Spiking neural network system including current scaling integrator and signal processing method thereof

The current scaling integrator-based neuron circuit addresses high power consumption and low integration density in C-TIA circuits by scaling input current and using a super source follower, achieving reduced power and increased integration with maintained accuracy.

KR103046503B1Active Publication Date: 2026-11-11DONGGUK UNIVERSITY INDUSTRY ACADEMIC COOPERATION FOUNDATION +1
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
KR1020250181733
Authority / Receiving Office
KR · KR
Patent Type
Patents
Current Assignee / Owner
Filing Date
2025-11-26
Publication Date
2026-11-11
Estimated Expiration
2045-11-26

AI Technical Summary

Technical Problem

Conventional C-TIA-based neuron circuits face high power consumption and low integration density issues, with operational amplifiers consuming 60% of the total power and membrane capacitors occupying 45% of the area, limiting the efficiency and integration of spiking neural networks.

Method used

A current scaling integrator-based neuron circuit that scales input current by a factor of 'n' (>1) and uses a super source follower structure to maintain current linearity without operational amplifiers, reducing power consumption and capacitor area.

Benefits of technology

The new neuron circuit achieves a 60% reduction in power consumption and a 40% increase in integration density while maintaining high current linearity, ensuring accurate inference performance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 112025132753293-PAT00005_ABST
    Figure 112025132753293-PAT00005_ABST
Patent Text Reader

Abstract

The present invention relates to a spiking neural network system and is characterized by replacing a conventional C-TIA (Capacitance Transimpedance Amplifier), which consumes a large amount of power, with a current scaling integrator (CSI). The current scaling integrator outputs the input current by dividing it by a scaling factor 'n' (1 / n), thereby reducing the size of the membrane capacitor to 1 / n. Through this, it is possible to drastically reduce the power consumption of the neuron circuit and significantly improve integration density while maintaining high linearity without using an operational amplifier.
Need to check novelty before this filing date? Find Prior Art

Description

Technology Field

[0001] The present invention relates to a neuromorphic computing system, and more specifically, to a current scaling integrator-based neuron circuit capable of improving the integration density of neuron circuits and reducing power consumption when implementing a spiking neural network (SNN) in hardware, and a signal processing method using the same. Background Technology

[0003] With the recent advancement of artificial intelligence, neuromorphic computing technology, which mimics the human brain, is gaining attention to address the bottlenecks and high power consumption issues of the existing Von Neumann architecture. Among these, the Spiking Neural Network is a technology with significant potential for implementing energy-efficient systems by mimicking the activity of biological neurons and processing information based on discrete spike events.

[0004] When implementing such spiking neural networks in hardware, the neuron circuits are primarily based on the Leaky Integrate-and-Fire (LIF) model. The LIF neuron model is the current (I) corresponding to the spike input from the synapse, i.e., the synaptic weight. in ) block capacitor (C mem It accumulates (integrates) over time in ). Through this integration process, the neuron's membrane potential (V mem As ) rises, this membrane potential (V mem ) is a preset threshold voltage (V th If it exceeds ), the neuron generates an output spike (Fire) and the membrane potential (V mem ) reset voltage (V rst Initialize with ).

[0005] The initial method of implementing the above LIF neuron circuit involves a membrane capacitor (C) at the current mirror output (drain) of the synapse. mem It was to directly connect the ). However, this method, the membrane potential (Vmem There was a problem where the voltage of the node connected to the capacitor fluctuated along with the integral, causing the operation of the current mirror MOSFET to become unstable. This severely degraded the linearity of current transmission, resulting in a limitation where the synaptic weight (current) was inaccurately transmitted to the neuron.

[0006] To address this problem of reduced linearity, C-TIA (Capacitance Transimpedance Amplifier)-based neuron circuits utilizing Operational Amplifiers (Op-Amps) have conventionally been widely used. C-TIA circuits utilize the feedback characteristics of operational amplifiers to maintain the drain terminal voltage of the current mirror constant, acting like a virtual ground. Through this, the membrane potential (V mem Even if ) fluctuates, the current mirror operates stably, enabling high current linearity to be secured.

