NEURON ELEMENT, DIFFERENTIAL NEURON ELEMENT, A/D CONVERTER AND NEURAL NETWORK

Abstract

This neuron element comprises: a first charge storage unit in which charge is stored by an input signal; and a signal processing unit that, if the charge stored in the first charge storage unit exceeds a first prescribed amount, outputs a second prescribed amount of charge from the first charge storage unit and generates a pulse signal. The signal processing unit performs prescribed processing for the second prescribed amount of charge.

Description

BACKGROUND

Technical Field

The present invention relates to a neuron element, a differential neuron element, an A/D converter, and a neural network.

Description of Related Art

Conventionally, a spiking neural network circuit may comprise a synapse configured to generate an operation signal from a weighted input spike signal, and a plurality of neuron elements configured to generate an output spiking signal based on comparing a voltage of a membrane signal generated by an input of the operation signal with a predetermined threshold value (see U.S. Pat. Application Publication No. 2020/0160146).

Conventionally, the operation signal is given as a current signal. In each neuron element, its internal capacitor is charged by the current signals inputted from a plurality of neuron elements in a preceding stage. Then, when a voltage is generated at the internal capacitor and the voltage exceeds the predetermined threshold value, the neuron element generates an output spiking signal.

Aside from this, an A/D converter which converts an analog signal into a digital signal using a sigma-delta modulation is conventionally known (see Kenji Taniguchi, “Introduction to CMOS Analogue Circuitry for LSI-Design”, CQ Publishing Co.,Ltd., Dec. 1, 2004 and “Basics of Delta-Sigma ADCs: About Delta-Sigma Modulators″, retrieved in Jan. 12, 2021”, Internet URL: https://e2e.ti.eom/blogs_/japan/b/analog/archive/2017/02/03/1-adc).

FIG. 15 is a circuit diagram illustrating a basic configuration of the conventional sigma-delta A/D converter. The sigma-delta A/D converter includes a sigma-delta modulating unit that performs the sigma-delta modulation on an input analog signal and generates a 1-bit modulated pulse signal, and a digital filter that receives the modulated pulse signal and converts a number of pulses being received within a predetermined range on a time axis, for example, into an n-bit output digital signal.

BRIEF SUMMARY

The present invention has been proposed to solve the problems posed by such prior art.

In accordance with various embodiments of the present disclosure, a neuron element is provided. In some embodiments, the neuron element comprises: a first charge accumulating unit configured to accumulate charges by an input signal; a signal processing unit which, based on a condition when the charges accumulated in the first charge accumulating unit exceed a first predetermined amount of charges, is configured to: discharge a second predetermined amount of charges from the first charge accumulating unit; generate a pulse signal; and perform a predetermined processing in relation to the second predetermined amount of charges.

In accordance with various embodiments of the present disclosure, a differential neuron element is provided. In some embodiments, the differential neuron element comprises: a charge accumulating unit which accumulates charges by a first input signal and a second input signal, the second input signal obtained by inverting a polarity of the first input signal; a signal processing unit which, based on a condition when the charges accumulated in the charge accumulating unit exceed a first predetermined amount of charges, is configured to: discharge a second predetermined amount of charges from the charge accumulating unit; generate a first pulse signal and a second pulse signal, the second pulse signal obtained by inverting a polarity of the first pulse signal; and perform a predetermined processing in relation to the second predetermined amount of charges.

In accordance with various embodiments of the present disclosure, an A/D converter is provided. The A/D converter comprises: a charge accumulating unit which accumulates charges by an input signal; a signal processing unit which, based on a condition when the charges accumulated in the charge accumulating unit exceed a first predetermined amount of charges, is configured to: discharge a second predetermined amount of charges from the charge accumulating unit; generate a pulse signal; perform a predetermined processing in relation to the second predetermined amount of charges; and a digital filter configured to perform a predetermined digital processing on the pulse signal generated by the signal processing unit.

In accordance with various embodiments of the present disclosure, a neural network is provided. The neural network comprises a plurality of neuron elements, wherein an input signal of any one of the neuron elements is obtained by a weighting addition processing of at least one of signals of at least one external input signal and an output signal of at least another one of the neuron elements, wherein an output signal of any of the neuron element becomes at least one of signals of at least one external output signal and a weighted input signal of the at least another one of the neuron elements.

The present invention makes it possible to suppress input signal data loss and power consumption.

BRIEF DESCRIPTION OF THE FIGURES

FIG. 1 is a block diagram showing a functional configuration of a neuron element according to a first embodiment.

FIG. 2 is a flowchart showing operation steps of a neuron element.

FIG. 3 is a circuit diagram showing a configuration of a neuron element.

FIG. 4 is a timing chart showing temporal changes of respective signals in a neuron element.

FIG. 5 is a timing chart showing temporal changes of respective signals in a neuron element when a delay period becomes longer.

FIG. 6 is a circuit diagram showing a configuration of a neuron element according to a second embodiment.

FIG. 7 is a timing chart showing temporal changes of respective signals in a neuron element.

FIG. 8 is a timing chart showing temporal changes of respective signals in a neuron element when it is most likely to cause a condition that does not operate normally.

FIG. 9 is a circuit diagram showing a configuration of a neuron element according to a third embodiment.

FIG. 10 is a block diagram showing a functional configuration of an A/D converter according to a fourth embodiment.

FIG. 11 is a set of wave diagrams respectively showing an input analog signal, a modulated pulse signal, and an output digital signal, of a conventional sigma-delta A/D converter.

FIG. 12 is a set of wave diagrams respectively showing an input analog signal, a modulated pulse signal, and an output digital signal, of an A/D converter according to a fourth embodiment.

FIG. 13 is a diagram showing an exemplary configuration of a neural network according to a fifth embodiment.

FIG. 14 is a diagram showing a relation between an output signal x_i of a neuron element (i = 1,2,3) and a weighting coefficient w_i (i = 1,2,3).

FIG. 15 is a block diagram showing a functional configuration of a conventional sigma-delta A/D converter.

DETAILED DESCRIPTION OF VARIOUS EMBODIMENTS

Hereinafter, embodiments in the present invention will be described in detail below with reference to the figures.

FIG. 1 is a block diagram showing a functional configuration of a neuron element 10 according to the first embodiment.

The neuron element 10 comprises an input unit 11, a charge accumulating unit 12, a discharging unit 13, a pulse signal generating unit 14, an output unit 15, and a controlling unit 16.

The input unit 11 supplies a current signal input from an outside (an input signal) to the charge accumulating unit 12. When the input signal is a voltage signal, for example, the input unit 11 converts the voltage signal to a current signal equivalent to the voltage signal using a predetermined method.

The charge accumulating unit 12 is supplied with the current signal from the input unit 11 to accumulate charges.

The discharging unit 13 discharges the charges accumulated in the charge accumulating unit 12 according to an instruction of the controlling unit 16 or according to a pulse signal generated by the pulse signal generating unit 14.

The pulse signal generating unit 14 generates the pulse signal based on an instruction of the controlling unit 16 or in response to the charges discharged by the discharging unit 13, and supplies the pulse signal to the output unit 15.

The output unit 15 outputs the pulse signal supplied from the pulse signal generating unit 14 to the outside. The output unit 15, if required, may output the pulse signal to the outside after shaping a waveform of the pulse signal supplied from the pulse signal generating unit 14 to a predetermined shape.

The controlling unit 16 measures, for example, a voltage generated at the charge accumulating unit 12 as an amount of accumulated charges in the charge accumulating unit 12. The controlling unit 16 instructs, to the pulse signal generating unit 14, a generation of a pulse signal when an amount of accumulated charges being measured (a measured amount of charges) exceeds a predetermined threshold value, and a generation stop of the pulse signal when the measured amount of charges no longer exceeds the threshold value.

Alternatively, the controlling unit 16 instructs, to the discharging unit 13, a discharge of a predetermined amount of charges from the charge accumulating unit 12 when a measured amount of charges exceeds a predetermined threshold value, and a discharge stop of charges from the charge accumulating unit 12 when the measured amount of charges no longer exceeds the threshold value.

The discharging unit 13 and the pulse signal generating unit 14 do not immediately execute a predetermined processing upon receiving an instruction from the controlling unit 16, but execute the predetermined processing after a certain period of time elapses. Also, the discharging unit 13, the pulse signal generating unit 14 and the controlling unit 16 are linked to each other as described above. For example, the discharging unit 13, the pulse signal generating unit 14 and the controlling unit 16 collectively function as the signal processing unit 20, which, when a measured amount of charges in the charge accumulating unit 12 exceeds a predetermined threshold value, discharges a predetermined amount of charges from the charge accumulating unit 12, and generates a pulse signal.

