Neural network device and signal processing method

Abstract

Minimize the loss of information transmitted. [Solution] The neural network device comprises a plurality of synaptic circuits and a plurality of neuron circuits. The first neuron circuit among the plurality of neuron circuits has an input circuit, a charge holding circuit, a comparison circuit, a firing circuit, a charge control circuit, and a control signal output circuit. The firing circuit outputs a spike signal when the judgment signal changes from a second value to a first value. The control signal output circuit outputs a control signal that represents a comparison voltage based on the excess component that exceeds the threshold potential at the membrane potential. The first synaptic circuit among the plurality of synaptic circuits that acquires the spike signal from the first neuron circuit outputs a control signal and a synaptic current of a current amount corresponding to the synaptic weight when it acquires the spike signal from the first neuron circuit.

Description

[Technical Field]

[0001] Embodiments of the present invention relate to a neural network device and a signal processing method. [Background technology]

[0002] A spiking neural network (SNN) is a neural network that performs information processing using spike signals. The neural network circuit that implements a spiking neural network includes multiple neuron circuits and multiple synaptic circuits, each connecting two neuron circuits. Each of the multiple neuron circuits holds charge from the synaptic current supplied by the preceding synaptic circuit using a charge-holding unit such as a capacitor. Each of the multiple neuron circuits also has a comparator that compares the amount of charge held with a threshold, and generates a spike signal when the amount of charge held exceeds the threshold. In response to the generation of a spike signal, each of the multiple neuron circuits discharges the charge held in the charge-holding unit to reset the amount of charge held by the charge-holding unit. Each of the multiple synaptic circuits also receives a spike signal from the preceding neuron circuit and outputs a synaptic current obtained by multiplying the spike signal by a synaptic weight using a variable resistor or the like.

[0003] In recent years, techniques have become publicly known for replacing tasks such as image processing, which were previously performed using arithmetic AI (Artificial Intelligence) such as CNNs (Convolutional Neural Networks), with spiking neural networks. However, spiking neural networks perform discrete information processing. Therefore, when image processing and similar tasks are performed using spiking neural networks, there is a problem of information loss.

[0004] The cause of this information loss is that when the amount of charge held in the preceding neuronal circuit exceeds a threshold, the excess charge exceeding the threshold is not transmitted to the subsequent neuronal circuit. To solve this problem, for example, a spiking neural network is known that adjusts the reset discharge amount of the held charge using the magnitude of the excess charge. This spiking neural network reduces information loss by increasing the frequency of spike signals generated by the excess charge. However, since this spiking neural network applies the effect of the excess charge to the processing after the spike signal is generated, there is a time delay in the reflection of the excess charge.

[0005] Therefore, when performing tasks such as image processing with high accuracy using spiking neural networks, it is necessary to minimize such information loss and information transmission delays. [Prior art documents] [Non-patent literature]

[0006] [Non-Patent Document 1] Dengyu Wu et al., “A Little Energy Goes a Long Way: Build an Energy-Efficient, Accurate Spiking Neural Network From Convolutional Neural Network”, Frontiers in Neuroscience, May 2022 [Non-Patent Document 2] Jaesung Kim et al., “Design of a 180 nm CMOS Neuron Circuit with Soft-Reset and Underflow Allowing for Loss-Less Hardware Spiking Neural Networks”, Adv. Intell. Syst., Jun 2024 [Overview of the Initiative] [Problems that the invention aims to solve]

[0007] The problem that this invention aims to solve is to provide a highly accurate spiking-type neural network device and signal processing method that minimizes the loss of transmitted information and reduces information transmission delay. [Means for solving the problem]