[0007] However, these conventional C-TIA-based neuron circuits have the following two clear limitations.

[0008] First, there is the issue of high power consumption. Operational amplifiers, which are core components of the C-TIA structure, consume very high static current. This accounts for about 60% of the total power consumption of the neuron circuit, and has become a critical factor in reducing the overall energy efficiency of large-scale spiking neural network chips in which numerous neurons are integrated.

[0009] Second is the problem of low integration density (large area). To ensure stable operation and linearity, the C-TIA structure uses relatively large film capacitors (C mem It requires ). This membrane capacitor occupied a large area, amounting to about 45% of the total area occupied by a single neuron. This became a serious obstacle to high integration, which increases the number of neurons per unit area of ​​the chip. The problem to be solved

[0011] One objective of the present invention is to provide a low-power neuron circuit and a signal processing method thereof that can solve the problem of high power consumption of existing C-TIA (capacitance transimpedance amplifier)-based neuron circuits when implementing spiking neural network hardware.

[0012] Another objective of the present invention is to provide a highly integrated neuron circuit and a signal processing method thereof that can dramatically improve the integration density of neuron circuits by solving the problem of the large area occupied by the C-TIA structure.

[0013] Another objective of the present invention is to provide a new neuron circuit structure and a signal processing method thereof that can guarantee the inference accuracy of an SNN system by maintaining the high current linearity, which was an advantage of C-TIA, while achieving the aforementioned low power and high integration. means of solving the problem

[0015] To solve the problem described above, the present invention provides the following means of solution.

[0016] A spiking neural network system according to one embodiment of the present invention comprises an input current (I) corresponding to a synapse weight. in A synapse providing ); and a neuron circuit receiving the input current; wherein the neuron circuit comprises the input current (I in ) receiving the scaled output current (I) in which the input current is divided by a scaling factor 'n' (n>1). out = I in A current scaling integrator that generates ( / n); connected to the current scaling integrator, the scaled output current (I out Integrating ) to obtain the membrane potential (V mem A film capacitor generating ); and the film potential (V mem ) critical voltage (V thIt is characterized by including a comparator that generates an output spike by comparing with ).

[0017] In one embodiment, the current scaling integrator may be characterized by including a first MOSFET and a second MOSFET, which are sized to have a ratio of 'n' to generate the scaled output current.

[0018] In one embodiment, the current scaling integrator may be characterized by including a super source follower structure for maintaining current linearity and driving the film capacitor.

[0019] In one embodiment, the super source follower may be characterized by including a first MOSFET acting as a source follower; and a second MOSFET forming a negative feedback loop with the first MOSFET, thereby lowering the output impedance of the current scaling integrator.

[0020] In one embodiment, the scaling factor 'n' is 2, the scaled output current is 1 / 2 of the input current, and the capacitance of the film capacitor is reduced to 1 / 2 compared to when the scaling factor 'n' is 1.

[0021] In one embodiment, the neuron circuit does not include an operational amplifier-based capacitance transimpedance amplifier, and the current scaling integrator may be characterized by stabilizing the voltage at the current mirror drain terminal of the synapse to ensure current linearity.

[0022] In one embodiment, the neuron circuit may be characterized by having an average power consumption reduced by 60% or more compared to a neuron circuit using a capacitance transimpedance amplifier.

[0023] In one embodiment, the neuron circuit may be characterized by having a neuron integration density improved by more than 40% compared to a neuron circuit using a capacitance transimpedance amplifier due to a reduction in the area of ​​the membrane capacitor.

[0024] A spiking neural network signal processing method according to another embodiment of the present invention comprises an input current (I) corresponding to a synapse weight from a synapse. in A step of receiving ); using a current scaling integrator, the input current (I in The scaled output current (I) obtained by dividing ) by the scaling factor 'n' (n>1). out = I in A step of generating ( / n); the scaled output current (I) on the film capacitor out Integrating the membrane potential (V mem A step of generating ); and the membrane potential (V mem ) critical voltage (V th It is characterized by including the step of generating an output spike by comparing with ).

