KR20210077159A - Spiking neural network and operation method thereof - Google Patents

Abstract

The present invention provides a spiking neural network and an operating method thereof, in which the spiking neural network includes: a weight synapse unit for storing a weight including a plurality of memory cells defined by a plurality of word lines and a plurality of bit lines; an input neuron unit for generating a plurality of input spikes in response to input data to activate a corresponding word line among the word lines; and an output neuron unit for receiving and accumulating weights that are bit-scaled into upper (n-k) bits by excluding lower k bits among weights of n bits in which bit values are stored in each of the memory cells disposed in an area where the activated word line intersects the bit lines, obtaining a membrane potential value by adding a membrane potential value in which a weight that is previously bit-scaled is accumulated, spiking an output spike when the obtained membrane potential value is greater than or equal to a spiking threshold that is bit-scaled to correspond to the bit-scaled weight, and counting a number of spiked output spikes to adjust the bit-scaled spiking threshold. Accordingly, a degree of integration and power efficiency are increased.

Description

SPIKING NEURAL NETWORK AND OPERATION METHOD THEREOF

The present invention relates to a spiking neural network and an operating method thereof, and to a spiking neural network capable of increasing integration and power efficiency and an operating method thereof.

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 the synapse between neurons transmits information in the form of spikes of electrical signals. This SNN processes information based on the timing at which the spike signal having a binary value is transmitted, that is, the time difference.

Since SNN is implemented in software, it does not require multiplication operation compared to deep neural networks represented by Convolution Neural Network (CNN) and Recurrent Neural Network (RNN), so it has a simple hardware structure. Since it can be implemented, it does not require a high-performance computer system and has the advantage of very low power consumption.

However, in order to mimic the human brain using SNN, at least hundreds of thousands to as many as trillions of neurons and synapses must be integrated. Therefore, integration and power efficiency are very important issues in the implementation of SNN.

SNNs are generally manufactured in the form of complementary metal-oxide semiconductor (CMOS)-based memory devices, but in recent years, they have been developed to utilize the characteristics of new types of memory devices such as RRAM (Resistive Random Access Memory) or PRAM (Phase-change Memory). Attempts are continuing. In the case of using such a new memory device, since the behavioral characteristics of neurons are imitated according to the material characteristics of each device, the degree of integration can be increased and power efficiency can be improved, but it is difficult to precisely control the characteristics of each device, and the characteristics between devices There is a limit in that reliability is low due to large deviation.

Accordingly, a method for increasing the degree of integration and power efficiency in a CMOS-based digital circuit is required.

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

SUMMARY OF THE INVENTION An object of the present invention is to provide a spiking neural network capable of increasing integration and power efficiency, and a method for operating the same.

Another object of the present invention is to provide a spiking neural network capable of maintaining accuracy of outputted spiking timing while increasing integration and power efficiency, and an operating method thereof.

A spiking neural network according to an embodiment of the present invention for achieving the above object includes a weight synapse unit for storing weights including a plurality of memory cells defined by a plurality of word lines and a plurality of bit lines; an input neuron unit generating a plurality of input spikes in response to input data to activate a corresponding word line among the plurality of word lines; and excluding the lower k bits among the weights of n bits in which bit values are stored in each of a plurality of memory cells disposed in a region where the activated word line and the plurality of bit lines intersect, and the weight scaled to upper nk bits is applied. Accumulated, the previous bit-scaled weight is added to the accumulated membrane potential value to obtain a membrane potential value, and if the obtained membrane potential value is equal to or greater than the bit-scaled firing threshold corresponding to the bit-scaled weight, an output spike is fired, and an output neuron unit that counts the number of output spikes and adjusts a bit-scaled firing threshold.

The output neuron unit calculates the weight applied through the bit line corresponding to the upper nk bit in each bit line group among a plurality of bit line groups in which a plurality of bit lines of the weighted synaptic unit are grouped in units of n bit lines. It can be obtained by weight.

The output neuron unit may bit scale the utterance threshold to be less than or equal to a bit scaling ratio of the bit scaled weight.

The output neuron unit includes a plurality of output neurons corresponding to each of the plurality of bit line groups, and each of the plurality of output neurons obtains a bit scaled weight from a bit line corresponding to an upper nk bit of the corresponding bit line group. a weight accumulator for accumulating and adding a previously obtained membrane potential to obtain a membrane potential; a fire determination unit igniting an output spike when the obtained film potential value is equal to or greater than the bit-scaled fire threshold; and a firing threshold adjusting unit configured to adjust the bit-scaled firing threshold by counting the number of output spikes fired during a predetermined period.

The output neuron further includes a leakage unit that obtains a leakage membrane potential value by weighting a predetermined leakage factor to the previously obtained membrane potential value, and the weight accumulator adds the leakage membrane potential value to the accumulated bit-scaled weight to obtain the membrane potential value. can be obtained

The output neuron may further include a reset selector configured to initialize the membrane potential to a predetermined level so as not to fire an output switch during a predetermined suppression period or refractory period.