[0008] The neural network device according to the embodiment comprises a plurality of synaptic circuits, each with a set synaptic weight, and a plurality of neuron circuits, each outputting a spike signal which is a voltage pulse. Each of the plurality of synaptic circuits acquires the spike signal output from any one of the plurality of neuron circuits, and when it acquires the spike signal, it outputs a synaptic current of an amount corresponding to the set synaptic weight. The first neuron circuit among the plurality of neuron circuits has an input circuit, a charge holding circuit, a comparison circuit, a firing circuit, a charge control circuit, and a control signal output circuit. The input circuit acquires the synaptic current from each of at least one of the plurality of synaptic circuits. The charge holding circuit accumulates a charge corresponding to the synaptic current acquired by the input circuit and generates a membrane potential corresponding to the accumulated charge. The comparison circuit generates a determination signal which becomes a first value when the sign of the difference voltage between a preset threshold potential and the membrane potential is a first sign, and becomes a second value when the sign of the difference voltage is a second sign different from the first sign. The firing circuit outputs the spike signal when the determination signal changes from the second value to the first value. The charge control circuit, after outputting the spike signal, charges and discharges the charge stored in the charge holding circuit so that the sign of the differential voltage becomes the second sign. The control signal output circuit outputs a control signal representing a comparison voltage based on the excess component that exceeds the threshold potential at the membrane potential. Each of the at least one first synaptic circuit among the plurality of synaptic circuits that acquires the spike signal from the first neuron circuit outputs the control signal and the synaptic current of an amount corresponding to the synaptic weight when it acquires the spike signal from the first neuron circuit. [Brief explanation of the drawing]

[0009] [Figure 1] A diagram showing an example of the configuration of a neural network device. [Figure 2] A diagram showing the connection relationships of the peripheral circuits of the first neuron circuit in the first embodiment. [Figure 3] A diagram showing the configuration of the first neuron circuit according to the first embodiment. [Figure 4] A diagram showing the first example of the configuration of a control signal output circuit. [Figure 5] This figure shows a second example of the configuration of a control signal output circuit. [Figure 6] A diagram showing the first example of the configuration of the first synaptic circuit. [Figure 7] A diagram showing a second example of the configuration of the first synaptic circuit. [Figure 8] A figure showing an example of membrane potential and synaptic current waveforms in the first embodiment. [Figure 9] A diagram showing the connection relationships of the peripheral circuits of the first neuron circuit in the second embodiment. [Figure 10] A diagram showing the configuration of the first neuron circuit according to the second embodiment. [Figure 11] A figure showing an example of membrane potential and synaptic current waveforms in the second embodiment. [Figure 12] A diagram showing the connection relationships of the peripheral circuits of the first neuron circuit in the third embodiment. [Figure 13] A diagram showing the configuration of the first neuron circuit according to the third embodiment. [Figure 14] A figure showing an example of membrane potential and synaptic current waveforms in the third embodiment. [Figure 15] A diagram showing the configuration of the first neuron circuit according to the first modified example. [Figure 16] A diagram showing the configuration of the first neuron circuit in the second modified example. [Figure 17] A diagram showing the configuration of the first neuron circuit in the third modified example. [Figure 18] A diagram showing the configuration of the first neuron circuit according to the fourth modified example. [Figure 19] A diagram showing the configuration of the first neuron circuit according to the fifth modified example. [Figure 20] A diagram showing the configuration of the first neuron circuit according to the sixth modified example. [Figure 21] A diagram showing the configuration of the first neuron circuit according to the seventh modified example. [Figure 22]A diagram showing the configuration of the first neuron circuit according to the eighth modified example. [Figure 23] A diagram showing the configuration of the first neuron circuit according to the ninth modified example. [Figure 24] A diagram showing the configuration of the first neuron circuit according to the 10th modified example. [Figure 25] A diagram showing a neural network device in the first example of a hierarchical configuration. [Figure 26] A diagram showing a neural network device in the second example of a hierarchical configuration. [Figure 27] A diagram showing a neural network device in the third hierarchical configuration example. [Figure 28] A diagram showing a neural network device in the fourth hierarchical configuration example. [Figure 29] A diagram showing a neural network device in the fifth hierarchical configuration example. [Figure 30] A diagram showing the connections around the first neuron circuit in the sixth example of a hierarchical configuration. [Figure 31] A diagram showing the connections around the first neuron circuit in the seventh example of a hierarchical configuration. [Figure 32] A diagram showing the connections around the first neuron circuit in the eighth example of a hierarchical configuration. [Figure 33] A diagram showing the connections around the first neuron circuit in the ninth example of hierarchical configuration. [Modes for carrying out the invention]

[0010] The neural network device 10 according to an embodiment will be described below with reference to the drawings.