[0025] In another embodiment, the generating step may be characterized by being performed while maintaining current linearity using a super source follower structure within the current scaling integrator.

[0026] In another embodiment, the scaling factor 'n' is 2, and the integrating step may be characterized by using a film capacitor having half the capacitance capacity required when the scaling factor 'n' is 1.

[0027] In another embodiment, the generating and integrating steps may be performed without using an operational amplifier-based capacitance transimpedance amplifier to reduce power consumption. Effects of the invention

[0029] A spiking neural network system and its signal processing method according to one embodiment of the present invention has the effect of drastically reducing the average power consumption of a neuron circuit by more than 60% by replacing a conventional capacitance transimpedance amplifier (C-TIA) circuit, which includes an operational amplifier with high power consumption, with a current scaling integrator that operates at low power.

[0030] In addition, the present invention has the effect of reducing the total area of ​​the neuron circuit and improving the neuron integration density by about 40% or more by scaling the input current by 1 / n times (e.g., 1 / 2 times) through the current scaling integrator, thereby reducing the capacitance (area) of the membrane capacitor required for neuron operation by 1 / n times (e.g., 1 / 2 times).

[0031] Furthermore, the present invention can secure high current linearity equivalent to that of the existing C-TIA by stably maintaining the drain terminal voltage of the synapse current mirror through a super source follower structure within the current scaling integrator without an operational amplifier, thereby having the effect of maintaining high inference accuracy (e.g., MNIST 95.41%).

[0032] However, the effects of the present invention are not limited to those described above, but include all effects naturally realized through the various configurations proposed in the present invention. Brief explanation of the drawing

[0034] FIG. 1 is an overall block diagram of a current scaling integrator-based spiking neural network system according to one embodiment of the present invention. Figure 2 is a detailed circuit diagram of a current scaling integrator, which is a key component of the present invention. Figure 3 is a detailed circuit diagram of a super source follower applied to the present invention. Figure 4 is a detailed overall circuit diagram of a neuron circuit utilizing a current scaling integrator according to the present invention. Figure 5 is a timing diagram to explain the operation of the neuron circuit of Figure 4. Figure 6 is a layout comparison diagram illustrating the area reduction effect of a film capacitor when current scaling (n=2) according to the present invention is applied (Example). Figure 7 is a graph comparing the current linearity of the current scaling integrator of the present invention (Example) and a conventional C-TIA. Figure 8 is a graph comparing the inference energy consumption of the neuron circuit of the present invention (Example) and the conventional C-TIA neuron circuit. FIG. 9 is a schematic diagram of a general analog IF neuron circuit according to the prior art. FIG. 10 is a schematic diagram of a neuron circuit based on a C-TIA (capacitance transimpedance amplifier) ​​according to the prior art. Specific details for implementing the invention

[0035] The embodiments of the present invention are illustrative for the purpose of explaining the technical concept of the present invention. The scope of rights according to the present invention is not limited to the embodiments presented below or the specific description thereof.

[0036] All technical and scientific terms used in this invention, unless otherwise defined, have the meaning generally understood by those skilled in the art to which this invention pertains. All terms used in this invention are selected for the purpose of further explaining this invention and are not selected to limit the scope of rights according to this invention.

[0037] Expressions such as "comprising," "having," "having," etc. used in the present invention should be understood as open-ended terms implying the possibility of including other embodiments, unless otherwise stated in the phrase or sentence containing such expressions.

[0038] Unless otherwise stated, singular expressions described in the present invention may include the meaning of the plural form, and this applies likewise to singular expressions described in the claims.

[0039] As used in this document, the term "module" may include a unit implemented in hardware, software, or firmware, and may be used interchangeably with terms such as logic, logic block, component, or circuit. A module may be a component formed as a whole, or a minimum unit of said component or a part thereof that performs one or more functions.

[0040] In this document, "modules" or "nodes" perform operations such as moving, storing, and converting data using computing devices and storage devices, such as processors (e.g., CPU, AP, etc.) and memory. For instance, "modules" or "nodes" can be implemented as devices such as servers, PCs, tablet PCs, and smartphones.