The firing threshold adjusting unit counts the number of output spikes during a predetermined period, and if the counted number of output spikes is equal to or greater than the predetermined homeostatic up number, the firing threshold is increased by a predetermined threshold change amount, and the counted When the number of output spikes is less than or equal to the predetermined homeostatic down number, the firing threshold may be decreased by a predetermined threshold change amount.

The weighted synapse unit is connected between each bit line corresponding to the upper nk bit in each of the plurality of bit line groups and a corresponding output neuron, and transmits a bit signal applied through the corresponding bit line to the corresponding output neuron. may further include a sense amplifier of

The spiking neural network further includes a weight learning unit for learning the weights stored in the weight synapse unit, and the weight learning unit includes an input spike history including time information at which the plurality of input spikes are generated and a time at which the output spikes are fired. By storing an output spike history including information, the weight may be updated based on a time difference between an input spike generation time and an output spike firing time.

The weight learner may update the weight of n bits that are not bit scaled.

A method of operating a spiking neural network according to another embodiment of the present invention for achieving the above object generates a plurality of input spikes in response to input data, and a plurality of memories defined by a plurality of word lines and a plurality of bit lines activating a corresponding word line from among the plurality of word lines in a weight synapse that stores a weight including a cell; In each of a plurality of memory cells disposed in an area where the activated word line and the plurality of bit lines intersect, a weight scaled by bit scale to the upper nk bits is applied and accumulated by excluding the lower k bits among the weights of n bits in which bit values are stored. to do; obtaining a membrane potential value by adding a membrane potential value accumulated by a previous bit scaled weight to the accumulated bit scaled weight; igniting the output spike when the obtained membrane potential value is equal to or greater than the bit-scaled firing threshold corresponding to the bit-scaled weight; and counting the number of fired output spikes to adjust the bit-scaled firing threshold.

Therefore, in the spiking neural network and its operating method according to an embodiment of the present invention, the firing threshold is varied according to the number of spikes output from the spiking neural network based on the homeostasis of biological neurons, and the weight is based on the variable firing threshold. An object of the present invention is to provide a spiking neural network capable of significantly increasing integration and power efficiency while maintaining accuracy of spiking timing by varying the bit scale of , and a method for operating the same.

1 shows a schematic structure of a spiking neural network.
2 shows an example of the configuration of the output neuron of FIG. 1 .
3 and 4 are diagrams for explaining the operation of the weighted synapse and the output neuron of FIG. 1 .
5 illustrates an example of a speech threshold value that is changed by the speech threshold value adjusting unit.
6 shows a concept in which the SNN bit scales weights according to the present embodiment.
7 illustrates a change in a speech threshold according to weight bit scaling.
8 shows a comparison of learning results by weight bit scaling.
9 shows the learning accuracy according to the number of weight bit scaling bits.
10 shows a change in power consumption according to the number of weight bit scaling bits.
11 shows a method of operating a spiking neural network according to an embodiment of the present invention.

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. And, in order to clearly explain 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, FIG. 2 shows an example of the configuration of the output neuron of FIG. 1, and FIGS. 3 and 4 are diagrams for explaining the operation of the weighted synapse and the output neuron of FIG. 1 . And FIG. 5 shows an example of a speech threshold variable by the speech threshold value adjusting unit.

Referring to FIG. 1 , the SNN includes an input neuron unit 100 , a weighted synapse unit 200 , an output neuron unit 300 , and a weight learning unit 400 .

The input neuron unit 100 converts input data into a plurality of input spikes according to a predetermined method, and transmits the converted plurality of input spikes to the weighted synapse unit 200 . The human brain receives information sensed from various sensory organs in the form of electrical signals, and synapses between neurons transmit information between neurons as spike-type electrical signals. Accordingly, the input neuron unit 100 of the SNN configured to imitate the human brain also converts the input data into an input spike according to a predetermined method and transmits it to the weighted synapse unit 200 .

The input neuron unit 100 is composed of a plurality of input neurons, and may be converted into input spikes generated with a time difference according to values of input data corresponding to each. For example, a plurality of input neurons of the input neuron unit 100 may receive a pixel value 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 of the input neuron unit 100 to convert input data into input spikes have been studied in various ways. For example, input data can be converted into input spikes based on a Poisson distribution. That is, input spikes are generated by generating or not generating multiple spikes based on a probability based on a Poisson distribution for input data. At this time, as described above, the input neuron of the input neuron unit 100 generates a plurality of input spikes several times with a time difference based on the probability according to the input data, rather than converting the input data into one input spike. can be printed out.

That is, as shown in FIG. 3 , a plurality of input spikes may be generated several times at different times for each input data.

Here, the input neuron unit 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.

The weighted synapse unit 200 transmits weights to the input spike applied from the input neuron unit 100 to the output neuron unit 300 . A plurality of weights are stored in the weight synapse unit 200 , and the stored plurality of weights are updated by the weight learning unit 400 .

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

The weighted synapse unit 200 includes a plurality of word lines WL that are activated when an input spike is applied from a corresponding input neuron of the input neuron unit 100 and a plurality of bit lines BL crossing the plurality of word lines WL. ) and a plurality of memory cells MC defined by a plurality of word lines WL and a plurality of bit lines BL. In addition, each of the plurality of memory cells MC stores a corresponding bit value in the weight.