FIG. 2 is a flowchart showing operation steps of the neuron element 10.

The input unit 11, when there is an input signal from an outside (step S1), receives the input signal and supplies it to the charge accumulating unit 12 (step S2). Alternatively, the input unit 11, when the input signal is a voltage signal, converts the voltage signal to an equivalent current signal using a predetermined method, and supplies the converted current signal to the charge accumulating unit 12.

The charge accumulating unit 12 accumulates charges when the input signal is supplied from the input unit 11 (step S3).

The controlling unit 16 measures, for example, a voltage generated at the charge accumulating unit 12 as an amount of accumulated charges in the charge accumulating unit 12 (step S4) and, if the measured amount of accumulated charges exceeds a predetermined threshold value (step S5), the controlling unit 16 instructs, to the discharging unit 13, a discharge of a predetermined amount of charges. Accordingly, the discharging unit 13 discharges the predetermined amount of charges from the charge accumulating unit 12 (step S6). Alternatively, if the measured amount of charges in the charge accumulating unit 12 does not exceed the predetermined threshold value, the charges in the charge accumulating unit 12 is not discharged and the process returns to step S1.

Next, the controlling unit 16 instructs the pulse signal generating unit 14 to generate a pulse signal. Then, the pulse signal generating unit 14 generates the pulse signal (step S7). Alternatively, the pulse signal generating unit 14 may generate the pulse signal in accordance with an amount of charges discharged by the discharging unit 13 even without receiving the instruction from the controlling unit 16.

The output unit 15 outputs the pulse signal generated by the pulse signal generating unit 14 to the outside (step S7). Then, the process returns to step S4. Alternatively, the output unit 15, if required, may output the pulse signal to the outside after shaping a waveform of the pulse signal generated by the pulse signal generating unit 14 to a predetermined shape.

Alternatively, the controlling unit 16 may, when the measured amount of charges in the charge accumulating unit 12 exceeds the threshold value, firstly issue the instruction to the pulse signal generating unit 14, and then issue the instruction to the discharging unit 13. Alternatively, the discharging unit 13 may discharge the predetermined amount of charges in response to the pulse signal generated by the pulse signal generating unit 14, without receiving the instruction from the controlling unit 16.

Here, from a time when the measured amount of charges in the charge accumulating unit 12 exceeds the threshold value to a time when charges are actually discharged from the charge accumulating unit 12, there is a predetermined delay period. Therefore, charges are continuously accumulated in the charge accumulating unit 12 during this delay period. Similarly, there is a predetermined delay period between when the measured amount of charges in the charge accumulating unit 12 falls below the threshold value and when the discharge of charges from the charge accumulating unit 12 actually stops. Therefore, charges are continuously discharged from the charge accumulating unit 12 during the latter delay period.

As described above, when an amount of charges accumulated in the charge accumulating unit 12 exceeds a predetermined threshold value, the neuron element 10 outputs a pulse signal to an outside and discharges a predetermined amount of charges from the charge accumulating unit 12, thereby preventing the charge accumulating unit 12 from overflowing with the charges. In other words, by outputting the pulse signal, which corresponds to a quantity of charges overflowing from the charge accumulating unit 12 to the outside, the neuron element 10 can avoid a loss of data corresponding to an amount of the discarded charges.

FIG. 3 is a circuit diagram showing a configuration of a neuron element 10A according to the present embodiment. The neuron element 10A is an exemplary configuration of the neuron element 10 shown in FIG. 1.

The neuron element 10A includes a capacitor C_A, a capacitor C_B, a comparator CMP, switches SWA and SWB, and a resistive load R. The switch SWA is, for example, a SPDT (Single-Pole Double-Throw) switch using the MOS transistors, and has three terminals, SW1, SW2 and SW3. The switch SWA conducts between the terminals SW1 and SW3, or between the terminals SW2 and SW3, depending on a control (selecting) signal from the comparator CMP. The switch SWB also has a same configuration as the switch SWA.

One terminal of the capacitor C_A is connected to an input terminal IN, and the other terminal thereof is grounded. One terminal of the capacitor C_B is connected to the terminal SW3 of the switch SWA, and the other terminal thereof is connected to the terminal SW3 of the switch SWB.

The terminal SW1 of the switch SWA is connected to the input terminal IN. The terminal SW1 of the switch SWB is grounded.

The terminal SW2 of the switch SWA is connected to the resistive load R and an output terminal OUT. The terminal SW2 of the switch SWB is grounded.

One terminal of the resistive load R is connected to the terminal SW2 of the switch SWA and the output terminal OUT, the other terminal is grounded. For this reason, the resistive load R has a function to generate a pulse signal in accordance with an amount of charges discharged from the capacitor C_B via the switch SWA.

The comparator CMP measures an amount of charges accumulated in the capacitor C_A as a voltage of the input terminal IN (= a voltage generated at the capacitor C_A) and determines whether or not the voltage being measured (the measured voltage v_A) exceeds a predetermined threshold voltage V_th. The comparator CMP supplies a selecting signal SW1 to the switches SWA and SWB when the measured voltage v_A exceeds the threshold voltage V_th, or supplies a selecting signal SW2 to the switches SWA and SWB when the measured voltage v_A does not exceed the threshold voltage V_th.

When the selecting signal is supplied from the comparator CMP, the switches SWA and SWB are switched as follows. That is, the switches SWA and SWB, when the selecting signal SW1 is supplied, conduct between the terminals SW1 and SW3 (switched to SW1). Furthermore, the switches SWA and SWB, when the selecting signal SW2 is supplied, conduct between the terminals SW2 and SW3 (switched to SW2).

Therefore, when the selecting signal SW1 is supplied to the switches SWA and SWB, the capacitors C_A and C_B are connected in parallel. On the other hand, when the selecting signal SW2 is supplied to the switches SWA and SWB, the capacitor C_B and the resistive load R are connected in parallel, and the charges accumulated in the capacitor C_B are discharged through the resistive load R.

Note that the other terminal of the capacitor C_A, the terminals SW1 and SW2 of the switch SWB, and the other terminal of the resistive load R are sufficient if they are equipotential, and are not limited to be grounded as shown in the present embodiment.

By the way, the switches SWA and SWB do not switch immediately when the measured voltage v_A during its ramping up exceeds the threshold voltage V_th or when the measured voltage v_A during its dropping falls below the threshold voltage V_th. That is, there is a predetermined delay period D_cmp from a time when the ramping up or dropping of the measured voltage v_A exceeds or falls below the threshold voltage V_th to a time when the switches SWA and SWB are actually switched.

Therefore, the measured voltage v_A continues to ramp up from the time when the threshold voltage V_th is exceeded until the delay period D_cmp elapses. When the capacitors C_A and C_B are connected in parallel at a time when the switches SWA and SWB are switched to SW1 after the delay period D_cmp has elapsed, the charges accumulated in the capacitor C_A flow into the capacitor C_B, and as a result, the measured voltage v_A drops rapidly.

The same is true if the measured voltage v_A of the capacitor C_A falls below the threshold voltage V_th. That is, even if the switches SWA and SWB are switched to SW1 and the measured voltage v_A drops abruptly and falls below the threshold voltage V_th, the parallel connection state of the capacitors C_A and C_B is retained until the delay period D_cmp elapses from that time. When the switches SWA and SWB are switched to SW2 after the delay period D_cmp has elapsed, the capacitor C_B is separated from the capacitor C_A and connected in parallel with the resistive load R.

The delay period D_cmp is a total sum of delays such as a delay caused by: an interconnect delay; gate delays of the MOS transistors constituting the switches SWA and SWB; and a delay at the comparator CMP. In particular, the delay caused by the comparator CMPs is determined from its output-capacitance C_out, its transconductance g_m and so on. Thus, by designing them appropriately, the delay period D_cmp is optimized.

When a signal is inputted from an outside, the neuron element 10A configured as described above operates as follows.

FIG. 4 is a timing chart showing temporal changes of respective signals in the neuron element 10A. Here, at its initial state, charges accumulated in the capacitors C_A and C_B shown in FIG. 3 are zero. Therefore, the measured voltage v_A of the capacitor C_A is 0. Since the measured voltage v_A is lower than the threshold voltage V_th, both the switches SWA and SWB are switched to SW2.

In FIG. 4, at a time t = 0, a constant current i_IN = I is inputted to the input terminal IN. In practice, however, a current input to the input terminal IN is not constant.

