increased by the first level is again the input spike IS It is converted to the same state as before application.

_{Here, since it is assumed that the up change amount obtaining unit 410 adjusts the up change amount Δw UP} based on the quantization STDP function of FIG. 4 , the first level is 2, and each time the first and second time intervals pass - It is decreased by 1, and the time interval has been described as being divided into two time intervals, a first time interval and a second time interval. However, as described above, the quantization STDP function may be changed, and in this case, the time interval may be divided into a plurality of time intervals.

That is, the up-change amount obtaining unit 410 may classify a plurality of time sections based on the quantized STDP function, and may reduce the _{up-change amount Δw UP in a predetermined change unit according to the divided time sections.}

The up change amount obtaining unit 410 repeatedly performs the above-described operation whenever the input spike IS is applied.

_{Meanwhile, the down change amount obtaining unit 420 obtains the down change amount Δw UP} for reducing the weight change amount Δw according to the output spike OS applied to the weighted synapse 200 based on the quantized STDP function. When the output spike OS is applied, the down variation obtaining unit 420 increases the _{down variation Δw DN to a predetermined first level.} And after the output spike (OS) is applied, when a predetermined first time interval passes, the increased down change amount (Δw _DN ) is reduced by 1, which is a predetermined change unit, and after a predetermined 2 time interval, the increased The down change amount (Δw _UP ) is decreased by 1, which is a predetermined change unit.

That is, the down change amount obtaining unit 420 increases the down change amount _{Δw DN} by a first level corresponding to the maximum value of the quantized STDP function whenever the output spike OS is applied, and thereafter, the predetermined first and Each time the second time period passes, it is decreased by a predetermined change unit, and after the second time period, the output spike OS is applied and the down change amount Δw _DN increased by the first level is again the input spike IS It is converted to the same state as before application.

Here, the down variation obtaining unit 420 also divides a plurality of time sections based on the quantized STDP function like the up variation obtaining unit 410, and according to the divided time sections, the down variation (Δw _DN ) in a predetermined change unit according to the divided time section. can be reduced

In addition, the down change amount obtaining unit 420 also repeatedly performs the above-described operation whenever the output spike OS is applied.

As a result, the up-variance acquisition unit 410 and the down-variance acquisition unit 420 perform basically the same operation except for the difference in whether the input spike IS or the output spike OS is applied. Accordingly, the up-variance acquisition unit 410 and the down-variance acquisition unit 420 may be implemented with hardware having the same configuration, thereby improving the convenience of implementation.

_{The weight change acquisition unit 430 receives the up change amount (Δw UP} ) and the down change amount (Δw _DN ) from the up change amount acquisition unit 410 and the down change amount acquisition unit 420 , and the down change amount in the up change amount (Δw _UP ) (Δw _DN ) is subtracted to obtain a weight change amount (Δw). Here, the weight change acquisition unit 430 obtains the weight change amount Δw by subtracting the down change amount Δw _{DN from the up change amount Δw UP} . The up change amount acquisition unit 410 and the down change amount acquisition unit 420 are This is because it is implemented by hardware of the same configuration and outputs a value with the same sign.

_{If the down change amount obtaining unit 420 is configured to output the down variance (Δw DN} ) having a negative value, unlike the up variance obtaining unit 410 , the weight variance obtaining unit 430 is the up variance (Δw _UP ) in the The weight change amount Δw may be obtained by adding the down change amount Δw _{DN .}

Then, the weight change amount obtaining unit 430 outputs the obtained weight change amount Δw to the weight update unit 440 . Here, the weight change amount obtaining unit 430 may obtain the weight change amount Δw at every clock in real time and transmit the obtained weight change amount Δw to the weight update unit 440 .

When a spike (input spike (IS) or output spike (OS)) occurs, the weight update unit 440 receives the stored current weight w(t), and changes the weight to the applied weight w(t). (Δw) is added to obtain an update weight w(t+1). Then, the weight is updated by replacing the weight w(t) in the weight synapse 200 and storing the obtained update weight w(t+1). That is, the SNN is trained.

6 shows a schematic structure of the change amount obtaining unit of FIG. 5 , and FIG. 7 shows an example of a detailed configuration of the section delay unit of FIG. 6 .

6 illustrates a schematic structure of the up change amount obtaining unit 410 as an example, and it is assumed that the input spike IS is applied thereto. However, as described above, since the up variation acquiring unit 410 and the down variation acquiring unit 420 may be implemented with the same hardware, the down variation acquiring unit 420 may also have the same structure. However, in the case of the down change amount acquisition unit 420, the output spike OS, not the input spike IS, is applied.

Referring to FIG. 6 , the up change amount obtaining unit 410 may include a first interval delay unit 411 , a second interval delay unit 412 , and a change amount calculation unit 413 .

First, the first period delay unit 411 delays the applied input spike IS for a first predetermined time period and outputs it. In addition, the second section delay unit 422 delays the input spike IS output from the first section delay unit 411 for a second time section and outputs the delay.

