KR102495823B1 - Artificial neuron device using threshold switch - Google Patents

Abstract

An artificial neuron element according to an embodiment of the present invention includes a threshold switch connected between an input terminal and a first node; and a load resistance connected between the first node and a ground terminal, wherein when an input voltage of a constant level is applied to the input terminal over time, a spike current flows through the load resistance, When the level of the input voltage applied to the input terminal increases, a period in which the spike frequency of the spike current flowing through the load resistor increases occurs.

Description

Artificial neuron device using threshold switch {ARTIFICIAL NEURON DEVICE USING THRESHOLD SWITCH}

The present invention relates to an artificial neural circuit, and more particularly, to an artificial neuron device including a threshold switch.

Neurons are nerve cells specially differentiated for information processing in living organisms. A neuron consists of three main parts: a soma, an axon, and a dendrite. Axons are nerve fibers that carry signals from the cell body to other neurons. Dendrites are fibers that receive signals from other neurons. That is, the end of the axon of another neuron is connected to the dendrite. This connection is called a synapse.

On average, a single neuron receives signals from thousands and even tens of thousands of axons. Therefore, in order for one neuron to combine with multiple axons, it is necessary to complicate the dendrites of the neuron and expand the places where it can combine with the axons. The cell body performs calculations on a number of input signals, and the answers are relayed to other neurons via the axon. What kind of calculation (or information processing) to be performed in this neural network is distributed and stored in the connection part of the dendrites, that is, the synapse.

In general, the potential inside a neuron is lower than that outside. However, when an input signal arrives from the outside, the internal membrane potential gradually increases. When the elevated membrane potential reaches a specific potential, that is, a threshold or threshold, the neuron generates a current in the form of a pulse and the membrane potential is lowered again. At this time, the neuron is said to have fired, and this is transmitted to other neurons as an electrical signal.

An artificial neural circuit is a circuit that performs an operation similar to that of a biological nerve system. Artificial neural circuits can be used in various fields such as text recognition, image recognition, voice recognition, and face recognition. Artificial neural circuits can be implemented to some degree in terms of circuits using technologies such as very large scale integrated circuits (VLSI). However, it is not easy to implement complex characteristics and behaviors of the biological nervous system in a circuit using conventional methods. In addition, in order to implement by the conventional method, the area of the artificial neural circuit is widened, and there is a problem in that a lot of power is consumed even if implemented as a circuit.

The present invention is to solve the above-described technical problems, and an object of the present invention is to provide an artificial neuron device having a very simple structure that satisfies the characteristics and behaviors of biological neurons and enables high integration and low power consumption. there is.

An artificial neuron element according to an embodiment of the present invention includes a threshold switch connected between an input terminal and a first node; and a load resistance connected between the first node and a ground terminal, wherein when an input voltage of a constant level is applied to the input terminal over time, a spike current flows through the load resistance, When the level of the input voltage applied to the input terminal increases, a period in which the spike frequency of the spike current flowing through the load resistor increases occurs.

As an example, the threshold switch may be an ovonic threshold switch (OTS). The threshold switch includes an upper electrode, a TS material, and a lower electrode, and the TS material may be an oxide containing at least one of Nb, Ta, and V. In addition, the TS material may include at least one of Se and Te and at least one of Si, Ge, and As. The threshold switch has an on-resistance state or an off-resistance state according to a threshold voltage (Vth), a holding voltage (Vh), and a previous resistance state.

According to an embodiment of the present invention, an artificial neural network device, a large-scale brain-inspired computing system, and an artificial intelligence (AI) system because the artificial neuron device shows well the Integrate-and-Fire function and rate coding ability of biological neurons. etc. can be widely used. In addition, since the artificial neuron device according to an embodiment of the present invention is designed with only a threshold switch (TS) and a resistance device, power consumption can be drastically reduced compared to artificial neuron devices based on conventional Si-MOSFETs.