Incidentally, a normal operating condition of the neuron element 10A refers to a condition under which the neuron element 10A operates normally even in the most stringent state, specifically, the state in which an allowable maximum instantaneous input current I_max is continuously inputted as the constant current. As long as this normal operating condition is satisfied, it is ensured that the neuron element 10A operates normally even when a current that changes arbitrarily but does not exceed I_max is inputted. The normal operating condition will be described in detail later.

When the constant current i_IN = I is inputted to the input terminal IN, the charges are accumulated in the capacitor C_A, and the measured voltage v_A of the capacitor C_A gradually ramps up.

At a time t = T_tS1, when the measured voltage v_A exceeds the threshold voltage V_th, the comparator CMP supplies the selecting signal SW1 to the switches SWA and SWB. Consequently, the switches SWA and SWB start their switching operations from SW2 to SW1.

However, as described above, the switching operations from SW2 to SW1 for the switches SWA and SWB are not executed immediately. A time t at which SW2 is actually switched to SW1 is t = T_S1 after the delay period D_cmp from the time t = T_tS1.

That is, in a period from the time t = T_tS1 to the time t = T_S1, the charges are continuously accumulated in the capacitor C_A. As a result, the measured voltage v_A ramps up continuously even if it exceeds the threshold voltage V_th, and reaches a peak voltage V_pk just before the time t = T_S1.

At the time t = T_S1, when the switches SWA and SWB are switched from SW2 to SW1, the capacitors C_A and C_B are connected in parallel, and the charges accumulated in the capacitor C_A flow into the capacitor C_B. Therefore, the measured voltage v_A of the capacitor C_A drops exponentially. Also, the voltage v_B generated at the capacitor C_B ramps up exponentially.

Here, a parasitic resistance r between the capacitors C_A and C_B is so small that a time constant τ₁ between the capacitors C_A and C_B is negligible compared with the delay period D_cmp. As a consequence, at a moment when the time t = T_S1 is exceeded, a part of charges accumulated in the capacitors C_A flow into the capacitor C_B, and the voltage v_A and v_B respectively generated at the capacitors C_A and C_B have become equal to V_sh.

At this time, the measured voltage v_A (= v_B = V_sh) is less than the threshold voltage V_th, and so the comparator CMP supplies the selecting signal SW2 to the switches SWA and SWB. The switches SWA and SWB then start the switching operation from SW1 to SW2 at a time t = T_tS2.

Here, as described above, even when the measured voltage v_A (= v_B = V_sh) becomes lower than the threshold voltage V_th, the switching operations of the switches SWA and SWB from SW1 to SW2 are not executed immediately. A time t at which SW1 is actually switched to SW2 is t = T_S2 after the delay period D_cmp from the time t = T_tS2.

That is, in a period from the time t = T_tS2 to the time t = T_S2, the charges are continuously accumulated in both of the capacitors C_A and C_B. Consequently, the measured voltage v_A (= v_B) ramps up and reaches the voltage V_bk just before the time t = T_S2.

Note that the measured voltage v_A’'s ramp-up speed (gradient) at this time is I/(C_A + C_B) and is smaller than a ramp-up speed of the measured voltage v_A from the time t = 0 to the time t ═ T_tS1 (═ I/C_A). The reason for this is that the charges flowed by the input constant current i_IN = I are accumulated in both of the parallel connected capacitors C_A and C_B.

When the switches SWA and SWB are switched from SW1 to SW2, the capacitor C_B is connected in parallel with the resistive load R. As a result, the charges accumulated in the capacitor C_B are discharged through the resistive load R. Then, the voltage v_B generated at the capacitor C_B rapidly drops according to an exponential function having a time constant τ₂ (= R ▪ C_B). The voltage v_B is supplied to an external load as a voltage pulse signal through the output terminal OUT.

On the other hand, the capacitor C_A is separated from the capacitor C_B. Consequently, the capacitor C_A is charged by the input constant current i_IN = I, and the measured voltage v_A of the capacitor C_A ramps up. A ramp-up speed of the measured voltage v_A at this time is the same as the ramp-up speed of the measured voltage v_A from the time t = 0 to the time t = T_tS1 (═ I/C_A).

At a second time t = T_tS1, the measured voltage v_A again exceeds the threshold voltage V_th, and the switches SWA and SWB start the switching operations from SW2 to SW1. At a second time t = T_s1 after the delay period D_cmp has elapsed from the second time t = T_tS1, the switches SWA and SWB are switched from SW2 to SW1. As a consequence, the charges accumulated in the capacitor C_A flow into the capacitor C_B.

As described above, when the constant current i_IN = I is inputted to the input terminal IN, the neuron element 10A in the first embodiment outputs a pulsed voltage having a peak voltage of V_bk and a period of 2 ▪ D_cmp + (T_tS1 —T_S2) via the output terminal OUT.

Next, two normal operating conditions of the neuron element 10A when the constant current i_IN = I is inputted to the input terminal IN at the time t = 0 will be described.

The first normal operating condition is that the measured voltage v_A at the time t = T_S1 (= V_pk) does not exceed the source voltage V_DD. That is, the first normal operating condition is given by Equation (1).

${\text{V}}_{\text{DD}}>{\text{V}}_{\text{pk}}={\text{V}}_{\text{th}}+\left(\text{I}·{\text{D}}_{\text{Cmp}}\right)/{\text{C}}_{\text{A}}$

Here, the second term on the right-hand side of Equation (1) represents a ramping up of the measured voltage v_A due to the charges flowing into the capacitor C_A from the time t = T_tS1 to the time t = T_S1.

As for the second normal operating condition, it is assumed that the delay period D_cmp becomes sufficiently long. At this time, it is assumed that the threshold voltage V_th is sufficiently lowered so that the peak voltage V_pk of the capacitor C_A does not exceed the source voltage V_DD.

Consequently, however, in a situation in which the switches SWAs and SWBs are switched to SW1, or the capacitors C_A and C_B are connected in parallel, it could happen that the measured voltage v_A (= v_B) exceeds the threshold voltage V_th.

FIG. 5 is a timing chart showing temporal changes of respective signals in the neuron element 10A when the delay period D_cmp becomes longer.

In FIG. 5, even when the measured voltage v_A (= v_B) exceeds the threshold voltage V_th, a situation is shown in which the parallel connection state of the capacitors C_A and C_B is retained. Then, at the time t = T_S2, the capacitors C_A and C_B are separated. A value of the measured voltage v_A (= v_B) just before the separation is V_bk.

On the other hand, the comparator CMP outputs the selecting signal SW1 when the measured voltage v_A (= v_B) exceeds the threshold voltage V_th (at a time t = T_tS1), even when the parallel connection of the capacitors C_A and C_B is retained. Then, at a time t = T_S1, the switches SWA and SWB are switched to SW1.

Thus, at the switching from SW1 to SW2 and at that from SW2 to SW1 by the switches SWA and SWB, a period from a start of the SW switching operation to a time the SW switching is actually enabled overlaps. Specifically, in FIG. 5, a period from the time t = T_tS1 to the time t = T_S1 and a period from the time t = T_tS2 to the time t = T_S2 overlap each other from the time t = T_tS1 to the time t = T_S2.

Consequently, a duration when the switch SWA and SWB are actually switched to SW2 is shortened. Then, because a discharging time of the capacitor C_B is shortened, the charges are not fully discharged from the capacitor C_B and thus there is residual charges in the capacitor C_B.

Subsequently, when the switches SWA and SWB are actually switched to SW1 and the capacitors C_A and C_B are connected in parallel, the residual charges are shared by the capacitors C_A and C_B. Consequently, a voltage V_sh generated at the capacitors C_A and C_B ramps up by the residual charges.

Consequently, the time t = T_tS1 at which the measured voltage v_A (= v_B) exceeds the threshold voltage V_th is moved forward (becomes earlier). However, the time t = T_S2 does not change. This is because that the time t = T_S2 is determined by the time t = T_tS2, which is a time at which the measured voltage v_A falls below the threshold voltage V_th after an abrupt drop of the measured voltage v_A caused by the switching of the switches SWA and SWB from SW2 to SW1.

Thus, the overlap period described above becomes longer, and the period in which the switches SWA and SWB are actually switched to SW2 is further shortened. In this way, the residual charges in the capacitor C_B gradually increases, and eventually, even if the capacitors C_A and C_B are connected in parallel, the voltage V_sh generated at both capacitors does not fall below the threshold voltage V_th, and thus the neuron element 10A no more operates normally.

That is, the second normal operating condition is that the charges accumulated in the capacitor C_B are completely discharged while the switches SWA and SWB are actually switched to SW2.

Here, let a voltage generated at the capacitors C_A and C_B just before the parallel connection of the capacitors C_A and C_B is separated by the switching of SWA and SWB from SW1 to SW2 at k-th time be V_bk(k). The second normal operating condition is as shown in Equation (2).