[0042] The present invention relates to a spiking neural network (SNN), which is one of the neuromorphic computing technologies that mimics the human brain. By mimicking the activity of biological neurons, spiking neural networks have great potential for implementing energy-efficient and low-power-consumption AI system architectures.

[0043] Hardware neurons in spiking neural networks are typically implemented based on the Leaky Integrate-and-Fire (LIF) model. In the LIF neuron model, when input spikes from synapses are transmitted as current according to weights, they act as a membrane capacitor (C mem Integrating ) over time gives the membrane potential (V mem It accumulates the ). The membrane potential (V) thus accumulated mem ) is a preset threshold voltage (V th If it exceeds ), the neuron generates an output spike (Fire) and the membrane potential (V mem ) specific reset voltage (Vrst It operates by initializing with ).

[0044] A conventional method for implementing such LIF neurons in a CMOS circuit was a general analog IF neuron circuit as shown in FIG. 9. FIG. 9 (prior art) shows the activation current (I) corresponding to the synaptic weight. exc ) and inhibition current (I inh A film capacitor (C) at the drain terminal of the current mirror that generates ) mem It is a structure in which ) is directly connected. However, this method uses a film capacitor (C mem Charge accumulates at the membrane potential (V). mem As the current changes, the drain terminal voltage of the current mirror fluctuates unstably. This causes the MOSFET of the current mirror to move out of the saturation region, resulting in a problem where the current corresponding to the synapse weight is not accurately transmitted, i.e., the current linearity is severely degraded.

[0045] To solve the aforementioned problem of degraded current linearity, a C-TIA (capacitance transimpedance amplifier)-based neuron circuit as shown in FIG. 10 was proposed as prior art. FIG. 10 (prior art) shows a current mirror and a membrane capacitor (C mem It is a structure in which an operational amplifier (Op-Amp)-based C-TIA is inserted between them. The C-TIA utilizes the feedback characteristics of the operational amplifier to strongly fix the drain terminal voltage of the current mirror to a constant voltage, acting like a virtual ground. Through this, the membrane potential (V mem Even if ) fluctuates, the current mirror operates stably, enabling high current linearity to be secured.

[0046] However, while these conventional C-TIA-based neuron circuits ensured linearity, they introduced two distinct limitations. First, there is the issue of high power consumption. Operational amplifiers, which are the core components of C-TIA circuits, consume high static power themselves, which is significant enough to account for approximately 60% of the total power consumption of the neuron circuit. This became a critical factor in reducing the overall energy efficiency of spiking neural network systems aimed at low power consumption. Second, there is the issue of low integration density. For stable operation, the C-TIA structure requires relatively large film capacitors (C mem It requires a membrane capacitor, and this capacitor occupies a large area, amounting to about 45% of the total area occupied by a single neuron. This became a serious obstacle to high integration, which increases the number of neurons per unit area of ​​the chip.

[0047] Therefore, in the hardware implementation of SNN systems, a new neuron circuit structure was required that could maintain high current linearity while solving the problems of high power consumption and large area of ​​C-TIA.

[0048] FIG. 1 is an overall block diagram of a current scaling integrator-based spiking neural network system according to one embodiment of the present invention.

[0049] Referring to FIG. 1, the spiking neural network system of the present invention comprises a synaptic current mirror that converts and transmits the weights of synapses into current, a neuron circuit that receives current from the synaptic current mirror and performs operations, and an output (V) of the neuron circuit. mem Includes a main comparator that generates spikes based on ).

[0050] The key feature of the present invention lies in the neuron circuit. Unlike the prior art (Fig. 9), which used a C-TIA that consumed a large amount of power and occupied a large area, the present invention replaces the C-TIA with a neuron circuit with a current scaling integrator.

[0051] Looking at the operation process, first, the input from the synapse array passes through the synapse current mirror to the active current (I exc ) and inhibition current (I inh The total input current (I) which is the difference of ) in It is converted into ) and transmitted to a neuron circuit with a built-in current scaling integrator. The neuron circuit scales and integrates the input current to obtain the membrane potential (V mem Creates ).