Here, the word lines WL may be arranged in a number corresponding to the number of input neurons, and the bit lines BL are arranged in a number unit bit line group BLG corresponding to the number of bits of the weight, and the bit line group (BLG) may be arranged in a number corresponding to the number of output neurons of the output neuron unit 300 to be described later. That is, the number of bit line groups BLG corresponding to the number of output spikes that the output neuron unit 300 can output in parallel is arranged. For example, if the weight is defined as an n-bit binary value and the number of output neurons of the output neuron unit 300 is m, n bit lines BL constitute a bit line group BLG, and m bits Since the line group BLG needs to be arranged, the weighted synaptic unit 200 may include n * m bit lines BL. In addition, as described above, the corresponding bit value of the weight is stored in the memory cell at the position where each word line WL and the bit line group BLG intersect. If the input neuron unit 100 includes N input neurons and outputs input spikes (x ₁ to x _N ) in response to input data, respectively, the weighted synaptic unit 200 includes N * n * m memory cells. (MC) may be included.

In the weighted synapse unit 200, when an input spike occurs in a specific input neuron, a corresponding word line WL is activated, and a memory cell at a position where the activated word line WL and a plurality of bit lines BL intersect The bit value stored in the MC is output to the output neuron unit 300 . That is, it may be configured to output a weight corresponding to the applied input spike to the output neuron unit 300 .

The weight synapse unit 200 repeatedly outputs a corresponding weight to the output neuron unit 300 in response to each of a plurality of input spikes generated and applied with a time difference from the input neuron unit 100 .

The output neuron unit 300 fires a plurality of output spikes based on the cumulative sum of the weights output from the weighted synapse unit 200 . The output neuron unit 300 may include a plurality of output neurons, and each of the plurality of output neurons 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 the membrane potential obtained by the accumulated weights is at a predetermined reference level. If it is abnormal, the output spike is fired, and the membrane potential value is leaked as time goes by and is gradually weakened.

Accordingly, each of the plurality of output neurons of the output neuron unit 300 configured based on the LIF neuron model also accumulates the weights applied from the weighted synaptic unit 200 to obtain a membrane potential value, and the obtained membrane potential value is a predetermined firing threshold. Above the value, it is configured to fire an output spike. And the output neuron may be configured to gradually decrease the membrane potential value over time.

2 is a diagram showing in detail the configuration of one of a plurality of output neurons of the output neuron unit 300. The output neurons include a weight accumulator 310, a leak unit 320, a reset selector 330, and a utterance discrimination. It may include a unit 340 , a spike counter 350 , an up/down determining unit 360 , and an ignition threshold generating unit 370 .

As shown in FIG. 3 , the weight accumulator 310 responds to a word line WL activated by each of a plurality of input spikes generated from a plurality of input neurons of the input neuron unit 100, and a weighted synaptic unit ( 200), all weights (w _i ) and the previously obtained membrane potential value (V(t-1)) are added. As shown in FIG. 3 , the weight synapse unit 200 _{applies a plurality of input spikes (x 1} to x _N ) to the weights (w ₁ to) corresponding to each of the input spikes (x ₁ to x _{N ).} obtaining w _N), and (and all of the x _i w _i) the addition, membrane potential values (V (t-1 obtained prior to the weighted summation of the input spikes (x _i w _i)) is the weighted input spike weighted weight) add up

As described above, the input spike (x ₁ to x _N ) in the form of a binary value activates a word line (WL) corresponding to it, and the weight synapse unit 200 calculates a weight corresponding to the activated word line (WL). Since it passes to the output neuron, the sum of the _{weighted input spikes (x i} w _{i ) (∑x i} w _i ) can actually be viewed as the sum of the weights.

On the other hand, the weight accumulator 310 does not add the previously obtained membrane potential value (V(t-1)) to _{the sum of input spikes (∑x i} w _{i ) as it is, but the leaker 320 uses the previously acquired membrane potential value (} The membrane potential value V(t) is obtained by adding the leakage membrane potential value λV(t-1)) weighted by the leakage factor λ (where 0 < λ < 1) to V(t-1). can

The leakage factor 320 weighting the leakage factor λ to the previous membrane potential value V(t-1) reflects the characteristic that the nerve cells of the brain gradually lose the information stored over time, and the weight The addition of the previous leakage membrane potential (λV(t-1)) to the sum of input spikes (∑x _i w _{i ) by the accumulator 310 reflects the characteristic of memorizing information through repeated stimulation.} That is, the leak unit 320 and the weight accumulator 310 are responsible for leakage and integration functions in the LIF neuron model, respectively.

In addition, the reset selector 330 selects one of the film potential value V(t) obtained from the weight accumulator 310 or the reset voltage V _rst having a predetermined level and transmits it to the ignition determination unit 340 . do. Here, the reset voltage V _rst is a voltage for initializing the membrane potential value V(t), and may be set to 0V, for example. For example, the reset selector 330 may include a refractory signal that prevents an output neuron that has fired an output spike from firing for a predetermined refract period or an output neuron that has not fired an output spike during a predetermined inhibition period. _{A reset voltage V rst} for initializing the membrane potential value V(t) in response to a suppression signal that prevents ignition during a period may be selected.