${\text{D}}_{\text{cmp}}-\text{-}\left(\left({\text{C}}_{\text{A}}{\text{+C}}_{\text{B}}\right)\cdot \left({\text{V}}_{\text{bk}}\left(\text{k}\right)-\text{-}{\text{V}}_{\text{th}}\right)\right)\text{/}\text{I}>{\tau }_2$

Here, the second term on the left-hand side of Equation (2) corresponds to a duration from a time at which the measured voltage v_A (= v_B) exceeds the threshold voltage V_th to a time at which the measured voltage reaches V_bk(k), that is, it is a period from the time t = T_tS1 to the time t = T_S2 shown in FIG. 5. Furthermore, τ₂ is a time constant of a discharging circuit composed of the capacitor C_B and the resistive load R as described above, and is expressed by Equation (3).

${\tau }_2=\text{R}\cdot {\text{C}}_{\text{B}}$

Incidentally, the voltage V_bk(k) converges to a predetermined value while oscillating each time as a number of switching times of SWA and SWB increases, with a first voltage V_bk(1) shown in FIG. 5 as the largest value. On the other hand, a total amount of charges accumulated at a time of the measured voltage v_A(= v_B) = V_bk(1) is a sum of charges C_A ▪ V_th accumulated in the capacitor C_A at a time of the first t = T_tS1 and charges 2 ▪ I ▪ D_cmp flowed during a subsequent duration of 2 ▪ D_cmp. Therefore, a relation between V_bk(k) and V_bk(1) is expressed by Equation (4).

${\text{V}}_{\text{bk}}\left(\text{k}\right)\leqq {\text{V}}_{\text{bk}}\left(1\right)=\left({\text{C}}_{\text{A}}\cdot {\text{V}}_{\text{th}}+2\cdot \text{I}\cdot {\text{D}}_{\text{cmp}}\right)/\left({\text{C}}_{\text{A}}\text{+}{\text{C}}_{\text{B}}\right)$

On the other hand, the left-hand side of Equation (2) is minimized when V_bk(k) is a maximum value, or V_bk(1). Thus, when V_bk(k) in Equation (2) is set to V_bk(1), Equations (3) and (4) are substituted into Equation (2) and arranged, Equation (5) is obtained.

$\left({\text{C}}_{\text{B}}\cdot {\text{V}}_{\text{th}}\right)/\text{I}-\text{-}{\text{D}}_{\text{cmp}}>\text{R}\cdot {\text{C}}_{\text{B}}\text{I}.$

Suppose the most severe condition where a maximum instantaneous input current I_max is continuously inputted as the constant current. Then, Equations (1) and (5) provide the normal operating conditions among the device parameters for a stable operation of the neuron element 10A. Therefore, when Equations (1) and (5) are arranged in terms of V_th with I = I_max, Equation (6) is obtained.

${\text{I}}_{\max }\cdot \left({\text{R+D}}_{\text{cmp}}{\text{/C}}_{\text{B}}\right)<{\text{V}}_{\text{th}}<{\text{V}}_{\text{DD}}-\text{-}\left({\text{I}}_{\text{max}}\cdot {\text{D}}_{\text{cmp}}\right)/{\text{C}}_{\text{I}}$

From Equation (6), an allowable setting range of the threshold voltage V_th is obtained. Furthermore, if an attention is paid to the leftmost side and the rightmost side of Equation (6), a constraint condition to be satisfied by the delay period D_cmp is expressed by Equation (7).

${\text{D}}_{\text{cmp}}<\left({\text{C}}_{\text{A}}\cdot {\text{C}}_{\text{B}}\right)/\left({\text{C}}_{\text{A}}+{\text{C}}_{\text{B}}\right)\left({\text{V}}_{\text{DD}}/{\text{I}}_{\text{m-x}}\text{I}\text{R}\right)$

Equations (6) and (7) provide the normal operating conditions that the threshold voltage V_th and the delay period D_cmp should respectively satisfy in order for the neuron element 10A to operate normally, assuming the maximum instantaneous input current to be I_max.

Next, a second embodiment will be described. The same portions as those of the first embodiment are denoted by the same reference numerals, and repetitive descriptions thereof are omitted.

FIG. 6 is a circuit diagram showing a configuration of a neuron element 10B according to the second embodiment. The neuron element 10B is an exemplary configuration of the neuron element 10 shown in FIG. 1.

The neuron element 10B includes a capacitor C_A, a comparator CMP, a pulse signal generating unit PG, and a voltage controlled current source VCCS. One terminal of the capacitor C_A is connected to an input terminal IN, and the other terminal thereof is grounded.

The comparator CMP measures an amount of charges accumulated in the capacitor C_A as a voltage of the input terminal IN (= a voltage generated at the capacitor C_A), and determines whether the voltage being measured (the measured voltage v_A) exceeds a predetermined threshold voltage V_th. The comparator CMP supplies a high-level signal (H-signal) to the pulse signal generating unit PG when the measured voltage v_A exceeds the threshold voltage V_th, and supplies a low-level signal (L-signal) to the pulse signal generating unit PG when the measured voltage v_A does not exceed the threshold voltage V_th.

The pulse signal generating unit PG generates a pulse signal that is a rectangular wave with its duty ratio α (0 < α < 1) when the H-signal is supplied from the comparator CMP. On the other hand, the pulse signal generating unit PG does not generate the pulse signal when the L-signal is supplied from the comparator CMP. The pulse signal generated by the pulse signal generating unit PG is outputted to an outside via an output terminal OUT, and is also supplied to the voltage controlled current source VCCS.

The voltage controlled current source VCCS is connected in parallel with the capacitor C_A. Specifically, a current input terminal of the voltage controlled current source VCCS is connected to the input terminal IN. Its current output terminal is grounded. Then, the voltage controlled current source VCCS controls the current flowing from the current input terminal to the current output terminal based on the pulse signal generated by the pulse signal generating unit PG.

Specifically, the voltage controlled current source VCCS forcibly causes the constant current i_s = I_F to flow in response to a high-level of the pulse signal generated by the pulse signal generating unit PG, and interrupts the current in response to a low-level of the pulse signal.

When the voltage controlled current source VCCS forces a constant current is = I_F and the constant current I_F is greater than the current coming from the input terminal IN, charges are discharged from the capacitor C_A connected to the input terminal IN and the measured voltage v_A of the capacitor C_A drops. When the voltage controlled current source VCCS interrupts the constant current I_F, all the current from the input terminal IN flows into the capacitor C_A, and the measured voltage v_A of the capacitor C_A ramps up.

The other terminal of the capacitor C_A and the current output terminal of the voltage controlled current source VCCS may be equipotential, and are not limited to be grounded as shown in the present embodiment.

Here, the pulse signal generating unit PG does not initiate or halt the pulse signal generation operation as soon as a ramping up measured voltage v_A exceeds a threshold voltage V_th or a dropping measured voltage v_A falls below the threshold voltage V_th. That is, there is a predetermined delay D_cmp from a time when the ramping up or dropping of the measured voltage v_A exceeds or falls below the threshold voltage V_th to a time when the pulse signal generating unit PG actually starts or stops the operation.

Therefore, the measured voltage v_A continues to ramp up from the time when the threshold voltage V_th is exceeded until the delay period D_cmp elapses, and drops when the operation of the pulse signal generating unit PG starts after the delay period D_cmp elapses and the voltage controlled current source VCCS discharges charges from the capacitor C_A in conjunction therewith.

The same is true when the measured voltage v_A of the capacitor C_A falls below the threshold voltage V_th. That is, the measured voltage v_A continues to drop from a time when it falls below the threshold voltage V_th until the delay period D_cmp elapses. The operation of the pulse signal generating unit PG then stops after the delay period D_cmp elapses, and the measured voltage v_A ramps up when the voltage controlled current source VCCS stops discharging the charges from the capacitor C_A in conjunction therewith.

The delay period D_cmp is a total sum of an interconnect delay, an operation delay of the pulse signal generating unit PG, a delay caused by the comparator CMP, and the like. In particular, the delay caused by the comparator CMP is determined from its output-capacitance C_out, its transconductance g_m and so on. Thus, by designing them appropriately, the delay period D_cmp is optimized.

When a signal is inputted from an outside, the neuron element 10B configured as described above operates as follows.

FIG. 7 is a timing chart showing temporal changes of respective signals in the neuron element 10B. In its initialization, charges accumulated in the capacitor C_A shown in FIG. 6 are zero. Therefore, the measured voltage v_A of the capacitor C_A is 0. Since the measured voltage v_A is lower than the threshold voltage V_th, the comparator CMP outputs the L-signal, and the operation of the pulse signal generating unit PG is stopped.