1 is a graph showing stimulation and response of biological neurons.
2 is a circuit diagram showing an artificial neuron device according to an embodiment of the present invention.
FIG. 3 is a graph showing a voltage-current change of a threshold switch (TS) of the artificial neuron device shown in FIG. 2 as an example.
FIG. 4 is a schematic diagram showing the structure of the threshold switch TS shown in FIG. 2 as an example.
5 and 6 are graphs for exemplarily explaining the load resistance dependence (Rload-dependence) of the threshold switch shown in FIG. 2 .
FIG. 7 is a graph for explaining the rate coding ability of the artificial neuron device shown in FIG. 2 .
FIG. 8 is a block diagram showing an artificial neural chip including the artificial neuron device shown in FIG. 2 .
FIG. 9 is a block diagram showing a user device including the artificial neural chip shown in FIG. 8 .

Hereinafter, embodiments of the present invention will be described clearly and in detail to the extent that those skilled in the art can easily practice the present invention.

1 is a graph showing stimulation and response of biological neurons. Figure 1 shows stimuli and responses expressed in neurons present in the optic nerve of locusts as an example. Referring to FIG. 1 , biological neurons tend to show certain response patterns to stimuli. When a temporally constant stimulus is applied to a biological neuron, a membrane potential (Vmem) increases. When the membrane potential increases and exceeds a threshold, neurons generate a spike current (Is). In Figure 1, the response (response) shows the change of the spike current (Is) for the intensity of several stimuli.

For example, when a certain level of stimulation a is applied to a biological neuron, the membrane potential (Vmem) increases until time t1, and then a spike current Is is generated at time t1. This neural response pattern is called Integrate-and-Fire. Biological neurons generally have an Integrate-and-Fire function (I&F function).

Biological neurons show a characteristic in which a spike rate gradually increases as the intensity of stimulation increases. Here, the spike rate is also referred to as the firing rate. Referring to FIG. 1 , it can be seen that the number of spikes generated in biological neurons increases as the intensity of stimulation increases to a, b, c, and d. For example, it can be seen that 1 spike occurs for the a stimulus, 4 spikes occur for the b stimulus, and more spikes occur for the c and d stimuli. A response of a biological nerve in which the spike rate increases as the intensity of stimulation increases is called rate coding. Biological neurons generally have rate coding ability.

Since the characteristics of these biological neurons are complex and delicate, it is not easy to implement an artificial neuron device that satisfies all the characteristics of biological neurons or to manufacture an artificial neuron network chip including the same. However, the artificial neuron device described below may satisfy various characteristics of the biological neuron described above.

2 is a circuit diagram showing an artificial neuron device according to an embodiment of the present invention. The artificial neuron device 100 shown in FIG. 2 well shows the stimulation and response characteristics of the biological neuron shown in FIG. 1 . Therefore, the artificial neuron device 100 of the present invention can be used in a large-scale brain-inspired computing system or an artificial intelligence (AI) system.

Referring to FIG. 2 , the artificial neuron device 100 may be implemented by including a threshold switch 110 and a load resistor 120 . Threshold switch 110 is connected between the input terminal and the N node. The input terminal receives the input voltage (Vin). A load resistor 120 is connected between the N node and the ground terminal. Here, the load resistor 120 may be implemented as an active resistive element (eg, a field effect transistor (FET)) or a passive resistive element (eg, a resistor). When the cupping resistor 120 is implemented as an active resistor, the resistance value may be adjusted according to the value of the gate voltage. Also, the load resistance 120 may be an equivalent resistance of an external device viewed from the threshold switch 110 (TS). Also, the current I flowing through the load resistor 120 may be an output current.

The voltage (Vin) between the input terminal and the ground terminal may be expressed as the sum of the voltage (V TS ) across the threshold switch 110 ( _TS ) and the voltage (V _load ) across the load resistor 120 . That is, Vin=V _TS +V _load . Here, V _load =I*R _load . If the resistance of the threshold switch 110 (TS) is R _TS , then V _TS =I*R _TS .