[0011] (First Embodiment) The neural network device 10 according to the first embodiment is a spiking-type neural network configured with hardware. For example, the neural network device 10 is mounted on a semiconductor device using a process such as CMOS (Complementary Metal Oxide Semiconductor).

[0012] Figure 1 shows an example of the configuration of the neural network device 10. The neural network device 10 according to the first embodiment includes, as an example, M layers 12 (where M is an integer of 2 or more) and (M-1) synaptic groups 14.

[0013] Each of the (M-1) synaptic groups 14 contains multiple synaptic circuits 20. Each of the multiple synaptic circuits 20 is assigned a synaptic weight. The synaptic weights assigned to the multiple synaptic circuits 20 are set by a learning process or random numbers, etc. For example, the synaptic weights assigned to the multiple synaptic circuits 20 may be updated by a predetermined update rule such as STDP (Spike Timing Dependent Plasticity) or SDSP (Spike Driven Synaptic Plasticity).

[0014] Each of the M-stage layers 12 contains multiple neuron circuits 22. Each of the multiple neuron circuits 22 outputs a spike signal. The spike signal is a voltage pulse that changes from a second voltage to a first voltage, and then returns to the second voltage after a certain period of time has elapsed since the change from the second voltage to the first voltage.

[0015] The m-th synaptic group 14 (where m is an integer between 1 and (M-1)) of the (M-1) synaptic groups 14 is located between the m-th layer 12 of the M-level layer 12 and the (m+1)-th layer 12 of the M-level layer 12.

[0016] Each of the multiple synaptic circuits 20 included in the m-th synapse group 14 acquires a spike signal output from one of the multiple neuron circuits 22 included in the m-th layer 12. When each of the multiple synaptic circuits 20 included in the m-th synapse group 14 acquires a spike signal, it outputs a synaptic current corresponding to the set synaptic weight and the acquired spike signal. The synaptic weight may be represented as a binary value or as a discrete value with three or more values. The synaptic weight may also be represented as an analog value by the amount of charge stored in a capacitor or the resistance value of a variable resistor.

[0017] Then, each of the multiple synaptic circuits 20 included in the m-th synaptic group 14 supplies a synaptic current to one of the multiple neuronal circuits 22 included in the (m+1)-th layer 12.

[0018] Each of the multiple neuronal circuits 22 contained in the (m+1)th layer 12 of the M-stage layer 12 acquires multiple synaptic currents output from the m-stage synapse group 14 and performs a process equivalent to a sum-of-accumulate operation on the acquired synaptic currents. The first layer 12 of the M-stage layer 12 acquires multiple signals from an external device or input layer. Each of the multiple neuronal circuits 22 then outputs a spike signal that has been processed with an activation function equivalent to the signal representing the calculation result.

[0019] In such a neural network device 10, the first layer 12 receives one or more signals from an external device or input layer. The neural network device 10 then outputs one or more signals from the Mth layer 12 that represent the result of performing calculations by the neural network on the received one or more signals.

[0020] Such a neural network device 10 executes arithmetic neural network operations such as CNN. As a result, the neural network device 10 can execute tasks such as image recognition and classification processing with low energy consumption and a small-scale circuit without using a CPU and a GPU.

[0021] FIG. 2 is a diagram showing the connection relationship of the peripheral circuits of the first neuron circuit 32 in the first embodiment.

[0022] Each of the plurality of neuron circuits 22 holds an internal potential called membrane potential V mem When the neuron circuit 22 obtains a synaptic current from any of the plurality of synaptic circuits 20 connected to the previous stage, the membrane potential V mem is increased according to the magnitude of the obtained synaptic current. As a result, each of the plurality of neuron circuits 22 can execute a process corresponding to a sum-of-products operation on the plurality of obtained synaptic currents.

[0023] If each of the plurality of neuron circuits 22 does not obtain a synaptic current, the membrane potential V mem may be decreased over time. In this case, if each of the plurality of neuron circuits 22 continuously obtains synaptic currents at short time intervals, the membrane potential V mem is increased, and if the synaptic current is not obtained for a long time, the membrane potential V mem is decreased. Note that when the membrane potential V mem is decreased over time and the membrane potential V mem reaches a predetermined initial potential, the decrease of the membrane potential V mem is stopped.