In FIG. 7, at a time t = 0, a constant current i_IN = I is inputted to the input terminal IN. In practice, however, a current input to the input terminal IN is not constant. Details of normal operating conditions of the neuron element 10B in the present embodiment will be described later.

When the constant current i_IN = I is inputted to the input terminal IN, the charges are accumulated in the capacitor C_A, and the measured voltage v_A of the capacitor C_A gradually ramps up.

At a time t = T_tOn, when the measured voltage v_A exceeds the threshold voltage V_th, the output signal of the comparator CMP switches from the L-signal to the H-signal. Consequently, the pulse signal generating unit PG initiates the pulse signal generation operation. When the pulse signal is supplied from the pulse signal generating unit PG, the voltage controlled current source VCCS forcibly causes the constant current i_s = I_F to flow in response to the high-level signal (see a case when a time t = T_On described later). As a consequence, the charges are discharged from the capacitor C_A and the measured voltage v_A drops.

However, as described above, the pulse signal generating unit PG does not operate immediately even when the measured voltage v_A exceeds the threshold voltage V_th. A time t at which the pulse signal generating unit PG starts the pulse signal generation operation is t = Ton after the delay period D_cmp has elapsed from the time t = T_tOn.

That is, in a period from the time t = T_tOn to the time t = T_On, the charges are continuously accumulated in the capacitor C_A. As a result, the measured voltage v_A ramps up as it is even if it exceeds the voltage threshold voltage V_th, and reaches a peak voltage V_pk just before the time t = T_On.

At the time t = T_On, when the pulse signal generating unit PG starts the pulse signal generation operation, the voltage controlled current source VCCS forcibly causes the constant current i_s = I_F to flow in response to the high-level of the pulse signal supplied from the pulse signal generating unit PG. As a consequence, the charges are discharged from the capacitor C_A and the measured voltage v_A drops.

On the other hand, the voltage controlled current source VCCS cuts off the current flow in response to the low-level of the pulse signal. At this time, the charges are accumulated in the capacitor C_A and the measured voltage v_A ramps up again. The ramp-up speed (slope) of the measured voltage v_A at this time is a same as the ramp-up speed of the measured voltage v_A from the time t = 0 to the time t = T_tOn (═ I/C_A).

As a result, during the generation of the pulse signal by the pulse signal generating unit PG, the measured voltage v_A drops on the average while periodically repeating the dropping and the ramping-up.

At a time t = T_tOff, the measured voltage v_A falls below the threshold voltage V_th. At this time, the signal outputted from the comparator CMP and supplied to the pulse signal generating unit PG is switched from the H-signal to the L-signal. Then, the pulse signal generating unit PG stops the pulse signal generation operation. Consequently, the voltage controlled current source VCCS cuts off the current flow.

However, as described above, the pulse signal generating unit PG does not stop as soon as the measured voltage v_A falls below the threshold voltage V_th. A time t at which the pulse signal generating unit PG stops the pulse signal generation operation is t = Toff after the delay period D_cmp has elapsed from the time t = T_tOff. That is, in a period from the time t = T_tOff to the time t = T_Off, the charges are continuously discharged from the capacitor C_A. Consequently, the measured voltage v_A drops as is, even if it falls below the threshold voltage V_th.

Note that, because of its nature, even if the L-signal is supplied from the comparator CMP during the pulse signal generation operation, the pulse signal generating unit PG generates a complete pulse signal for one period and then stops, without immediately interrupting the generation operation.

At the time t = Toff, the operation of the pulse signal generating unit PG stops, the voltage controlled current source VCCS stops and interrupts the current flow is. Then, a constant current i_IN = I from the input terminal IN accumulates the charges in the capacitor C_A, and the measured voltage v_A of the capacitor C_A ramps up again.

Thereafter, at a second time t = T_tOn, the measured voltage v_A again exceeds the threshold voltage V_th. At a second time t = T_On after the second time t = T_tOn and the delay period D_cmp, the pulse signal generating unit PG and the voltage controlled current source VCCS resume their operations. Then, the above-described operations are repeated.

As described above, when the constant current i_IN = I is inputted to the input terminal IN, the neuron element 10B according to the second embodiment intermittently outputs a pulse signal having a predetermined voltage, a period D_p and a rectangular wave with a duty ratio α, via the output terminal OUT.

Next, two normal operating conditions of the neuron element 10B will be described.

In a first normal operating condition, consider a case where a constant current i_IN = I is inputted to the input terminal IN at the time t = 0. Here, a maximum instantaneous input current I = I_max is used instead of the constant current i_IN = I to derive the normal operating condition.

As described above, in order for the measured voltage v_A of the capacitor C_A to continue to drop on average, the amount of charges discharged from the capacitor C_A by the voltage controlled current source VCCS needs to be larger than the amount of charges flowing into the capacitor C_A by the constant current i_IN = I_max.

Consider a pulse signal for one cycle generated by the pulse signal generating unit PG immediately after the time t = T_On.

In the one cyclic period of the pulse signal, a total amount of charges flowed into the capacitor C_A is given by D_p ▪ I_max because one cyclic period of the pulse signal is D_p. In the one cyclic period, a total amount of charges discharged from the capacitor C_A by the voltage controlled current source VCCS is given by α ▪ D_p ▪ I_F because an operation time of the voltage controlled current source VCCS is α ▪ D_p. Since the latter is larger than the former, the first normal operating condition is expressed by Equation (8).

${\text{I}}_{\text{F}}>1/\text{α}\cdot {\text{I}}_{\max }$

On the other hand, if the voltage controlled current source VCCS forcibly discharges the charges from the capacitor C_A even when the amount of charges accumulated in the capacitor C_A becomes zero, it is assumed that the voltage controlled current source VCCS does not operate normally. Therefore, the second normal operating condition is a condition for avoiding such the unnormal operation.

FIG. 8 is a timing chart showing temporal changes of respective signals in the neuron element 10B when it is most likely to cause a condition that does not operate normally.

In FIG. 8, at the time t = 0, the constant current i_IN = I_max is inputted to the input terminal IN. Thus, the charges are accumulated in the capacitor C_A and the measured voltage v_A of the capacitor C_A ramps up at a rate I_max/C_A.

At the time t = T_tOn, suppose that the constant current i_IN = I_max stops immediately after the measured voltage v_A exceeds the threshold voltage V_th. Consequently, the amount of accumulated charges in the capacitor C_A remains constant. Thereafter, when the pulse signal generating unit PG starts the pulse signal generation operation and the voltage controlled current source VCCS responds to the pulse signal generation operation by causing the constant current i_s = I_F to flow, the charges are forcibly discharged from the capacitor C_A.

However, the voltage controlled current source VCCS actually starts the operation at which a time t becomes t = T_On after the delay period D_cmp elapses from the time t = T_tOn, and after the pulse signal generating unit PG starts the pulse signal generation operation. Here, just before the time t = Ton, the amount of charges C_A ▪ V_th is accumulated in the capacitor C_A.

At the time t = T_On, the voltage controlled current source VCCS causes the constant current i_s = I_F to flow in response to the high-level of the pulse signal supplied from the pulse signal generating unit PG. Consequently, in one cyclic period of the pulse signal, the amount of charges α ▪ D _p ▪ I_F is forcibly discharged from the capacitor C_A, and the measured voltage v_A drops by α ▪ D p ▪ I_F/C_A. The voltage controlled current source VCCS also cuts off the current flow in response to the low-level of the pulse signal. At this time, no charge is discharged from the capacitor C_A, and the measured voltage v_A does not change.

As described above, when the pulse signal is at the high-level, the measured voltage v_A drops, and when the pulse signal is at the low-level, the measured voltage v_A does not drop and remains constant. Therefore, as shown in FIG. 8, after the time t = T_on, the measured voltage v_A periodically and intermittently drops.

On the other hand, as described above, immediately before the time t = T_On, the amount of charges C_A ▪ V_th is accumulated in the capacitor C_A. Then, since the charges are discharged from the capacitor C_A by the pulse signal immediately after the time t = T_On, the measured voltage v_A immediately falls below the threshold voltage V_th and the signal provided from the comparator CMP to the pulse signal generating unit PG switches from the H-signal to the L-signal. Let a time at this time be t = T_toff.

However, as described above, the pulse signal generating operation of the pulse signal generating unit PG is not stopped immediately even if the measured voltage v_A falls below the threshold voltage V_th. A time t at which the pulse signal generating unit PG stops the operation is t = T_Off after the delay period D_cmp has elapsed from the time t = T_tOff. That is, from the time t = T_tOff to the time t = T_Off, charges are periodically and intermittently discharged from the capacitor C_A by the voltage controlled current source VCCS. In other words, the drop of the measured voltage v_A described above continues until the time t = T_Off.