Here, R _TS may be defined as a resistance value of a variable resistor that changes according to a bias voltage applied to TS or a previous state. R _TS can be expressed by Equation 1 as follows.

Referring to Equation 1, it can be seen that the resistance state of R _TS changes according to Vth and V _H . Here, Vth is the threshold voltage of TS and V _H is the holding voltage. When V _TS is higher than Vth, TS is in the on-resistance state (Ron). When V _TS is lower than V _H , TS is in an off resistance state (Roff).

Meanwhile, the resistance state of R _TS may change according to the previous resistance state. When V _TS is higher than V _H and lower than Vth (V _H < V _TS < Vth), the previous resistance state is maintained. That is, if the previous resistance state of the TS is the on-resistance state (Ron), the on-resistance state is maintained, and if the previous resistance state of the TS is the off-resistance state (Roff), the off-resistance state is maintained. The TS resistance (R _TS ) expressed in Equation 1 is due to the material characteristics of TS and its voltage-current change. The voltage-current change of the TS will be described in more detail with reference to FIG. 3 .

FIG. 3 is a graph showing a voltage-current change of a threshold switch (TS) of the artificial neuron element shown in FIG. 2 exemplarily. In FIG. 3 , the horizontal axis represents the voltage (V _TS ) applied to the TS, and the vertical axis represents the current (I _TS ) flowing through the TS.

Referring to FIG. 3 , when the TS voltage (V _TS ) is lower than the threshold voltage (Vth), TS is in a non-conductive state. No current flows through TS. In the example of FIG. 3 , no current flows through the TS until the TS voltage (V _TS ) reaches about 5V. TS is in an insulator state, that is, in an off resistance state (Roff). When V _TS reaches the threshold voltage (Vth, about 5V), TS is in a conductive state. The current flowing through the TS increases rapidly. TS changes to a conductive state in which I _TS rapidly increases, that is, an on-resistance state (Ron).

In this on-resistance state (Ron), when V _TS is lowered, I _TS starts to decrease linearly. When I _TS decreases and reaches a holding current (Ih), TS changes to an off resistance state (Roff) again. That is, when V _TS becomes the holding voltage (V _H , about 1V), TS changes back to a non-conductive state, that is, to an off-resistance state (Roff). In this way, TS changes from an off resistance state Roff to an on resistance state Ron, or vice versa, from an on resistance state Ron to an off resistance state Roff according to the voltage V _TS .

Referring to the graph of FIG. 3 , the following relational expression may be established according to the resistance state of the threshold switch TS 110 of FIG. 2 .

(1) If TS is off-storage state (Roff), then R _TS >> R _load , so V _TS >>V _load .

(2) When TS is in the on-resistance state (Ron), R _TS << R _load , so V _TS << V _load .

(3) When TS is in the on resistance state and V _TS < V _H , TS performs one cycle of oscillation and returns to the off state.

If the load resistance (R _load ) sufficiently satisfies the relationship of (2) and (3) above, the threshold switch (TS, 110) may serve as an oscillator.

FIG. 4 is a schematic diagram showing the structure of the threshold switch TS shown in FIG. 2 as an example. Referring to FIG. 4 , the TS may have a cross-point structure including an upper electrode TE, a threshold switch material TS, and a lower electrode BE.

A conductive material or a semiconductor material may be used as the upper and lower electrodes TE and BE. The materials of the threshold switch (TS) showing current-voltage characteristics as shown in FIG. 3 include a Ge-Se-Te compound, a Si-Se-Te compound, a Si-As-Te-GP compound, and a Nb-Ta-O compound having a special composition. , Mo-WO compounds and the like can be used. For example, the TS material is an oxide containing at least one of Nb, Ta, and V (eg, Nb ₂ O ₅ ), or a compound containing at least one of Se and Te and at least one of Si, Ge, and As (eg, Nb 2 O 5 ). , Ge ₆₀ Se ₄₀ ). Also, the TS material may be an ovonic threshold switch (OTS).