[0024] Then, when the membrane potential V mem of each of the plurality of neuron circuits 22 rises and becomes equal to or higher than a predetermined threshold potential V th , it fires and outputs a spike signal to the synaptic circuit 20 in the subsequent stage. After each of the plurality of neuron circuits 22 fires, the membrane potential Vmem Return to the initial potential.

[0025] Furthermore, each of the multiple neuronal circuits 22 maintains its membrane potential V even when a synaptic current is applied during a predetermined refractory period after firing. mem The threshold potential V may not be increased, and further firing may be stopped. In this case, each of the multiple neuronal circuits 22 begins accumulating charge according to the synaptic current after the refractory period ends. The initial potential is the threshold potential V th It is smaller than that.

[0026] Each of the multiple synaptic circuits 20 acquires a spike signal output from one of the multiple neuronal circuits 22.

[0027] Each of the multiple synaptic circuits 20 has a circuit that generates current. When each of the multiple synaptic circuits 20 acquires a spike signal, it uses the current-generating circuit to output a synaptic current to the subsequent neuron circuit 22, with a current amount corresponding to the set synaptic weight and the acquired spike signal.

[0028] Here, the first neuron circuit 32 among the multiple neuron circuits 22 outputs a control signal along with a spike signal.

[0029] The control signal is the membrane potential V mem The threshold potential V th This represents the comparative voltage based on the excess component when it exceeds the membrane potential V. The excess component is the membrane potential V. mem In this case, the threshold potential V th This is a component that exceeds the limit. For example, the comparison voltage is the membrane potential V mem Furthermore, for example, the comparison voltage is the membrane potential V mem From the threshold potential V th The differential voltage obtained by subtracting the voltage may also be used. Such a control signal is the membrane potential V at the time the spike signal is fired. mem Threshold potential V th It can represent excess components that exceed a certain threshold.

[0030] The control signal may be an analog voltage or digital data.

[0031] Furthermore, each of the multiple synaptic circuits 20, at least one of which acquires a spike signal from the first neuron circuit 32, acquires a control signal along with the spike signal from the first neuron circuit 32.

[0032] Each of the at least one first synaptic circuit 30, upon receiving a spike signal from the first neuron circuit 32, outputs a synaptic current with a current amount corresponding to the acquired control signal and synaptic weight. For example, each of the at least one first synaptic circuit 30 increases the amplitude of the output synaptic current as the comparison voltage represented by the control signal increases. As a result, each of the at least one first synaptic circuit 30 can increase the amount of synaptic current as the excess component at firing of the first neuron circuit 32 increases, assuming the synaptic weight is a fixed value.

[0033] Figure 3 shows the configuration of the first neuron circuit 32 according to the first embodiment, along with the first synaptic circuit 30 connected to the first neuron circuit 32.

[0034] Furthermore, in the neural network device 10, all of the multiple neuron circuits 22 may have the same configuration as the first neuron circuit 32, or some of the multiple neuron circuits 22 may have the same configuration as the first neuron circuit 32.

[0035] The first neuron circuit 32 is connected to one or more of the synaptic circuits 20 in its preceding stage. The first neuron circuit 32 is supplied with synaptic current from each of the one or more synaptic circuits 20 connected to its preceding stage.

[0036] Furthermore, one or more first synaptic circuits 30 are connected to the first neuron circuit 32. The first neuron circuit 32 outputs a spike signal and a control signal to each of the one or more first synaptic circuits 30.

[0037] The first neuron circuit 32 comprises an input circuit 38, a charge holding circuit 40, a comparison circuit 42, a firing circuit 44, a charge control circuit 46, and a control signal output circuit 48.

[0038] The input circuit 38 acquires synaptic currents from each of at least one synaptic circuit 20 connected to the first neuron circuit 32 among the multiple synaptic circuits 20. The input circuit 38 may be simply an electric wire or an electrode terminal, etc.

[0039] The charge retention circuit 40 stores charge corresponding to the amount of synaptic current acquired by the input circuit 38. The charge retention circuit 40 stores the membrane potential V corresponding to the stored charge. mem Therefore, each time a synaptic current is supplied, the charge holding circuit 40 adjusts the membrane potential V in accordance with the amount of the supplied synaptic current. mem This increases the voltage. For example, in the charge holding circuit 40, one terminal is connected to the ground terminal, and the other terminal is connected to the membrane potential V mem This is a capacitor that generates [something].