Suppose that the pulse signal supplied to the voltage controlled current source VCCS goes through n periods and the measured voltage v_A drops to a near-zero value. At this time, a total amount of charges emitted from the capacitor C_A is n ▪ α ▪ D_p ▪ I_F. On the other hand, as mentioned previously, an amount of charges accumulated in the capacitor C_A before the pulse signal generating unit PG initiates the pulse signal generation operation is C_A ▪ V_th.

The total amount of charge discharged from the capacitor C_A must not exceed the amount of charges accumulated in the capacitor C_A. That is, it is necessary to satisfy Equation (9).

$\left(\text{n}\cdot \text{α}\cdot {\text{D}}_{\text{p}}\cdot {\text{I}}_{\text{F}}\right)/{\text{C}}_{\text{A}}<{\text{V}}_{\text{th}}<\text{}{\text{V}}_{\text{DD}}$

Here, an inequality on the right-hand side of Equation (9) indicates that an upper limit of the threshold voltage V_th is given by the source voltage V_DD.

Next, a relation between the discharging period and the delay period D_cmp by a continuous operation of the voltage controlled current source VCCS will be described.

In FIG. 8, if the delay period D_cmp is shorter than the discharging period, a position of the time t = Toff moves leftward, and as a result, the charges accumulated in the capacitor C_A does not become zero. Thus, the neuron element 10B operates normally.

Conversely, if the delay period D_cmp is longer than the discharging period, the charges are discharged beyond the amount of charges accumulated in the capacitor C_A. Therefore, the neuron element 10B does not operate normally. Thus, the condition to be satisfied by the delay period D_cmp is given by Equation (10) when the cyclic period D_p of the pulse signal supplied to the voltage controlled current source VCCS is sufficiently smaller than the delay period D_cmp.

${\text{D}}_{\text{cmp}}\leqq \text{n}\cdot {\text{D}}_{\text{p}}$

In Equation (10), n is given by Equation (11) where the largest integer less than or equal to x for a real number x is represented by [x] using the Gaussian symbol [].

$\text{n}=\left[\left({\text{C}}_{\text{A}}\cdot {\text{V}}_{\text{th}}\right)/\left(\text{α}\cdot {\text{D}}_{\text{p}}\cdot {\text{I}}_{\text{F}}\right)\right]$

Substituting Equation (11) into Equation (10), and further considering the inequality on the right-hand side of Equations (8) and (9), Equation (10) becomes Equation (12).

${\text{D}}_{\text{cmp}}<{\text{C}}_{\text{A}}\cdot {\text{V}}_{\text{DD}}/{\text{I}}_{\max }$

In addition, when Equation (10) is substituted into the inequality on the left-hand side of Equation (9), and Equation (8) is further considered, Equation (9) becomes Equation (13).

$\left({\text{D}}_{\text{cmp}}\cdot {\text{I}}_{\max }\right)/{\text{C}}_{\text{A}}<{\text{V}}_{\text{th}}<{\text{V}}_{\text{DD}}$

That is to say, when a maximum instantaneous input current is assumed to be I_max, Equations (12) and (13) are the normal operating conditions both the delay period D_cmp and the threshold voltage V_th need to respectively satisfy in order to normally operate the neuron element 10B.

The pulse signal generating unit PG may be designed so as to generate a pulse signal having: a duty ratio α with which the constant current i_S = I_F controlled by the voltage controlled current source VCCS is set so as to satisfy Equation (8); and a cyclic period D_p which is set so as to satisfy Equation (10) using a desired n.

Incidentally, in the first and second embodiments, a case has been described in which a reference potential of the capacitor C_A is set to 0V, a current flows into the capacitor C_A, and the measured voltage v_A generated at the capacitor C_A ramps up toward the source voltage V_DD. However, the present invention is not limited to these embodiments.

For example, as will be described later, while the reference potential of the capacitor C_A remains at 0V, the current may flow out of the capacitor C_A, and the measured voltage v_A generated at the capacitor C_A may fall toward a negative source voltage -V_DD.

Furthermore, the source voltage V_DD may be used as the reference potential of the capacitor C_A, and as a result of the current flowing into the capacitor C_A, the measured voltage v_A generated at the capacitor C_A may fall toward 0V.

Next, a third embodiment will be described. It should be noted that the same reference numerals are assigned to the same portions as those in the above-described embodiments, and repetitive descriptions thereof are omitted.

FIG. 9 is a circuit diagram showing an exemplary configuration of a neuron element 10C according to the third embodiment. The neuron element 10C has a circuit corresponding to the neuron element 10A shown in FIG. 3 and a circuit obtained by inverting the circuit corresponding to the neuron element 10A. That is, the neuron element 10C includes capacitors C_A, C_B1 and C_B2, a comparator CMP, switches SWA1, SWB1, SWA2 and SWB2, and resistive loads R₁ and R₂.

One terminal of the capacitor C_A is connected to an input terminal IN1, and the other terminal is connected to an input terminal IN2. That is, the capacity C_A is not grounded. Therefore, when a current flows from the input terminal IN1 to the capacitor C_A, a current having a magnitude equal to that flowing into the capacitor C_A flows out from the input terminal IN2. Therefore, a voltage v_A based on a potential of the input terminal IN2 is generated at both terminals of the capacity C_A. In this way, the capacitor C_A is integrated into one in order to suppress the number of the components.

Each one terminal of the capacitors C_B1 and C_B2 is connected to a terminal SW3 of the switches SWA1 and SWA2, respectively. The other terminal of the capacitors C_B1 and C_B2 is connected to the terminals SW3 of the switches SWB1 and SWB2, respectively.

Each terminal SW1 of the switches SWA1 and SWA2 is connected to the input terminal IN1 and IN2, respectively. Each terminal SW1 of the switches SWB1 and SWB2 is connected to each other.

A terminal SW2 of the switch SWA1 is connected to the resistive load R₁ and an output terminal OUT1. A terminal SW2 of the switch SWA2 is connected to the resistive load R₂ and an output terminal OUT2. Each terminal SW2 of the switches SWB1 and SWB2 is grounded.

One terminal of the resistive load R₁ is connected to the terminal SW2 of the switch SWA1 and the output terminal OUT1. The other terminal of the resistive load R₁ is grounded. One terminal of the resistive load R₂ is connected to the terminal SW2 of the switch SWA2 and the output terminal OUT2. The other terminal of the resistive load R₂ is grounded.

An inverted input terminal of the comparator CMP is connected to a positive terminal of a constant voltage source V_th which supplies a predetermined reference voltage V_th. A noninverted input terminal of the comparator CMP is connected to the input terminal IN1. A negative terminal of the constant voltage source V_th is connected to the input terminal IN2. Then, the constant voltage source V_th supplies, to the inverted input terminal of the comparator CMP, a reference voltage V_th based on a potential of the input terminal IN2 as a threshold voltage V_th.

The comparator CMP measures an amount of charges accumulated in the capacitor C_A as a voltage of the input terminal IN1 based on a potential of the input terminal IN2 (a voltage of the capacitor C_A), and determines whether or not the voltage being measured (the measured voltage v_A) exceeds the predetermined threshold voltage V_th. The comparator CMP supplies a selecting signal SW1 to the switches SWA1, SWB1, SWA2 and SWB2 when the measured voltage v_A exceeds the threshold voltage V_th, and supplies a selecting signal SW2 to the switches SWA1, SWB1, SWA2 and SWB2 when the measured voltage v_A does not exceed the threshold voltage V_th.

The switches SWA1, SWB1, SWA2 and SWB2 switch as follows when the selecting signal is supplied from the comparator CMP. That is, the switches SWA1, SWB1, SWA2 and SWB2, when the selecting signal SW1 is supplied, conduct between the terminals SW1 and SW3 (switched to SW1). On the other hand, the switches SWA1, SWB1, SWA2 and SWB2, when the selecting signal SW2 is supplied, conduct between the terminals SW2 and SW3 (switched to SW2).

Therefore, when the selecting signal SW1 is supplied to the switches SWA1, SWB1, SWA2 and SWB2, the capacitor C_A and a serially connected capacitors C_B1 and C_B2 are connected in parallel. On the other hand, when the selecting signal SW2 is supplied to the switches SWA1, SWB1, SWA2 and SWB2, the capacitors C_B1 and C_B2 are respectively connected to the resistive load R₁ and R₂ in parallel. The charges accumulated in the capacitors C_B1 and C_B2 are discharged through the resistive load R₁ and R₂, respectively.