The upper and lower electrodes (TE, BE) of the TS and the threshold switch material may be manufactured by photolithography and lift-off technology. The lower electrode, the TS material, and the upper electrode are deposited by RF magnetron-sputtering technology, and each may have a thin film thickness of about 100 nm.

5 and 6 are graphs for exemplarily explaining the load resistance dependence (Rload-dependence) of the threshold switch shown in FIG. 2 . 5 and 6 show changes in V _TS and I _TS according to a change in the load resistance (R _laod ) of FIG. 2 . 5 is a graph of experimental results measured for 30 μ when the load resistance (R _load ) is 1KΩ, 2KΩ, 5.6KΩ, and 10KΩ. 6 is a graph of experimental results measured for 100μ when the load resistance (R _load ) is 23.67KΩ and 51KΩ. Referring to FIGS. 5 and 6 , it can be seen that as the magnitude of the load resistance R _load increases, the current spike value gradually decreases and the interval between spikes increases.

Continuously referring to FIGS. 5 and 6 , it can be seen that the threshold switch shown in FIG. 2 performs an Integrate-and-Fire function. That is, when a square pulse DC voltage is applied to the input terminal, the threshold switch TS 110 has characteristics similar to the membrane potential and spike current of a biological neuron. The Integrate-and-Fire function of the threshold switch (TS, 110) may occur through the following mechanism.

When V _TS reaches Vth, TS turns into an on-resistance state (Ron), and I _TS increases rapidly. As shown in FIG. 5, the artificial neuron device 100 generates a first spike current (Is) at about 5 μs. After the first spike current occurs, V _TS drops to the holding voltage (V _H ), and TS turns into an off resistance state (Roff). The artificial neuron device 100 may repeatedly generate a spike current Is while repeating this operation for 30 μs up to about 35 μs.

In the artificial neuron device 100 according to an embodiment of the present invention, V _TS may correspond to a membrane potential of a biological neuron, and I _TS may correspond to a spike current. Similar to a biological neuron, the artificial neuron device 100 may generate a spike current when V _TS increases and TS reaches the threshold voltage Vth. That is, the artificial neuron device 100 can perform the integral-and-fire function seen in biological neurons.

FIG. 7 is a graph for explaining the rate coding ability of the artificial neuron device shown in FIG. 2 . 7 shows a change in spike current according to a change in the input voltage Vin of the artificial neuron device 100 .

In FIG. 7, looking at the change of I _TS with time for each input voltage Vin, the spike frequency increases when the input voltage Vin increases by 0.2V from 4.2V to 6.0V. Able to know. Here, the increase in the input voltage Vin may correspond to the increase in the intensity of stimulation of the biological neuron as a, b, c, and d in FIG. 1A. The increase in the spike frequency of I _TS whenever the input voltage (Vin) of the artificial neuron device 100 increases by 0.2V is consistent with the increase in the spike rate as the intensity of stimulation increases in the biological neuron in FIG. 1a. can respond That is, it can be seen that the rate coding ability of biological neurons also occurs in the artificial neuron device 100 shown in FIG. 2 .

The artificial neuron device 100 according to an embodiment of the present invention shows the Integrate-and-Fire function and rate coding capability of biological neurons shown in FIGS. 1A to 1C. For this reason, the artificial neuron device 100 of the present invention can be widely used in an artificial neural network device, a large-scale brain-inspired computing system or an artificial intelligence (AI) system.

FIG. 8 is a block diagram showing an artificial neural chip including the artificial neuron device shown in FIG. 2 . Referring to FIG. 8 , an artificial neural chip 1000 includes a first artificial neuron element 1110, a second artificial neuron element 1120, and an artificial synapse 1200. Although the artificial neural chip 1000 is simply shown in FIG. 8 , row lines (RL) extending horizontally from the first artificial neuron element 1110 and vertically extending from the second artificial neuron element 1120 An artificial synapse 1200 may be disposed at an intersection of a column line (CL) to be formed.