[0040] The comparison circuit 42 uses the membrane potential V generated from the charge holding circuit 40. mem The comparison circuit 42 obtains the preset threshold potential V. th The comparison circuit 42 obtains the membrane potential V. mem The threshold potential V th When it is greater than this, it becomes the first value, and the membrane potential V mem The threshold potential V th If the value is not greater than the second value, a judgment signal is generated.

[0041] For example, the comparison circuit 42 is a comparator mounted on a semiconductor device. The comparator has a threshold potential V at its inverting input terminal. th A membrane potential V is given to the non-inverting input terminal.mem The following is given. Then, the comparator determines the membrane potential V mem The threshold potential V th It outputs a judgment signal indicating whether it is greater than or less than the membrane potential V. For example, a comparator will output a judgment signal indicating whether it is greater than or less than the membrane potential V. mem The threshold potential V th If it is determined to be greater than the first value (e.g., H logic), the membrane potential V mem The threshold potential V th If it is determined that the value is not greater than the specified value, a judgment signal with a second value (e.g., L logic) is output.

[0042] The ignition circuit 44 acquires a determination signal from the comparison circuit 42. When the determination signal changes from the second value to the first value, the ignition circuit 44 outputs a spike signal, which is a voltage pulse with a predetermined time width. In other words, the ignition circuit 44 determines that the comparison circuit 42 has a membrane potential V mem The threshold potential V th From a state where the membrane potential V is not large, mem The threshold potential V th When the state changes to a larger value, it outputs a spike signal, which is a voltage pulse. For example, the firing circuit 44 outputs a spike signal when the membrane potential V mem The threshold potential V th From a state where the membrane potential V is not large, mem The threshold potential V th When the voltage changes to a larger state, it outputs a spike signal that changes from the second voltage to the first voltage, and then returns to the second voltage after a certain period of time has elapsed since the change from the second voltage to the first voltage.

[0043] The charge control circuit 46 releases the charge stored in the charge holding circuit 40 after the ignition circuit 44 has output a spike signal. For example, the charge control circuit 46 releases the charge stored in the charge holding circuit 40 for a predetermined period after the ignition circuit 44 has output a spike signal.

[0044] For example, the charge control circuit 46 controls the membrane potential V in the charge holding circuit 40. mem By connecting the terminal that generates the charge to the ground terminal, the charge accumulated in the charge holding circuit 40 is released. For example, the charge control circuit 46 controls the membrane potential V in the charge holding circuit 40.mem This is a switch that turns on or off the connection between the terminal generating the current and the ground terminal. For example, the switch can be implemented using a MOSFET (metal-oxide-semiconductor field-effect transistor) mounted on a semiconductor device.

[0045] This charge control circuit 46 releases the charge accumulated in the charge holding circuit 40 to the ground terminal, thereby generating the membrane potential V from the charge holding circuit 40. mem The membrane potential V can be returned to its initial potential. Furthermore, the charge control circuit 46 can prevent charge accumulation in the charge holding circuit 40 by directing the synaptic current supplied from the preceding synaptic circuit 20 to the ground terminal. In addition, the charge control circuit 46 can control the membrane potential V generated from the charge holding circuit 40. mem Alternatively, the potential of the ground terminal can be supplied to the comparison circuit 42 to prevent the generation of spike signals. The charge control circuit 46 can then stop the discharge of charge from the charge holding circuit 40 after a predetermined period has elapsed since the start of charge discharge, allowing the charge holding circuit 40 to accumulate charge corresponding to the synaptic current acquired by the input circuit 38.

[0046] The control signal output circuit 48 acquires the comparison voltage. Then, the control signal output circuit 48 outputs a control signal representing the acquired comparison voltage. The comparison voltage is the membrane potential V mem Threshold potential V th This is the voltage based on the excess component that exceeds a certain threshold.

[0047] The control signal output circuit 48 may output an analog voltage representing the comparison voltage as the control signal. Alternatively, the control signal output circuit 48 may output digital data representing the comparison voltage as the control signal.