The terminals SW2 of the switches SWB1 and SWB2, and the other terminals of the resistive loads R₁ and R₂ may be equipotential, and are not limited to be grounded as shown in the present embodiment.

When signals are inputted from an outside, the neuron element 10C configured as described above operates as follows. Here, at its initial state, charges accumulated in the capacitors C_A, C_B1 and C_B2 are zero. Therefore, the measured voltage v_A of the capacitor C_A is 0. Since the measured voltage v_A is below the threshold voltage V_th, all switches SWA1, SWB1, SWA2 and SWB2 have switched SW2. Thus, the capacitors C_B1 and C_B2 are separated from the capacitor C_A.

Now, when a constant current i₁ = I is inputted to the input terminal IN1, a constant current i₂ = I equivalent to the constant current i₁ flows out from the input terminal IN2 and the charges are accumulated in the capacitors C_A, and the measured voltage v_A at the capacitor C_A gradually ramps up.

If the measured voltage v_A exceeds the threshold voltage V_th, the comparator CMP supplies the selecting signal SW1 to the switches SWA1, SWB1, SWA2 and SWB2. Consequently, the switches SWA1, SWB1, SWA2 and SWB2 initiate their switching operations from SW2 to SW1.

However, as described in the first embodiment, the switching operations from SW2 to SW1 for the switches SWA1, SWB1, SWA2 and SWB2 are not executed immediately. A time t at which SW2 is switched to SW1 is t = T_S1 which is a time after a delay period D_cmp elapses from a time t = T_tS1 at which the measured voltage v_A exceeds the threshold voltage V_th.

That is, in a period from the time t = T_tS1 to the time t = T_S1, the charges are continuously accumulated in the capacitor C_A. As a result, the measured voltage v_A ramps up as it is, even if it exceeds the threshold voltage V_th, and peaks just before the time t = T_S1.

At the time t = T_S1, when the switches SWA1, SWB1, SWA2 and SWB2 are switched from SW2 to SW1, the capacitors C_A and the serially connected capacitors C_B1 and C_B2 are connected in parallel. Then, the charges accumulated in the capacitor C_A flow into the serially connected capacitors C_B1 and C_B2. Therefore, the measured voltage v_A of the capacitor C_A drops exponentially. Voltages v_B1 and v_B2 generated at the respective capacitors C_B1 and C_B2 ramp up exponentially.

However, suppose parasitic resistances r among the capacitors C_A, C_B1 and C_B2 be small enough, and a time constant τ₁ among the capacitors C_A, C_B1 and C_B2 be negligible compared with the delay period D_cmp. As a result, at an instant when a time t exceeds T_S1, a part of the charges accumulated in the capacitor C_A flows into the capacitors C_B1 and C_B2. Then, the voltage v_A generated at the capacitor C_A and, a voltage generated at the serially connected capacitors C_B1 and C_B2 (= v_B1 + v_B2) become equal.

At this time, the measured voltage v_A (= v_B1 + v_B2) is lower than the threshold voltage V_th. For this reason, the comparator CMP provides the selecting signal SW2 to the switches SWA1, SWB1, SWA2 and SWB2. The switches SWA1, SWB1, SWA2 and SWB2 start their switching operations from SW1 to SW2.

When the delay period D_cmp elapses from a time when the measured voltage v_A falls below the threshold voltage V_th, the switches SWA1, SWB1, SWA2 and SWB2 are switched from SW1 to SW2. Thus, the serially connected capacitors C_B1 and C_B2 are separated from the capacitor C_A. The charges accumulated in the capacitors C_B1 and C_B2 are then discharged through the respective resistive loads R₁ and R₂.

The voltages v_B1 and v_B2 generated at the respective capacitors C_B1 and C_B2 drop abruptly according to exponential functions having respective time constants τ₂₁ (= R₁ · C_B1) and τ₂₂ (= R₂ · C_B2). The voltages v_B1 and v_B2 are supplied as pulse signals to external loads via the output terminals OUT1 and OUT2, respectively.

Here, let C_A, C_B1 and C_B2 be respective capacitances of the capacitors C_A, C_B1 and C_B2, and R₁ and R₂ be respective resistances of the resistive loads R₁ and R₂. At this time, Equations (14) and (15) are respectively satisfied between those device parameters and, the capacitance C_B of the capacitor C_B and the resistance R of the resistive load R, used in the first embodiment.

${\text{C}}_{\text{B1}}={\text{C}}_{\text{B2}}=2\cdot {\text{C}}_{\text{B}}$

${\text{R}}_{\text{1}}={\text{R}}_{\text{2}}=\text{R/2}$

If Equations (14) and (15) are satisfied, an upper half of the circuit shown in FIG. 9 (the capacitors C_A and C_B1, the comparator CMP and the resistive load R₁) is equivalent to the circuit shown in FIG. 3 (the capacitors C_A and C_B, the comparator CMP and the resistive load R). Furthermore, a lower half of the circuit shown in FIG. 9 (the capacitors C_A and C_B2, the comparator CMP and the resistive load R₂) is equivalent to that inverted polarity of the circuit shown in FIG. 3. That is, a signal output from the output terminal OUT2 is equivalent to that obtained by inverting a polarity of a signal output from the output terminal OUT1.

Consequently, the neuron element 10C can cancel effects of noises applied equally to the upper and the lower circuits compared to the case in FIG. 3, and thus can generate a lower noise, higher precision pulse signal.

Note that the present embodiment is not limited to the one based on the neuron element 10A in the first embodiment, and can be applied to the one based on the neuron element 10B in the second embodiment. That is, the present embodiment can be applied to all of the devices having the same functions as those of the neuron element 10 in FIG. 1.

Next, a fourth embodiment will be described. It should be noted that the same reference numerals are assigned to the same portions as those in the above-described embodiments, and repetitive descriptions thereof are omitted.

FIG. 10 is a block diagram showing a functional configuration of an A/D converter 100 according to the fourth embodiment. The A/D converter 100 includes a modulated pulse signal generating unit 50, a pulse waveform shaping unit 60, and a digital filter 70. The modulated pulse signal generating unit 50 has a same configuration as the neuron element 10 shown in FIG. 1.

If a pulse signal received from the modulated pulse signal generating unit 50 is not square, the pulse waveform shaping unit 60 shapes it and outputs it as a rectangular wave. The digital filter 70 is, for example, a pulse counter composed of a T-type flip-flop multistage connection. The digital filter 70, for the rectangular wave received from the pulse waveform shaping unit 60, for example, performs processing such as counting a number of the rectangular waves being received within a predetermined range on a time axis and converting the number into an n-bit output digital signal.

FIG. 11 is a set of wave diagrams respectively showing an input analog signal, a modulated pulse signal and an output digital signal, of a conventional sigma-delta A/D converter (e.g., shown in FIG. 15).

In FIG. 11, an upper input analog signal is a sine wave voltage signal with a frequency of 1 Hz and an amplitude of 0.5 V (0.0 V∼1.0 V). A waveform diagram in a middle row shows a 1-bit modulated pulse signal obtained by performing a sigma-delta modulation at a sampling frequency of 500 Hz for this input analog signal. A lower waveform diagram is an output digital signal as a result of the A/D conversion of the input analog signal obtained by performing a predetermined digital processing for this 1-bit modulated pulse signal in a subsequent digital filter.

As shown in the waveform diagram in the middle row of FIG. 11, in the conventional sigma-delta A/D converter, a pulse density near a reference amplitude of the input analog signal (0.5V in FIG. 11) is high. Thus, in the conventional sigma-delta A/D converter, when the amplitude of the input signal is set to-1.0 V~1.0 V, for example, the pulse density is highest in the vicinity of 0 V that is the reference value.

In other words, in the conventional sigma-delta A/D converter, even when there is little input analog signal (when the input analog signal is near the reference value (zero)), the high density pulse signal occurs. Therefore, there is a problem that a power is consumed unnecessarily even though there is almost no input analog signal.

FIG. 12 is a set of wave diagrams respectively showing an input analog signal, a modulated pulse signal and an output digital signal, of the A/D converter 100 shown in FIG. 10.

In FIG. 12, an upper input analog signal is a sine wave current signal with a frequency of 1 Hz and an amplitude of 5 nA (0 nA to 10 nA). A waveform diagram in a middle row shows a 1-bit modulated pulse signal obtained by the modulated pulse signal generating unit 50 of the A/D converter 100 for this input analog signal. It should be noted that although the modulated pulse signal generating unit 50 originally performs an event-driven type operation which performs a processing passively only when there is an input signal, the present embodiment shows a result of performing a synchronous operation with a sampling frequency of 500 Hz in order to directly compare the present embodiment with the conventional sigma-delta A/D converter operating with the same sampling frequency. A lower waveform diagram is an output digital signal as a result of the AD-conversion of the input analog signal obtained by performing a predetermined digital processing for this 1-bit modulated pulse signal by the digital filter 70.