As shown in FIG. 7 , the first section delay unit 411 and the second section delay unit 412 may be implemented as a plurality of flip-flops sequentially connected, respectively. The first interval delay unit 411 is implemented with the number of flip-flops FF1 to FF20 corresponding to the first time interval, and the second interval delay unit 412 includes the number of flip-flops corresponding to the second time interval. FF21 to FF60). Here, for example, based on the quantization STDP function shown in FIG. 4 , the first section delay unit 411 includes 20 flip-flops FF1 to FF20 to delay the applied input spike IS by 20 clocks, and , the second section delay unit 412 is illustrated as including 40 flip-flops FF1 to FF40 to be delayed by 40 clocks.

When the input spike IS is applied, the change amount calculator 413 _{increases (+2) the stored up change amount Δw UP} by a predetermined first level (+2), and the first section delay unit 411 and the second section Each time the input spike IS is output from the delay unit 412, the stored up change amount Δw _UP is decreased (-1) by a predetermined change unit.

Here, as described above, it is assumed that the up-change amount obtaining unit 410 divides the time interval into two based on the quantized STDP function, including a first interval delay unit 411 and a second interval delay unit 412 . Although shown as doing, the up-change amount obtaining unit 410 may include a plurality of interval delay units, and in this case, the variance calculation unit 413 stores the input spikes IS from each of the plurality of interval delay units each time it is output. The up change amount Δw _UP may be decreased (-1) by a predetermined change unit.

As shown in FIG. 7 , when the first section delay unit 411 and the second section delay unit 412 are implemented as a plurality of flip-flops FF1 to FF60, respectively, the plurality of flip-flops FF1 to FF60 is an up-change amount obtaining unit. At 410, the history of the input spike IS applied for a predetermined period (here, 60 clocks) is always stored. That is, it is possible to acquire the input spike history with a very simple hardware configuration without using a separate memory unlike the existing one.

However, in the present embodiment, whenever an input spike is applied to the up change amount obtaining unit 410 or the delayed input spike IS is output from the first section delay unit 411 and the second section delay unit 412, the up change amount ( Since Δw _UP ) is updated in real time, it is not necessary to check the separately stored input spike history at the time of updating the weight w later. Therefore, the up change amount (Δw _UP ) can be updated in real time with a very simple hardware configuration.

And as described above, since the down change amount obtaining unit 420 can also be configured in the same way as the up change amount obtaining unit 410, even if the output spike history is not stored in a separate memory, the down change amount (Δw _DN ) is displayed in real time. can be updated.

As a result, in this embodiment, up variation of the input spikes history and output spike history using a plurality of flip-flops (Δw _UP) and down variation (Δw _DN) a can be obtained in real time, the acquired-up variation (Δw _UP ) and the weight change amount Δw from the down change amount Δw _DN can also be obtained in real time.

FIG. 8 shows another example of the detailed configuration of the interval delay unit of FIG. 6 .

Referring to the section delay unit of FIG. 8 , in FIG. 8 , the first section delay unit 411 and the second section delay unit 412 include four flip-flops FF1 to FF4 and eight flip-flops FF5 to FF12, respectively. do. That is, it has a configuration in which the number of flip-flops is reduced by 1/5 compared to the first section delay unit 411 and the second section delay unit 412 of FIG. 7 . In addition, each of the plurality of flip-flops FF1 to FF12 is applied with a clock CLK_X5 whose period is increased by 5 times compared to FIG.

7, when the plurality of flip-flops FF1 to FF60 are configured to transfer the input spike IS to the flip-flop of the next adjacent stage at every clock CLK, the history of all the input spikes IS is first Although the detailed spike history is obtained by being stored in the first section delay unit 411 and the second section delay unit 412, it is inefficient in terms of manufacturing cost, size, and power consumption as it includes a plurality of flip-flops.

Accordingly, in FIG. 8, by applying a time step scaling technique of increasing the clock period at a predetermined ratio, the flip-flops constituting each of the first section delay unit 411 and the second section delay unit 412 are shown. Allows the number to be reduced by the time step scaling factor. Here, as an example, the flip-flops of the first section delay unit 411 and the second section delay unit 412 are performed by increasing the cycle of the clock CLK by 5 times and in response to the scaling clock CLK_X5 whose cycle is increased by 5 times By operating it, the first section delay unit 411 and the second section delay unit 412 can each be configured with the number of flip-flops reduced to 1/5 compared to FIG. 7 .

However, according to the time step scaling technique, the flip-flop, which previously transmitted a spike in units of every clock, is transmitted every 5 clocks, and this may reduce the accuracy of the spike history. However, since the input and output spikes do not occur every clock, the error caused by this is not large. That is, the learning performance is not significantly degraded.

The above-described time step scaling technique may be equally applied to each section delay unit of the up variation obtaining unit 410 and the down variation obtaining unit 420, but in some cases, the up variation obtaining unit 410 and the down variation obtaining unit It may be applied to only one of (420). In addition, the time step scaling technique may be applied to the up variation obtaining unit 410 and the down variation obtaining unit 420 at different ratios.

9 shows the result of analyzing the learning accuracy according to the time step scaling ratio, and FIG. 10 shows the result of learning by applying the time step scaling.