[0048] The first neuronal circuit 32 with this configuration has a membrane potential V mem The threshold potential V th Along with the spike signal indicating the timing of the increase in intensity, the membrane potential V mem Threshold potential Vth It can output a control signal representing the excess component that exceeds a certain threshold.

[0049] Furthermore, the first neuron circuit 32 shown in Figure 3 has a threshold potential V th However, this configuration is for cases where the threshold potential V is greater than a predetermined reference potential, such as the ground potential. However, the first neuron circuit 32 has a threshold potential V th However, a configuration with a negative sign inverted and a smaller positive potential than the reference potential, such as the ground potential, is also acceptable.

[0050] In the case of a configuration where the positive and negative signs are reversed, the charge holding circuit 40 adjusts the membrane potential V according to the amount of the supplied synaptic current each time a synaptic current is supplied. mem The membrane potential V is lowered. Also, the comparison circuit 42 controls the membrane potential V. mem The threshold potential V th When it is smaller, it becomes the first value, and the membrane potential V mem The threshold potential V th If the value is not smaller, a determination signal is generated that becomes the second value. In addition, the charge control circuit 46 causes the charge holding circuit 40 to charge after the ignition circuit 44 has output a spike signal.

[0051] Therefore, whether the configuration is one in which the positive and negative signs are not inverted, or one in which the positive and negative signs are inverted, the comparison circuit 42 is at the threshold potential V th and membrane potential V mem A determination signal is generated which becomes the first value when the sign of the differential voltage is the first sign, and the second value when the sign of the differential voltage is the second sign, which is different from the first sign. In addition, after outputting the spike signal, the charge control circuit 46 charges and discharges the charge stored in the charge holding circuit 40 so that the sign of the differential voltage becomes the second sign. The same applies to the second embodiment and subsequent embodiments.

[0052] Figure 4 shows a first example of the configuration of the control signal output circuit 48. The control signal output circuit 48 may include, for example, a first resistor 50 and a voltage-controlled MOSFET 52, as shown in Figure 4. One terminal of the first resistor 50 is connected to the ground terminal.

[0053] The voltage-controlled MOSFET 52 is, for example, a p-type MOSFET. The drain of the voltage-controlled MOSFET 52 is connected to the power supply voltage terminal, and the source is connected to the terminal of the first resistor 50 that is not connected to the ground terminal. A comparison voltage is then applied to the gate of the voltage-controlled MOSFET 52. The voltage-controlled MOSFET 52 changes the amount of current flowing between the drain and source according to the comparison voltage. Therefore, the voltage-controlled MOSFET 52 can change the voltage generated across the first resistor 50 according to the comparison voltage.

[0054] In this configuration, the control signal output circuit 48 outputs the voltage generated from the terminal of the first resistor 50 that is not connected to ground as a control signal. As a result, the control signal output circuit 48 can output a control signal represented by an analog voltage.

[0055] Figure 5 shows a second example of the configuration of the control signal output circuit 48. The control signal output circuit 48 may also include an AD conversion circuit 54, as shown in Figure 5.

[0056] The AD conversion circuit 54 acquires a comparison voltage, converts the comparison voltage from analog to digital, and generates digital data representing the comparison voltage. For example, the AD conversion circuit 54 is an integral-type AD converter that alternately repeats a sampling period and a hold period at a predetermined cycle. In this case, the AD conversion circuit 54 stores the comparison voltage in the sampling capacitor during the sampling period. Subsequently, during the hold period, the AD conversion circuit 54 uses a counter to measure the time it takes for the charge stored in the sampling capacitor to be released and the voltage of the sampling capacitor to drop to a predetermined voltage. The AD conversion circuit 54 then outputs digital data representing the counter's count value as a control signal.

[0057] As a result, the control signal output circuit 48 can output a control signal represented by digital data. Note that the AD conversion circuit 54 is not limited to an integrating type; it may be an AD converter of another type.

[0058] Figure 6 shows a first example of the configuration of the first synaptic circuit 30. For example, the first synaptic circuit 30 may have the configuration shown in Figure 6. For example, the first synaptic circuit 30 includes a synaptic weight holding circuit 56, a control signal input circuit 58, and a propagation circuit 60.