The ignition determination unit 340 includes a film potential V(t) or a reset voltage V _{rst applied from the reset selection unit 330 and a firing threshold value V th} applied from the ignition threshold generator 370 . are compared, and an output spike is fired according to the comparison result.

The ignition determining unit 340 does not output an output spike when the reset voltage Vrst is applied from the reset selection unit 330 . However, the membrane potential value (V(t)) output from the weight accumulator 310 is _{the sum of the weighted input spikes (x i} w _i ) (∑x _i w _i ) and the previous leakage membrane potential value (λV(t-1)) As a signal obtained by adding , it may be higher or lower than the _{firing threshold value (V th ).}

As shown in FIG. 4 , when the film potential value V(t) _{is lower than the ignition threshold value V th} , the fire determination unit 340 does not fire an output spike, and the film potential value V(t) )) is higher than the firing threshold (V _th ), the output spike is fired.

The spike counter 350, the up-down determining unit 360 and the firing threshold generating unit 370 constitute a firing threshold adjusting unit that adjusts the voltage level _{of the firing threshold (V th ) according to the number of output input spikes,} , the firing threshold control unit is a component added to implement biologically plausible, such as homeostasis of living neurons. Homeostasis is applied to ensure stability and increase accuracy by preventing problems such as overfitting in unsupervised learning by maintaining the excitability of output neurons at a predetermined level during SNN learning.

The spike counter 350 counts the number of output spikes output from the utterance discrimination unit 340 during a predetermined homeostasis epoch and outputs a counting value cnt. And the spike counter 350 initializes the counting value cnt in response to the reset signal rst every homeostatic period.

When the spike counter 350 outputs the counting value cnt, the up/down determining unit 360 determines that the counting value cnt _{is greater than the homeostasis up threshold H th_up} or smaller than the homeostasis down threshold H _{th_dn} ). not determined, and the determination result output homeostasis up signal (H _up) to increase the ignition threshold value (V _th) in accordance with, or outputs a homeostasis-down signal (H _dn) to decrease the ignition threshold value (V _th) .

The up-down determining unit 360 may include a homeostasis up determining unit 361 and a homeostatic down determining unit 362 as shown in FIG. 2 . The homeostasis up determining unit 361 compares the counting value cnt with a predetermined homeostasis up threshold H _{th_up} , and when the counting value cnt _{is greater than the homeostasis up threshold H th_up} , the ignition threshold V _th ) to output the homeostasis up signal (H _{up ) to increase.} And the homeostasis down determining unit 362 compares the counting value cnt with a predetermined homeostasis down threshold (H _{th_dn} ), and if the counting value (cnt) is _{less than the homeostasis down threshold (H th_dn} ), the firing threshold ( V _th ) to output a homeostatic down signal (H _dn ) to decrease. Here, the homeostasis up threshold (H _{th_up} ) and homeostasis down threshold (H _{th_dn} ) may be a predetermined level higher value and a predetermined level lower value based on the preset homeostasis threshold value (H _{T ).}

The firing threshold generator 370 determines the threshold change amount ΔV _{th in response to the homeostatic up signal H up} or the homeostatic down signal H _dn applied from the up/down determining unit 360 , and the determined threshold By adding the value change ΔV _th to the previous ignition threshold V _th , the ignition threshold V _th = V _th + ΔV _th is varied and output to the utterance determination unit 340 .

As an example the ignition threshold value generating unit 370 homeostasis up signal period of the two in response to (H _up) may increase the ignition threshold value (V _th) by the threshold change amount (ΔV _th) of the specified size, homeostasis-down In response to the signal H _dn , the firing threshold value V _th may be decreased by a threshold change amount ΔV _{th of a predetermined negative magnitude.} That is, the firing threshold adjusting unit increases or decreases the _{firing threshold V th} according to the number of output spikes fired by the firing determining unit 340 during the homeostasis period as shown in FIG. 5A , The number of output spikes fired during the same homeostatic period should be as constant as possible.

In this case, the threshold change amount ΔV _{th according to the homeostasis up signal H up} and the homeostasis down signal H _dn may have the same magnitude, but may be specified differently.

As described above, the firing threshold adjusting unit calculates the homeostasis up threshold (H _{th_up} ) and homeostasis down threshold, which _{is set based on the homeostasis threshold (H T} ), a counting value (cnt), which is the number of output spikes counted during the homeostasis period. Compared to (H _{th_dn} ), output a homeostatic up signal (H _up ) or a homeostatic down signal (H _dn ), and output a homeostatic up signal (H _up ) or a homeostatic down signal (H _dn ) to a firing threshold (V) _th ) can be adjusted in various ways. In addition, since the firing threshold (V _th ) can be adjusted in various ways, it is possible to suppress the error caused by bit scaling of weights as much as possible, and also to prevent overfitting, so that the SNN can accurately measure various input data. make learning possible.

Here, the homeostasis period may be set based on the number of times the same input data is repeatedly input during learning, and in FIG. 5 ( a ), as an example, a period in which input data is repeatedly input 500 times and two input data is input is homeostasis. A case in which the period was set is shown. The homeostasis period may be set in various ways, for example, it may be set to a 2,500 cycle period that is input repeatedly 500 times for each of 50 pieces of input data.