[0052] The generated membrane potential (V mem ) is transmitted to the main comparator. The main comparator is the membrane potential (V mem ) preset threshold voltage (V th Compare with ). If membrane potential (V mem ) is the critical voltage (V th When it rises above ), the main comparator generates an output spike (Post Spike) and transmits it to the next layer. At the same time, the main comparator sends a latch reset signal (Lat rstn It generates ), controls the latch (Lat) signal inside the neuron circuit, and the membrane potential (V mem ) reset voltage (V rst Initialize with ).

[0053] As such, the present invention can replace a power-consuming operational amplifier-based C-TIA with a current-scaling integrator-based neuron circuit capable of low power consumption and high integration, while still performing the integration and firing functions, which are the core operations of the entire spiking neural network system, in the same way.

[0054] FIG. 2 is a detailed circuit diagram of a current scaling integrator configured in the present invention, and FIG. 3 is a detailed circuit diagram of a super source follower applied in the present invention.

[0055] The neuron circuit of the present invention uses a current scaling integrator shown in FIG. 2 to replace the C-TIA (capacitance transimpedance amplifier) ​​of the prior art (Fig. 10). This circuit performs two functions. First, it scales the input current by a factor of 'n' to reduce the size of the film capacitor (high integration), and second, it ensures current linearity without the operational amplifier of the C-TIA (low power consumption).

[0056] First, the principle of current scaling is explained. Referring to Fig. 2, the current scaling integrator is V C1 and V C2 The voltage of each node is current source nI B2 and I B2 It is determined by. At this time, MOSFET M n1 is M n2 It is 'n' times the size. Here, 'n' represents a scaling factor greater than 1.

[0057] V C 1 and V C The equations for Kirchhoff's Current Law (KCL) at node 2 are as follows [Equation 1] and [Equation 2], respectively.

[0059] [Mathematical Formula 1]

[0060] I in + nI B2 = I1

[0062] Here, I in is the input current from the synapse, and nI B2 is a scaled current source, and I_1 is M nx It is the current flowing through 1.

[0064] [Mathematical Formula 2]

[0065] I out + I B2 = I1 / n

[0067] Here, I out is the scaled output current, and I B2 is a current source, and I1 / n is M nx It is the current flowing through 2.

[0068] Substituting the above [Equation 1] into [Equation 2] and rearranging, the film capacitor (C mem The final scaled output current I flowing into ). out The input current I is as shown in the following [Equation 3]. in It becomes 1 / n times.

[0070] [Mathematical Formula 3]

[0071] I out = I in / n

[0073] Through this principle, the input current I transmitted from the synapse in It is scaled down by 1 / n and transferred to the film capacitor. Therefore, the same film potential (V mem The film capacitor (C) required to achieve the change mem The capacity (size) of ) can also be reduced by a factor of 1 / n.

[0074] Next, the principle of ensuring linearity and stability by replacing the operational amplifier of the C-TIA is explained. The current scaling integrator of the present invention applies a super source follower structure as shown in FIG. 3.

[0075] Figure 3 is a detailed circuit diagram of a super source follower. A super source follower is a circuit that applies a negative feedback loop by adding a MOSFET (M2) to a general source follower (M1). This negative feedback mechanism stabilizes the output voltage of the circuit and significantly reduces the output impedance.

[0076] Referring again to the circuit diagram in Fig. 2, V xIt can be seen that super source followers are configured on both sides of the node. This structure can prevent input current loss due to low input impedance, and with low output impedance, similar to the operational amplifier of a C-TIA, a large load capacitor (C mem It can stably drive the synapse current mirror. In addition, this structure maintains a constant drain terminal voltage of the synapse current mirror, ensuring high current linearity without the need for a power-consuming operational amplifier.

[0077] FIG. 4 is an overall detailed circuit diagram of a neuron circuit utilizing a current scaling integrator according to the present invention, and FIG. 5 is a timing diagram to explain the operation of the neuron circuit of FIG. 4.