[0059] The synaptic weight holding circuit 56 holds a preset synaptic weight (W). The synaptic weight holding circuit 56 may include, for example, a variable resistor whose resistance value changes according to the synaptic weight (W). The synaptic weight holding circuit 56 may also include, for example, a holding capacitor that stores charge according to the synaptic weight (W). Furthermore, the synaptic weight holding circuit 56 may have a memory or the like to store a digital value according to the synaptic weight (W). In this embodiment, the synaptic weight holding circuit 56 holds a weight voltage (V) according to the set synaptic weight (W). W ) will occur.

[0060] The control signal input circuit 58 receives a control signal from the first neuron circuit 32. The control signal input circuit 58 receives a control voltage (V) corresponding to the control signal. cont ) outputs. If the control signal is an analog voltage, the control signal input circuit 58 buffers or amplifies the control signal to output a control voltage (V cont ) can be output as a control voltage (V cont ) may be output as a control voltage (V). If the control signal is digital data, the control signal input circuit 58 converts the control signal to an analog voltage and outputs the converted analog voltage as a control voltage (V). cont Output as ).

[0061] The propagation circuit 60 receives a spike signal (S) from the first neuron circuit 32. in The propagation circuit 60 also receives a weight voltage (V) from the synaptic weight holding circuit 56. W The propagation circuit 60 receives a control voltage (V) from the control signal input circuit 58. cont ) receive.

[0062] The propagation circuit 60 outputs a synaptic current (I in ) with a current amount corresponding to a spike signal (S cont ), a synaptic weight (W), and a control voltage (V s ). For example, when the control voltage (V cont ) is a predetermined value greater than 0 and the spike signal (S in ) is the first voltage, the propagation circuit 60 outputs a synaptic current (I cont ) with a larger current amount as the control voltage (V s ) increases. Also, for example, when the control voltage (V cont ) is a predetermined value and the spike signal (S in ) is the first voltage, the propagation circuit 60 outputs a synaptic current (I s ) with a larger current amount as the synaptic weight (W) increases. Further, when the spike signal (S in ) is the second voltage, the propagation circuit 60 does not output a synaptic current (I cont ) regardless of the synaptic weight (W) and the control voltage (V s ).

[0063] For example, the propagation circuit 60 includes a weight current circuit 62, a spike input circuit 64, a first capacitor 66, an output amplifier circuit 68, a charge adjustment circuit 72, a current supply circuit 74, and a current control circuit 76.

[0064] When the spike signal (S in ) is the first voltage, the weight current circuit 62 passes a weight current (I W ) with a current value corresponding to the synaptic weight (W) set in the synaptic weight holding circuit 56. For example, the weight current circuit 62 passes a weight current (I W ) proportional to the synaptic weight (W).

[0065] The weight current circuit 62 is, for example, a MOSFET. In the example of FIG. 6, the weight current circuit 62 is an N-channel MOSFET. The weight current circuit 62 that is a MOSFET has a weight voltage (V W) is applied, and the drain is connected to node A. Then, the weight current circuit 62, which is a MOSFET, allows a weight current (I W ) corresponding to the weight voltage (V W ) to flow between the drain and the source.

[0066] The spike input circuit 64 receives the spike signal (S in ) output from the first neuron circuit 32. The spike input circuit 64 switches whether to allow the weight current (I in ) to flow through the weight current circuit 62 according to the spike signal (S W ) output from the first neuron circuit 32. For example, the spike input circuit 64 allows the weight current (I in ) to flow when the spike signal (S W ) is at the first voltage. For example, the spike input circuit 64 does not allow the weight current (I in ) to flow when the spike signal (S W ) is at the second voltage. That is, the spike input circuit 64 sets the weight current (I in ) to 0 when the spike signal (S W ) is at the second voltage.

[0067] In this embodiment, the spike input circuit 64 is a MOSFET that performs a switching operation. In the example of FIG. 6, the spike input circuit 64 is an N-channel MOSFET. The spike input circuit 64, which is a MOSFET, has a spike signal (S in ) applied to its gate, its drain connected to the source of the weight current circuit 62, and its source connected to the reference potential (ground).