5 (b) shows the change in the _{speech threshold value (V th} ) during the learning period in which the SNN is learned. Here, it is assumed that the SNN is implemented based on a digital CMOS circuit, and thus the firing threshold (V _th ) is also expressed as a digital value rather than a voltage level. As shown in FIG. 5B , the fire threshold V _th may have a very large wide variety of values by the fire threshold adjuster. Here, the firing threshold adjusting unit adjusts the firing threshold value (V _th ) whenever an output spike is fired or not fired or in proportion to the number of output spikes fired during the homeostasis period without increasing or decreasing _{the firing threshold value (V th )} , _{increasing or decreasing the ignition threshold V th} by a predetermined threshold change amount ΔV _th is to facilitate the implementation of the ignition threshold adjustment unit.

In the above, the spike counter 350 outputs a counting value cnt corresponding to the number of counted output spikes, and the homeostasis up determining unit 361 and the homeostasis down determining unit 362 of the up/down determining unit 360 are By comparing the counting value (cnt) with the homeostasis up threshold (H _{th_dn} ) and the homeostatic down threshold (H _{th_dn} ), respectively, the homeostasis up signal (H _up ) or the homeostatic down signal (H _dn ) was described as outputting, but the spike The counter 350 may be simply configured to output _{a homeostatic up signal (H up} ) or a homeostatic down signal (H _dn ) by comparing the counted number of output spikes with a predetermined homeostatic up number and homeostatic down number.

The weight learning unit 400 learns a plurality of weights of the weighted synaptic unit 200 according to a spike-timing-dependent plasticity (STDP) technique. The weight learning unit 400 for learning the weights according to the STDP technique increases the weights based on the time information at which the input spikes are generated in the input neuron unit 100 for a previously predetermined period based on the time when the output spikes are fired. On the other hand, the weight may be reduced based on information about the time when the output spike is fired for a previously predetermined period based on the time when the input spike is generated in the input neuron unit 100 from the input neuron unit 100 .

First, when the output spike is output from the output neuron unit 300, the weight learning unit 400 corresponds to a predetermined learning window size (eg, 60) before the output spike is fired, and the weighted synapse unit ( 200) and analyzes the input spike history indicating information on the applied input spike. Then, by comparing the generation time of the previous input spike stored in the input spike history with respect to the time when the output spike is fired, the amount of change in the weight is calculated according to the time difference, and the weight is increased.

On the other hand, when an input spike is generated in the input neuron unit 100 and applied to the weight synaptic unit 200 , the weight learning unit 400 corresponds to a predetermined learning window size before the input spike occurs during a predetermined period in the past. Analyze the output spike history indicating the information of the fired output spike. Then, the weight is reduced by calculating the change amount of the weight according to the time difference between the time when the output spike is fired based on the time when the input spike is generated.

If the time difference from the time when the output spike is fired to the time when the previous input spike is generated is short, the weight learner 400 increases the weight 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 generated to the time when the previous output spike is fired is short, the weight is reduced so that the correlation between the input spike and the output spike is low. This is because the output spike should be fired in response to the previously inputted input spike, whereas the incoming input spike should be generated as independent of the previously fired output spike as possible.

That is, the weight learning unit 400 increases or decreases the weight based on the time difference between the applied input spike and the output spike.

A method for the weight learning unit 400 to learn the weight synapse unit 200 based on the STDP technique is a well-known technique and will not be described in detail here.

On the other hand, as described above, when the membrane potential value V(t _{) is higher than the firing threshold value V th} , the firing determination unit 340 of the output neuron unit 300 ignites an output spike, and the firing threshold value of the output neuron unit 300 . (V _th ) is variously varied. Therefore, in reality, the value of the upper bit of the membrane potential value V(t) dominates, while the value of the lower bit may have an error negligible due to the _{variable firing threshold value V th .}

In general, in SNN, the more neurons are implemented, the higher the accuracy, but since the power consumption of neurons is large, the simplest way to reduce power consumption while increasing the density of neurons is to reduce the number of bits in weights. When the number of bits of the weight is reduced, as each of the plurality of output neurons of the output neuron unit 300 performs an operation on fewer bits, the circuit structure is simplified and the degree of integration of the SNN can be improved. Also, when the number of bits of the weight is reduced, the weight synapse unit 200 is connected to each of the plurality of bit lines, and the sense amplifier transmits the bit value of the weight transmitted through the corresponding bit line to the output neuron unit 300 . By reducing the number (not shown), the degree of integration of the SNN can be further increased.

However, when using a weight with low precision by reducing the number of bits in the weight, there is a problem in that the stability of learning is lowered.

Therefore, in the present embodiment, by bit scaling the weights in the process of calculating the sum (∑x _i w _{i ) of the weighted input spikes (x i} w _i ), the output spikes are fired normally, but the integration of the SNN is increased. to reduce power consumption.

Although it has been described above that the output neuron unit 300 includes a plurality of output neurons corresponding to the plurality of bit line groups BLG, in some cases, the output neuron unit 300 includes at least one output neuron, It may be configured to obtain only weights.