The first artificial neuron element 1110 may transmit an electrical signal to the artificial synapse 1200 through the row line RL. The second artificial neuron element 1120 may receive an electrical signal from the artificial synapse 1200 through the column line CL. Conversely, the electrical signal may be transferred from the second artificial neuron element 1120 to the first artificial neuron element 1110 via the artificial synapse 1200 .

The artificial synapse 1200 may include a variable resistive layer and two electrodes. For example, the artificial synapse 1200 may include a first electrode electrically connected to the first artificial neuron element 1110 and a second electrode electrically connected to the second artificial neuron element 1120 . The variable resistance layer of the artificial synapse 1120 may have multiple resistance levels. The variable resistance layer of the artificial synapse 1200 includes metal oxides such as transition metal oxides or perovskite-based materials, phase change materials such as chalcogenide-based materials, It may include at least one of a ferroelectric material and a ferromagnetic material. The artificial synapse 1200 may be changed into an insulator state or a conduction state according to the number of inputs of electrical signals inputted from the first and second artificial neuron elements 1110 and 1120, time difference, and/or voltage difference.

FIG. 9 is a block diagram showing a user device including the artificial neural chip shown in FIG. 8 . Referring to FIG. 9 , a user device 2000 includes an artificial neural chip 2100, a control unit 2200, and a power unit 2300. The user device 2000 including the artificial neural chip 2100 may be used in various application fields such as voice recognition or artificial intelligence. The control unit 2200 may include a processing unit for controlling the operation of the artificial neural chip 2100, a signal processing unit for signal processing, and a memory unit for storing data or driving an algorithm. The power unit 2300 may provide power necessary for the operation of the artificial nerve chip 2100 and the control unit 2200 .

As described above, since the artificial neuron device 100 according to an embodiment of the present invention shows the Integrate-and-Fire function and rate coding ability of biological neurons well, an artificial neural network device, large-scale brain-inspired computing It can be widely used in systems, artificial intelligence (AI) systems, etc. In addition, since the artificial neuron device 100 according to an embodiment of the present invention is designed with only a TS and a resistance device, power consumption can be drastically reduced compared to artificial neuron devices based on conventional Si-MOSFETs.

100: artificial neuron element
1000: artificial neural chip
1110, 1120: artificial neuron element
1200: artificial synapse
2000: user device
2100: artificial neural chip
2200: control unit
2300: power unit

Claims (7)

a threshold switch connected between the input terminal and the first node; and
Including a load resistor connected between the first node and a ground terminal,
When a constant level of input voltage is applied to the input terminal over time, a spike current flows through the load resistor, and when the level of the input voltage applied to the input terminal increases, a spike current flows through the load resistor. A period in which the spike frequency of the spike current increases,
The resistance of the off resistance state of the threshold switch is greater than the load resistance,
The load resistance is greater than the resistance of the on-resistance state of the threshold switch,
The artificial neuron device of claim 1 , wherein the threshold switch includes an upper electrode, a TS material, and a lower electrode.
According to claim 1,
The threshold switch has the off resistance state in which resistance is high in a state where no voltage is applied,
The threshold switch is in the on-resistance state with a small resistance when a voltage higher than the threshold switching voltage is applied in the off-resistance state,
When the voltage applied to the threshold switch is less than the holding voltage in the on-resistance state of the threshold switch, the threshold switch performs at least one cycle of oscillation to return to the off-resistance state;
The threshold switching voltage is greater than the holding voltage,
The artificial neuron device, characterized in that the threshold switch is an ovonic threshold switch (OTS).
According to claim 1,
The artificial neuron element of claim 1, wherein the TS material is an oxide containing at least one of Nb, Ta, and V. According to claim 1,
The artificial neuron element of claim 1, wherein the TS material is a compound containing at least one of Se and Te and at least one of Si, Ge and As. delete According to claim 1,
The load resistance is an artificial neuron device implemented as an active resistance device. According to claim 1,
The load resistor is an artificial neuron device implemented as a passive resistance device.

Images