[0078] Referring to FIG. 4, the neuron circuit of the present invention includes a synaptic current mirror section that generates a current corresponding to a synaptic weight, and a current scaling integrator section that receives the synaptic current and performs scaling and integration.

[0079] The synaptic current mirror part is the active current (I exc ) and inhibition current (I inh It generates ), and the input current (I) corresponding to the difference between these two currents in ) is a single node (V cs It is transferred to the current scaling integrator through ).

[0080] The current scaling integrator is M as described in Fig. 2 n1 , M n2 , M nx1 , M nx2 It includes MOSFETs such as current sources and a super source follower structure. The input current (I in ) is the output current (I) scaled to 1 / n according to the principle of [Equation 3] out It is converted into ). This scaled output current (I out ) is a film capacitor (C mem Accumulated in the membrane potential (V mem Changes )

[0081] Figure 5 shows the operation process of this neuron circuit over time.

[0082] Input spike (PRE <1> ...PRE <n>When ) is applied to a neuron, the synaptic current mirror of Fig. 4 produces the corresponding synaptic current (I syn It generates ) this synaptic current (I syn ) is transferred to the current scaling integrator of FIG. 4 and scaled to 1 / n, after which the film capacitor (C mem Starts charging ).

[0083] This process is called the integration state, and in this state, the membrane potential (V mem ) is synaptic current (I syn Whenever ) is input, it rises in a stepwise manner as shown in Fig. 5.

[0084] Membrane potential (V) mem As ) continues to rise, the preset threshold voltage (V th When it reaches ), the comparator shown in Fig. 1 detects this and generates an output spike (POST).

[0085] Simultaneously, the comparator activates the latch (Lat) signal. Referring to FIG. 4, the latch (Lat) signal is V C 2. The voltage of the node, i.e., the reset voltage (V rst ) is the membrane potential (V mem The switch is operated to connect to the ) node. As a result, the membrane potential (V mem ) is the critical voltage (V th Reset voltage (V) at ) rst It discharges rapidly, and this state is called the reset state.

[0086] After the reset state is complete and the latch (Lat) signal is deactivated, the neuron circuit returns to the integration state and begins accumulation for the next input spike. This process of integration, firing, and reset is performed repeatedly.

[0087] The effects of the present invention will be further clarified below through specific embodiments. However, the following embodiments are merely illustrative of the present invention, and the scope of the present invention is not limited to the following embodiments.

[0089] Example 1: Effect of improving neuron density (area reduction)

[0090] The effect of the current scaling integrator (CSI) of the present invention on the integration density of a neuron circuit was verified. In this embodiment, the scaling factor 'n' was set to 2, and the input current (I in Scaling ) to 1 / 2 (I out = I in It was assumed that / 2)

[0091] Figure 6 is a layout comparison diagram illustrating the effect of reducing the area of ​​a film capacitor when applying current scaling (n=2) according to the present invention. As shown in Figure 6, when implementing the same LIF neuron operation, a film capacitor of 800 fF is required when n=1 (no scaling), whereas the same operation is possible with only a film capacitor of 400 fF when applying current scaling with n=2 according to the present invention. As a result, it can be confirmed that the total area of ​​the neuron circuit, particularly the area occupied by the capacitor, is significantly reduced.

[0092] The following [Table 1] compares the integration density and power characteristics of a conventional C-TIA-based neuron circuit and a current scaling integrator-based neuron circuit of the present invention.

[0093] [Table 1]

[0094]

[0095] Referring to [Table 1], the C-TIA-based neuron circuit uses a membrane capacitance of 800 fF, whereas the current-scaling integrator-based neuron circuit of the present invention uses a membrane capacitance of 400 fF through n=2 scaling. Due to this reduction in membrane capacitance, the neuron density (number of neurons per unit area) is 2.498 x 10⁻⁶ for the C-TIA-based circuit. 4 cm -2 Based on the present invention 3.523x10 4 cm -2 It was improved by about 40%. This demonstrates that the present invention contributes to the high integration of neuron circuits.