6 illustrates a concept in which the SNN according to the present embodiment bit-scales a weight, and FIG. 7 illustrates a change in an utterance threshold according to the weight-bit scaling.

In FIG. 6, (a) shows a configuration in which weights are not bit scaled, and (b) is a configuration in which weights are not bit scaled. Comparing (a) and (b) of FIG. 6 , in (a), the sense amplifier SA in the weighted synaptic unit 200 corresponds to all bit lines for each of the plurality of bit line groups BLG, 8 each. It is provided and configured to apply an 8-bit weight to the weight accumulator 310 . On the other hand, in the case of bit scaling the weight according to the present embodiment, as shown in (b), a predetermined number (here, four) corresponding to the lower bit of the weight among the bit lines of the bit line group BLG. The sense amplifier SA corresponding to the bit line is not provided. Accordingly, the weight accumulator 310 receives only the upper bit of the weight. That is, in (a), an 8-bit weight is applied to the weight accumulator 310 , whereas in (b), a 4-bit weight is applied to the weight accumulator 310 .

Accordingly, the values of the membrane potential values V(t) calculated by the weight accumulator 310 of (a) and (b) are output differently from each other. However, as described above, the membrane potential (V(t)) is a value for firing the output spike, and firing of the output spike is generated by comparing the _{membrane potential (V(t)) and the firing threshold (V th ).} do. In addition, the fire threshold V _th is varied by the fire threshold adjustment unit.

_{Therefore, if the change in the firing threshold (V th} ) is greater than the error caused by the reduced number of bits by bit-scaling the weight, the error generated by bit-scaling the weight can be ignored, and the output spike can be fired at the same time. . As described above, the weight is bit scaled by 4 bits from 8 bits to 4 bits, and correspondingly, the firing threshold value (V _th ) having a 16-bit level shown in (a) of FIG. 7 is 12 as shown in (b). When bit-scaled by 4 bits, an error generated by bit-scaling the weight can be ignored.

The bit scaling ratio of the weight and the bit scaling ratio of the speech threshold (V _th ) are not necessarily the same, but when the bit scaling ratio of the speech threshold (V _th ) is increased to exceed the bit scaling ratio of the weight, output The change may occur at the point in time when the spike is fired. That is, an error may occur in the output spike. Therefore, it is preferable that the bit scaling ratio of the speech threshold V _th is set within the bit scaling ratio of the weight.

When bit scaling is performed in this way, each of the plurality of output neurons of the output neuron unit 300 does not perform an accumulative operation on an 8-bit weight, but performs an accumulative operation on a 4-bit weight to obtain a membrane potential value. and the output spike may be ignited according to the obtained membrane potential value. Accordingly, since each of the plurality of output neurons of the output neuron unit 300 is bit-scaled, the circuit structure can be simplified according to the number of bits of the reduced weight, thereby improving the degree of integration of the SNN. For example, the weight accumulator 310 may also be replaced with a 12-bit accumulator instead of a 16-bit adder for accumulating 8-bit weights. And the remaining components of the output neuron shown in FIG. 2 _{can also be implemented to process a smaller number of bits according to the bit-scaled firing threshold (V th} ) and the bit-scaled weight, thereby improving the degree of integration of the SNN. .

Also, as shown in (b), the number of sense amplifiers SA of the weighted synaptic unit 200 can be reduced. If the 8-bit weight is bit-scaled to 4 bits, the number of sense amplifiers can be reduced by half. Therefore, the degree of integration of the SNN can be further improved.

Meanwhile, in terms of power, the weighted synaptic unit 200 requires precharged power E _PRE because the bit lines of the bit line group BLG must be precharged. And when one input spike is applied, the word line (WL) must be activated, and since the weight corresponding to the activated word line (WL) is applied to the sense amplifier (SA) connected to each bit line, the sense amplifier (SA) is The wordline activation power E _WL is required to amplify the corresponding bit signal. _{In addition, power E ADD} for accumulating the weight applied by the weight accumulator 310 is required.

In the case of (a), n bit lines of the bit line group (BLG) must be precharged, n sense amplifiers must amplify the bit signal, and the adder must accumulate the weights of n bits into 16 bits. The power consumption for the output neuron is _{calculated as (E PRE} + E _WL ) * n * M + (E _{ADD_16} * M). where E _{ADD_16} represents the power consumption of the 16-bit accumulator.

On the other hand, in the case of (b), n/2 bit lines of the bit line group BLG need to be precharged, n/2 sense amplifiers need to amplify the bit signal, and the adder sets the weight of n/2 bits to 12. Since we need to accumulate in bits, the power consumption for M output neurons is _{calculated as (E PRE} + E _WL ) * n/2 * M + (E _{ADD_12} * M). where E _{ADD_12} represents the power consumption of the 12-bit accumulator. That is, it can be seen that the power consumption is greatly reduced to about 50% level.

However, even when bit scaling is performed as shown in FIG. 6B , the number of bit lines in the bit line group BLG should be maintained as it is. In this case, although the weight learning unit 400 updates the weight according to the time difference between the input spike and the output spike, the change in the updated weight appears as a very small value. However, if the weights updated by the weight learning unit 400 are bit-scaled, the stability of learning may be lowered and normal learning may not be performed due to an increase in the change range of the weights. Accordingly, by maintaining the number of bit lines in the bit line group BLG as it is, the weights stored in the weight synaptic unit 200 are not bit-scaled, so that the weight learning unit 400 performs the weight learning unit 400 despite the weight bit scaling in the output neurons. By allowing the weights to be updated with the same precision as before, the stability of learning can be maintained.