As shown in the waveform diagram in the middle row of FIG. 12, in the A/D converter 100, a pulse density when the input analog signal is near-zero is substantially reduced.

In other words, in the A/D converter 100, when there is almost no input analog signal, a power consumption can be suppressed by making a density of the pulse signal sparse depending on an amplitude of the input analog signal.

In the present embodiment, the A/D converter 100 is an example of an AD-converting of a sine wave current signal as the input analog signal, but it is not limited to such the example.

For example, as described above, the neuron element 10 shown in FIG. 1 originally functions as an event-driven type device that operates only when a signal is inputted. That is, because the neuron element 10, unlike a common synchronized processing, does not operate unnecessary circuitry when there is no input signal, it is possible to significantly suppress the power consumption.

Next, a fifth embodiment will be described. It should be noted that the same reference numerals are assigned to the same portions as those in the above-described embodiments, and repetitive descriptions thereof are omitted.

FIG. 13 is a diagram showing an exemplary configuration of a neural network 200 according to the fifth embodiment. The neural network 200 includes a plurality of the neuron element 10 configured in multiple stages. FIG. 14 is a diagram showing a relation between an output signal x_i of the neuron element 10 (i = 1, 2, 3) and a weighting coefficient w_i (i = 1, 2, 3).

An input signal of any neuron element 10 is a summation of a plurality of the neuron element 10 output signal x_i weighted by the weighting coefficient w_i. The input signal is accumulated in the charge accumulating unit 12 of the neuron element 10. Then, the neuron element 10 supplies the output signal to the neuron element 10 of a next stage through a processing shown in FIG. 2.

Note that the input signal of any neuron element 10 is not limited to the output signal of another neuron element 10, and may be a signal inputted externally (an external input signal). Also, the input signal of any neuron element 10 may be both the output signal of other neuron element 10 and the external input signal. Furthermore, the output signal and the external input signal may be a plurality or singular.

The output signal of any neuron element 10 is not limited to the input signal of another neuron element 10, and may be an output signal to an outside (an external output signal), or may be both the input signal to other neuron element 10 and the external output signal. Furthermore, the input signal and the external output signal may be a plurality or singular.

The neural network 200 configured as described above can realize a so-called deep learning by optimizing the weighting coefficient w_i based on teacher data prepared in advance for a desired purpose.

It should be noted that the present invention is not limited to the above-mentioned embodiments, and can be applied to those whose design is changed within a scope of the claims.

REFERENCE SIGNS LIST

• 10, 10A, 10B, 10C Neuron elements
• 12 Charge accumulating unit
• 13 Discharging unit
• 14 Pulse signal generating unit
• 16 Controlling unit
• 20 Signal processing unit
• 50 Modulated pulse signal generating unit
• 70 Digital filter
• 100 A/D converter
• 200 Neural network
• C_A, C_B, C_B1, C_B2 capacitors
• CMP comparator
• PG pulse signal generating unit
• R, R₁, R₂ resistive loads
• SWA, SWA1, SWA2, SWB, SWB1, SWB2 Switches
• VCCS Voltage controlled current source

Claims

1. A neuron element comprising: a first charge accumulating unit configured to accumulate charges by an input signal;a signal processing unit which, based on a condition when the charges accumulated in the first charge accumulating unit exceed a first predetermined amount of charges, is configured to:discharge a second predetermined amount of charges from the first charge accumulating unit; andgenerate a pulse signal,wherein the signal processing unit performs a predetermined processing in relation to the second predetermined amount of charges and the pulse signal.

2. The neuron element according to claim 1, wherein the signal processing unit comprises: a pulse signal generating unit configured to generate the pulse signal, anda discharging unit which, as the predetermined processing, is configured to discharge the second predetermined amount of charges from the first charge accumulating unit based on the pulse signal generated by the pulse signal generating unit.

3. The neuron element according to claim 2, wherein the discharging unit is connected in parallel with the first charge accumulating unit and further comprises: a constant current source configured to discharge the second predetermined amount of charges from the first charge accumulating unit by causing a predetermined current during a period in which the pulse signal is generated by the pulse signal generating unit; anda determining unit configured to output an on signal based on the condition when the charges accumulated in the first charge accumulating unit is determined to exceed the first predetermined amount of charges, andwherein the pulse signal generating unit is configured to generate the pulse signal with a predetermined duty ratio while the on signal is inputted to the pulse signal generating unit from the determining unit.

4. The neuron element according to claim 3, wherein, when a maximum instantaneous input current of the input signal is Imax, a capacitance of the first charge accumulating unit is CA, a threshold voltage for detecting the first predetermined amount of charges is Vth, a source voltage is VDD, and a delay period, from a time when a voltage generated at the first charge accumulating unit exceeds the threshold voltage Vth to a time when the constant current source causes the predetermined current, is Dcmp, then Dcmp is less than CA · VDD/Imax.

5. The neuron element according to claim 1, wherein the signal processing unit comprises: a discharging unit configured to discharge the second predetermined amount of charges from the first charge accumulating unit, anda pulse signal generating unit which, as the predetermined processing, is configured to generate the pulse signal by using a voltage signal generated based on the second predetermined amount of charges discharged from the first charge accumulating unit by the discharging unit.

6. The neuron element according to claim 5, wherein the discharging unit further comprises:a second charge accumulating unit; a determining unit configured to output a first selecting signal based on the condition when the charges accumulated in the first charge accumulating unit is determined to exceed the first predetermined amount of charges, and output a second selecting signal based on a condition when the charges accumulated in the first charge accumulating unit is determined to fall below the first predetermined amount of charges; anda connection switching unit configured to connect the first charge accumulating unit and the second charge accumulating unit in parallel based on outputting of the first selecting signal from the determining unit, and connect a resistive load, which corresponds to the pulse signal generating unit, and the second charge accumulating unit in parallel based on outputting of the second selecting signal from the determining unit,wherein the second charge accumulating unit, when connected in parallel with the first charge accumulating unit by the connection switching unit, is configured to accumulate the second predetermined amount of charges discharged from the first charge accumulating unit, andwherein the resistive load, when connected in parallel with the second charge accumulating unit by the connection switching unit, is configured to generate a voltage signal corresponding to the pulse signal based on the second predetermined amount of charges accumulated in the second charge accumulating unit.

7. The neuron element according to claim 6, wherein, when a maximum instantaneous input current of the input signal is Imax, a capacitance of the first charge accumulating unit is CA, a threshold voltage for detecting the first predetermined amount of charges is Vth, a source voltage is VDD, a capacitance of the second charge accumulating unit is CB, a resistance value of the resistive load is R, and a delay period, from a time when a voltage generated at the first charge accumulating unit exceeds the threshold voltage Vth to a time of the connecting by the connection switching unit, is Dcmp, then Dcmp is less than (CA · CB)/(CA + CB)(VDD/Imax - R).

8. A differential neuron element comprising: a charge accumulating unit which accumulates charges by a first input signal and a second input signal, the second input signal obtained by inverting a polarity of the first input signal;a signal processing unit which, based on a condition when the charges accumulated in the charge accumulating unit exceed a first predetermined amount of charges, is configured to:discharge a second predetermined amount of charges from the charge accumulating unit; andgenerate a first pulse signal and a second pulse signal, the second pulse signal obtained by inverting a polarity of the first pulse signal,wherein the signal processing unit performs a predetermined processing in relation to the second predetermined amount of charges and the pulse signal.

9. An A/D converter comprising: a charge accumulating unit which accumulates charges by an input signal;a signal processing unit which, based on a condition when the charges accumulated in the charge accumulating unit exceed a first predetermined amount of charges, is configured to: discharge a second predetermined amount of charges from the charge accumulating unit, andgenerate a pulse signal; anda digital filter configured to perform a predetermined digital processing on the pulse signal generated by the signal processing unit,wherein the signal processing unit performs a predetermined processing in relation to the second predetermined amount of charges and the pulse signal.

10. A neural network comprising a plurality of neuron elements including the neuron element of claim 1, wherein an input signal of any one of the neuron elements is obtained by a weighting addition processing of at least one of signals of at least one external input signal and an output signal of at least another one of the neuron elements,wherein an output signal of the any one of the neuron element becomes at least one of signals of at least one external output signal and a weighted input signal of the at least another one of the neuron elements.