[0097] Example 2: Effect of ensuring current linearity

[0098] A simulation was performed to verify whether the neuron circuit of the present invention secures high current linearity even without the operational amplifier of the C-TIA.

[0099] Figure 7 is a graph comparing the current linearity of the current scaling integrator (CSI) of the present invention with that of a conventional C-TIA. Referring to Figure 7, in the relationship between the ideal current on the x-axis and the real current on the y-axis, it can be seen that the present invention (CSI) exhibits high linearity at a level almost identical to that of the prior art (C-TIA).

[0100] The following [Table 2] quantitatively shows the error rate of the output current relative to the input current in the current scaling integrator (CSI) of the present invention.

[0102] [Table 2]

[0103]

[0104] Referring to [Table 2], it can be seen that the current scaling integrator (CSI) of the present invention maintains a low error rate of approximately 1% across the entire input current range from 0.1 μA to 1.4 μA. This demonstrates that the present invention has successfully achieved high current linearity equivalent to that of C-TIA through a super source follower structure, etc., without using an operational amplifier that consumes a large amount of power.

[0106] Example 3: Power consumption reduction effect

[0107] The power consumption reduction effect of the neuron circuit of the present invention was confirmed by comparing it with a conventional C-TIA-based neuron circuit.

[0108] The following [Table 3] compares the power consumption of the C-TIA-based neuron circuit and the current scaling integrator-based neuron circuit of the present invention under the same synaptic current input conditions.

[0110] [Table 3]

[0111]

[0112] Referring to [Table 3], it can be seen that under various synaptic current conditions of 0.1 μA to 0.7 μA, the power consumption of the current scaling integrator-based neuron circuit of the present invention is reduced by an average of about 57% or more compared to the C-TIA-based neuron circuit.

[0113] In addition, referring again to [Table 1], based on a specific synaptic current of 400 nA, the power consumption of the C-TIA-based neuron circuit is 42.006 μW, whereas the neuron circuit of the present invention is 16.950 μW, showing a power saving effect of about 61%.

[0114] Figure 8 is a graph comparing the total energy consumed during the entire process of inferring the number '7' in the MNIST dataset (Inference 7). Referring to Figure 8, the C-TIA-based neuron consumed 2173 pJ of energy, whereas the neuron with the current scaling integrator (CSI) of the present invention consumed 861 pJ (862 pJ in Scaling Mode). This demonstrates that power consumption was significantly reduced by more than 60% by the present invention during the actual inference process.

[0115] Despite these low power and high integration characteristics, the present invention achieved a high accuracy of 95.41% in MNIST dataset inference, confirming that there is no performance degradation.

[0116] The following [Table 4] summarizes the detailed performance indicators of the system to which the present invention is applied. (Unit: %)

[0118] [Table 4]

[0119]

[0121] The method and system described above may be implemented as a program (or application) comprising an executable algorithm that can be executed on a computer. The program may be provided by storing it on a non-transitory computer-readable medium. Here, a non-transitory computer-readable medium refers to a medium that stores data semi-permanently and is readable by a device, rather than a medium that stores data for a short period of time, such as a register, cache, or memory. Specifically, the various applications or programs described above may be provided by storing them on non-transitory computer-readable media such as CDs, DVDs, hard disks, Blu-ray discs, USBs, memory cards, ROMs, etc.

[0122] The foregoing description is merely an illustrative explanation of the technical concept of the present invention, and those skilled in the art to which the present invention pertains may make various modifications and variations within the scope of the essential characteristics of the present invention. Accordingly, the embodiments disclosed in the present invention are intended to explain, not limit, the technical concept of the present invention, and the scope of the technical concept of the present invention is not limited by these embodiments. The scope of protection of the present invention shall be interpreted by the claims below, and all technical concepts within an equivalent scope shall be interpreted as being included within the scope of rights of the present invention.< / n>