As shown in FIG. 9 , it can be seen that the learning accuracy is almost the same as when the time step scaling ratio is not applied until the time step scaling ratio N is 4. And, as shown in FIG. 8, even when the time step scaling ratio (N) is 5, it was found that the learning accuracy was reduced by only about 2%.

10 shows the result of reconstructing the weights after performing learning by applying the time step scaling ratio (N) to 5, and as shown in FIG. 10 , the learning performance of the SNN even when the time step scaling ratio (N) is applied. It can be seen that is not significantly reduced.

Therefore, if the number of flip-flops included in the section delay unit is reduced by applying time step scaling, manufacturing cost, size, and power consumption can be reduced while maintaining learning performance.

The method according to the present invention may be implemented as a computer program stored in a medium for execution by a computer. Here, the computer-readable medium may be any available medium that can be accessed by a computer, and may include all computer storage media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data, and read dedicated memory), RAM (Random Access Memory), CD (Compact Disk)-ROM, DVD (Digital Video Disk)-ROM, magnetic tape, floppy disk, optical data storage, and the like.

Although the present invention has been described with reference to the embodiment shown in the drawings, which is merely exemplary, those skilled in the art will understand that various modifications and equivalent other embodiments are possible therefrom.

Accordingly, the true technical protection scope of the present invention should be defined by the technical spirit of the appended claims.

100: input neuron 200: weighted synapse
300: output neuron 400: weight learning unit
410: up variation acquisition unit 420: down variation acquisition unit
430: weight change amount acquisition unit 440: weight update unit
411: first section delay unit 412: second section delay unit
413: change amount calculation unit

Claims (9)

In order to train a spiking neural network that ignites an output spike when a cumulative spike signal obtained by cumulatively adding a weight corresponding to an applied input spike among a plurality of stored weights is equal to or greater than a predetermined firing threshold, the input spike and the The amount of change for adjusting the weight based on the time difference between output spikes is divided into a plurality of time sections and based on a preset quantization STDP function, the input spike is applied or after the input spike is applied, the quantization STDP an up change amount obtaining unit that updates the up change amount every time a plurality of time sections divided by a function are reached;
a down change amount obtaining unit that updates the down change amount whenever the input spike is applied or a plurality of time intervals divided by the quantization STDP function after the input spike is applied;
a change amount obtaining unit for obtaining a weight change amount using the up change amount and the down change amount; and
and a weight updater configured to update a weight by adding the weight change amount to the weight. The method of claim 1, wherein the up change amount obtaining unit
a plurality of section delay units connected in series to delay and output the input spike by a predetermined time section when the input spike is applied; and
When the input spike is applied, the pre-stored up change amount is increased by the maximum value set by the quantization STDP function, and whenever the delayed input spike is output from each of the plurality of section delay units, a predetermined change set by the quantization STDP function A spiking neural network learning apparatus comprising a change amount calculation unit for decreasing the up change amount by a unit. The method of claim 2, wherein each of the plurality of interval delay units comprises:
A plurality of time sections divided by the quantization STDP function are serially connected in a number corresponding to a corresponding time section, and the input spike is applied in response to a clock of a predetermined period, respectively, or a previously applied input spike is delivered. A spiking neural network training device including a plurality of flip-flops. 4. The method of claim 3, wherein each of the plurality of interval delay units comprises:
and a plurality of flip-flops receiving the input spike in response to the scaling clock increased according to a predetermined time step scaling ratio or transferring the previously applied input spike, wherein the cycle of the clock includes a plurality of flip-flops,
A spiking neural network learning apparatus including the plurality of flip-flops reduced by the time step scaling ratio in a corresponding time interval. According to claim 1, wherein the down change amount obtaining unit
a plurality of section delay units connected in series to delay and output the output spike by a predetermined time section when the output spike is applied; and
When the output spike is applied, the pre-stored down change amount is increased by the maximum value set by the quantization STDP function, and whenever a delayed output spike is output from each of the plurality of interval delay units, a predetermined value set by the quantization STDP function is output. A spiking neural network learning apparatus comprising a change amount calculation unit for decreasing the down change amount by a change unit. The method of claim 5, wherein each of the plurality of interval delay units is
It is serially connected in a number corresponding to a corresponding time interval among a plurality of time intervals divided by the quantization STDP function, and receives the output spike in response to a clock of a predetermined period, or delivers the previously applied output spike A spiking neural network training device including a plurality of flip-flops. 7. The method of claim 6, wherein each of the plurality of interval delay units comprises:
and a plurality of flip-flops receiving the output spike in response to the scaling clock increased according to a predetermined time step scaling ratio or transferring the previously applied output spike, wherein the cycle of the clock includes a plurality of flip-flops,
The spiking neural network learning apparatus is included in the number of the plurality of flip-flops reduced by the time step scaling ratio in a corresponding time interval. The method of claim 1, wherein the change amount obtaining unit
A spiking neural network learning apparatus for obtaining a weight change amount by subtracting the down change amount from the up change amount. The method of claim 1, wherein the weight update unit
When the input spike or the output spike is applied, the spiking neural network learning apparatus is configured to update the weight by adding the weight change amount to the corresponding weight.

Images