8 shows a comparison of learning results by weight bit scaling.

In FIG. 8, (a) shows the results obtained by reconstructing the weights by performing learning without bit-scaling the weights, and (b) shows the results obtained by reconstructing the weights by performing learning by bit-scaling the weights.

Comparing (a) and (b), it can be seen that learning is performed normally without being affected by whether or not the weight bit is scaled.

FIG. 9 shows the learning accuracy according to the number of weight bit scaling bits, and FIG. 10 shows the change in power consumption according to the number of weight bit scaling bits.

Referring to FIG. 9 , since the initial value of the weight is set randomly and the encoded input spike is generated based on the probability, there is some deviation in accuracy depending on the input data on which learning is performed, but in bit scaling up to 4 bits It can be seen that the accuracy does not decrease. That is, even when bit scaling is performed on weights, an error in timing at which an output spike is fired appears almost the same as when bit scaling is not performed.

However, as shown in FIG. 10, when bit scaling is performed on weights, it can be seen that the energy consumption consumed when the SNN performs addition operation on weights is greatly reduced in proportion to the increase in the number of bits to be bit scaled. have. In particular, when comparing the case in which bit scaling is not performed and the case in which bit scaling of 4 bits is performed in FIG. 10 , it can be seen that the energy consumption is reduced to 1/2 level.

11 shows a method of operating a spiking neural network according to an embodiment of the present invention.

Referring to FIG. 11 , the operating method of the spiking neural network according to the present embodiment will be described. First, a plurality of input spikes corresponding to input data are generated ( S10 ). The generated input spike activates a corresponding word line among a plurality of word lines of a weighted synapse in which a plurality of weights are stored.

Then, a weight stored in a memory cell defined by a plurality of bit lines intersecting the activated word line is applied and obtained through the bit line. At this time, the weight is stored in the form of n-bit binary data, but the weight is obtained by applying a bit-scaled weight to higher (n-k) bits except for the predetermined lower k bits (S20).

Then, a plurality of bit scaled weights corresponding to each of the plurality of input spikes applied are accumulated (S30). In addition, the membrane potential value (V(t)) by adding the previously obtained membrane potential value (V(t-1)) to the leakage membrane potential value (λV(t-1)) obtained by weighting the predetermined leakage factor (λ) to the accumulated bit scaling weight. )) is obtained (S40).

When the membrane potential V(t) is obtained, the obtained membrane potential V(t) is _{compared with the bit-scaled firing threshold V th} ( S50 ). If the membrane potential value V(t) is _{less than the firing threshold V th} , the bit scaling weight according to the input spike is again acquired and accumulated to obtain the membrane potential value V(t+1).

However, when the membrane potential value V(t) is _{equal to or greater than the ignition threshold value V th} , the output spike is fired and the membrane potential value V(t) is initialized ( S60 ).

When the output spike is fired, the number of output spikes fired during the predetermined homeostatic period is counted, and the bit-scaled firing threshold V _th is adjusted according to the counted number of output spikes ( S70 ). In this case, the bit-scaled firing threshold V _th may be increased or decreased regardless of the number of counted output spikes, and may be set to increase or decrease in proportion to the counted number of output spikes.

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 unit 200: weighted synaptic unit
300: output neuron unit 400: weight learning unit
SA: sense amplifier 310: weight accumulator
320: leakage 330: reset selection unit
340: ignition discrimination unit 350: spike counter
360: up-down determining unit 370: ignition threshold generating unit

Claims (18)