Claims

Claim 1 In a spiking neural network system, the input current (I) corresponding to the synaptic weight in A synaptic current mirror section providing ); and a neuron circuit receiving the input current; wherein the neuron circuit comprises, wherein the input current (I in ) receiving the scaled output current (I) in which the input current is divided by a scaling factor 'n' (n>1). out = I in A current scaling integrator that generates ( / n); connected to the current scaling integrator, the scaled output current (I out Integrating ) to obtain the membrane potential (V mem A film capacitor generating ); and the film potential (V mem ) critical voltage (V th A spiking neural network system comprising: a comparator that generates an output spike by comparing with ); wherein the current scaling integrator comprises a first scaling MOSFET and a second scaling MOSFET, the gate terminals of which are connected to a common bias voltage, the source terminals of which are connected to the input and output terminals of the current scaling integrator, respectively, and the drain terminals of which are connected to a first bias current source and a second bias current source, respectively, wherein the ratio of channel width to channel length (W / L) of the first scaling MOSFET is n times the ratio of channel width to channel length (W / L) of the second scaling MOSFET, and the current supplied by the first bias current source is n times the current supplied by the second bias current source. Claim 2 delete Claim 3 A spiking neural network system according to claim 1, wherein the current scaling integrator comprises a source follower MOSFET having a follower input node and a follower output node, wherein the source follower MOSFET has a gate terminal connected to the follower input node and a source terminal connected to the follower output node, and a feedback MOSFET having a gate terminal connected to the drain terminal of the source follower MOSFET, wherein the drain terminal is connected to the follower output node and the source terminal is connected to ground, and wherein the source follower MOSFET and the feedback MOSFET form a negative feedback loop. Claim 4 delete Claim 5 A spiking neural network system according to claim 1, characterized in that when the scaling factor n is 2, the current scaling integrator generates a scaled output current that is 1 / 2 of the input current, and the capacitance of the film capacitor is set to 1 / 2 of the capacitance required when n is 1 to achieve the same film potential change. Claim 6 A spiking neural network system according to claim 3, wherein the synaptic current mirror unit includes an output MOSFET, the drain terminal of the output MOSFET is connected to the input terminal of the current scaling integrator to which the input current is delivered, and the super source follower structure maintains the drain voltage of the output MOSFET constant by maintaining the input terminal voltage of the current scaling integrator constant through the negative feedback loop. Claim 7 delete Claim 8 delete Claim 9 In a signal processing method of a spiking neural network, the synaptic current mirror part has an input current (I) corresponding to a synaptic weight. in A step of providing ) to the input terminal of a current scaling integrator; wherein the current scaling integrator utilizes a first scaling MOSFET and a second scaling MOSFET, wherein, according to a scaling factor n (n>1), the gate terminals are connected to a common bias voltage, the source terminals are connected to the input and output terminals of the current scaling integrator, respectively, and the drain terminals are connected to a first bias current source and a second bias current source, respectively, wherein the ratio of channel width to channel length (W / L) of the first scaling MOSFET is n times the ratio of channel width to channel length (W / L) of the second scaling MOSFET, and the current supplied by the first bias current source is n times the current supplied by the second bias current source, and the input current (I in The scaled output current (I) is obtained by dividing ) by the scaling factor n. out =I in A step of generating / n); a film capacitor connected to the output terminal of the current scaling integrator, which generates the scaled output current (I out Integrating the membrane potential (V mem A step of generating ); and a comparator generating the membrane potential (V mem ) critical voltage (V th A spiking neural network signal processing method characterized by including the step of generating an output spike by comparing with ). Claim 10 A spiking neural network signal processing method according to claim 9, wherein the step of generating the scaled output current is performed by utilizing a super source follower in which the source follower MOSFET and the feedback MOSFET form a negative feedback loop, wherein the source follower MOSFET and the feedback MOSFET form a negative feedback loop, and the voltage of the follower output node is stabilized by the negative feedback loop. Claim 11 A spiking neural network signal processing method according to claim 9, wherein when the scaling factor n is 2, the current scaling integrator generates a scaled output current that is 1 / 2 of the input current, and the capacitance of the film capacitor is set to 1 / 2 of the capacitance required when n is 1 to achieve the same film potential change. Claim 12 delete

Citation Information

Patent Citations

  • Methods and systems for memristor-based neuron circuits

    US20120011092A1