a weight synaptic unit for storing weights including a plurality of memory cells defined by a plurality of word lines and a plurality of bit lines;
an input neuron unit generating a plurality of input spikes in response to input data to activate a corresponding word line among the plurality of word lines; and
In each of a plurality of memory cells disposed in a region where the activated word line and the plurality of bit lines intersect, a weight scaled by bits to the upper nk bits is applied and accumulated by excluding the lower k bits among the weights of n bits in which bit values are stored. and obtaining a membrane potential value by adding the membrane potential value accumulated by the previous bit-scaled weight, and if the obtained membrane potential value is equal to or greater than the bit-scaled firing threshold corresponding to the bit-scaled weight, an output spike is fired, and the fired output A spiking neural network comprising an output neuron unit that counts the number of spikes and adjusts a bit-scaled firing threshold. According to claim 1, wherein the output neuron unit
Obtaining, as the bit scaled weight, a weight applied through a bit line corresponding to the upper nk bit in each bit line group among a plurality of bit line groups in which a plurality of bit lines of the weight synapse are grouped in units of n bit lines Spiking Neural Network. The method of claim 2, wherein the output neuron unit
A spiking neural network that bit scales the firing threshold to less than or equal to a bit scaling ratio of the bit scaled weight. The method of claim 2, wherein the output neuron unit
a plurality of output neurons corresponding to each of the plurality of bit line groups;
Each of the plurality of output neurons is
a weight accumulator for acquiring and accumulating a bit-scaled weight from a bit line corresponding to an upper nk bit of a corresponding bit line group, and obtaining a membrane potential value by adding the previously acquired membrane potential;
a fire determination unit igniting an output spike when the obtained film potential value is equal to or greater than the bit-scaled fire threshold; and
A spiking neural network comprising: a firing threshold adjusting unit configured to count the number of output spikes fired during a predetermined period to adjust a bit-scaled firing threshold. 5. The method of claim 4, wherein the output neuron is
Further comprising a leakage unit for obtaining a leakage membrane potential value by weighting a predetermined leakage factor to the previously obtained membrane potential value,
The weight accumulator adds the leaked membrane potential value to the accumulated bit-scaled weight to obtain the membrane potential value. 5. The method of claim 4, wherein the output neuron is
The spiking neural network further comprising a reset selector configured to initialize the membrane potential to a predetermined level so as not to ignite the output switch during a predetermined suppression period or fire resistance period. 5. The method of claim 4, wherein the ignition threshold adjustment unit
Count the number of output spikes for a predetermined period, and if the counted number of output spikes is equal to or greater than a predetermined homeostatic up number, the firing threshold is increased by a predetermined threshold change amount, and the number of counted output spikes is A spiking neural network that decreases the firing threshold by a predetermined threshold change amount if the number of homeostatic downs is less than or equal to a specified number. According to claim 4, wherein the weight synaptic unit
A plurality of sense amplifiers connected between each bit line corresponding to the upper nk bit in each of the plurality of bit line groups and a corresponding output neuron to transmit a bit signal applied through the corresponding bit line to the corresponding output neuron; A spiking neural network that includes more. The method of claim 2, wherein the spiking neural network
Further comprising a weight learning unit for learning the weight stored in the weight synapse unit,
The weight learning unit stores an input spike history including time information at which the plurality of input spikes are generated and an output spike history including time information at which the output spikes are fired, so that the time when the input spikes are generated and the time when the output spikes are fired A spiking neural network that updates the weights based on the time difference. 10. The method of claim 9, wherein the weight learning unit
A spiking neural network that updates the weights of n bits that are not bit scaled. A corresponding word among the plurality of word lines in a weight synapse that generates a plurality of input spikes in response to input data and stores a weight including a plurality of memory cells defined by a plurality of word lines and a plurality of bit lines activating the line;
In each of a plurality of memory cells disposed in an area where the activated word line and the plurality of bit lines intersect, a weight scaled by bit scale to the upper nk bits is applied and accumulated by excluding the lower k bits among the weights of n bits in which bit values are stored. to do;
obtaining a membrane potential value by adding a membrane potential value accumulated by a previous bit scaled weight to the accumulated bit scaled weight;
igniting the output spike when the obtained membrane potential value is equal to or greater than the bit-scaled firing threshold corresponding to the bit-scaled weight; and
A method of operating a spiking neural network, comprising the step of counting the number of fired output spikes and adjusting a bit-scaled firing threshold. 12. The method of claim 11, wherein the accumulating comprises:
A weight applied through a bit line corresponding to the upper nk bit in a corresponding bit line group among a plurality of bit line groups in which a plurality of bit lines of the weight synapse are grouped in units of n bit lines is obtained as the bit scaled weight How to operate a spiking neural network. 13. The method of claim 12, wherein adjusting the firing threshold comprises:
A method of operating a spiking neural network by bit-scaling the speech threshold to be less than or equal to a bit-scaling ratio of the bit-scaled weight and adjusting the bit-scaled speech threshold. 14. The method of claim 13, wherein the obtaining the membrane potential value comprises:
obtaining a leakage membrane potential value by weighting a predetermined leakage factor to the previously obtained membrane potential value; and
A method of operating a spiking neural network, comprising the step of adding the leaked membrane potential to an accumulated bit-scaled weight to obtain the membrane potential. The method of claim 13, wherein the operating method of the spiking neural network is
The method of operating a spiking neural network further comprising the step of initializing the membrane potential to a predetermined level so as not to ignite the output switch during a predetermined suppression period or fire resistance period. 14. The method of claim 13, wherein adjusting the firing threshold comprises:
counting the number of output spikes;
if the counted number of output spikes is equal to or greater than a predetermined homeostatic up number, increasing the firing threshold by a predetermined threshold change amount; and
When the counted number of output spikes is less than or equal to a predetermined homeostatic down number, the method of operating a spiking neural network comprising the step of reducing the firing threshold by a predetermined threshold change amount. The method of claim 12, wherein the operating method of the spiking neural network is
Further comprising the step of learning the weight stored in the weight synaptic unit,
The learning step
storing an input spike history including time information at which the plurality of input spikes are generated and an output spike history including time information at which the output spikes are fired;
calculating a time difference between an input spike generation time and an output spike firing time using the stored input spike history and output spike history; and
A method of operating a spiking neural network comprising the step of updating the weight based on the calculated time difference. 18. The method of claim 17, wherein updating the weight comprises:
A method of operating a spiking neural network that updates the weights of n bits that are not bit scaled.

Images