[0068] Then, the spike input circuit 64, which is a MOSFET, turns on when the spike signal (S in ) is at the first voltage, and connects the source of the weight current circuit 62 to ground, thereby allowing the weight current (I W ) to flow through the weight current circuit 62. Also, the spike input circuit 64, which is a MOSFET, inWhen the second voltage is present, it turns off, and by disconnecting the source of the weight current circuit 62 from ground, the weight current (I W ) Do not let it flow.

[0069] The first capacitor 66 has a first terminal 66a and a second terminal 66b. The first terminal 66a of the first capacitor 66 is connected to a power supply voltage that generates a constant voltage. In such a first capacitor 66, a constant voltage is applied to the first terminal 66a. Also, the first capacitor 66 has a capacitor voltage (V) applied to the second terminal 66b. C ) generates a capacitor voltage (V C This value is obtained by subtracting the voltage generated by the first capacitor 66 from the power supply potential. The voltage generated by the first capacitor 66 is the voltage obtained by dividing the stored charge by the capacitance.

[0070] The output amplifier circuit 68 generates a capacitor voltage (V) at the second terminal 66b of the first capacitor 66. C Synaptic current (I) corresponding to s It outputs a capacitor voltage (V) to the gate. For example, the output amplifier circuit 68 is a MOSFET. In the example in Figure 6, the output amplifier circuit 68 is a P-channel MOSFET. The output amplifier circuit 68, which is a P-channel MOSFET, outputs a capacitor voltage (V) to the gate. C A voltage (V) is applied, the source is connected to the power supply potential, and the drain is connected to the output terminal of the synaptic current. Then, the output amplifier circuit 68, which is a MOSFET, receives the capacitor voltage (V) C Synaptic current (I) corresponding to s ) is then flowed between the drain and source.

[0071] The charge adjustment circuit 72 controls the weight current (I) via the spike input circuit 64. W If ) flows, the weight current (I W The charge adjustment circuit 72 decreases or increases the charge stored in the first capacitor 66 by a time-varying amount corresponding to the current value of the first capacitor 66. For example, when the spike input circuit 64 is ON, the charge adjustment circuit 72 decreases or increases the charge stored in the first capacitor 66.

[0072] Furthermore, the charge adjustment circuit 72 controls the weight current (I) via the spike input circuit 64. W If no current flows, the charge stored in the first capacitor 66 remains constant. For example, the charge adjustment circuit 72 does not change the charge stored in the first capacitor 66 when the spike input circuit 64 is in the off state.

[0073] For example, the charge regulating circuit 72 is a diode-connected MOSFET. In the example in Figure 6, the charge regulating circuit 72 is a diode-connected N-channel MOSFET. The charge regulating circuit 72, being an N-channel MOSFET, has its gate and drain connected, the second terminal 66b of the first capacitor 66 connected to the gate, and its source connected to node A. The charge regulating circuit 72, being a MOSFET, then supplies a weight current (I W When ) flows, the capacitor current (I C ) is drawn out and supplied to node A. Alternatively, the charge adjustment circuit 72, which is a MOSFET, draws out capacitor current (I C The capacitor extracts the ions and supplies them to the second terminal 66b of the first capacitor 66.

[0074] As a result, the charge adjustment circuit 72 receives the weight current (I) from the spike input circuit 64. W When a current flows through the capacitor, the capacitor voltage (V C ) reduces the weight current (I W If no current flows through the capacitor, the capacitor voltage (V C The charge adjustment circuit 72 can keep the spike signal (S) constant when the spike input circuit 64 is ON, that is, when the spike signal (S) is ON. in When ) is the first voltage, the capacitor voltage (V C ) can be reduced. Also, the charge adjustment circuit 72 can reduce the spike signal (S) when the spike input circuit 64 is off. in When ) is the second voltage, the capacitor voltage (V C ) can be kept constant.

[0075] The current supply circuit 74 receives the weight current (I) from the spike input circuit 64. W When a predetermined voltage (V) is applied, tau The resistance is determined by ), and the potential V of the second terminal 66b of the first capacitor 66. C Therefore, the first current (I tau ) changes. Weight current (I W If the first capacitor 66 is in a discharged state and the current does not flow, the first current (I tau The first capacitor 66 is charged by the spike signal (S). in ) against the spike current I S A delay occurs in the change in intensity. For example, the current supply circuit 74 is a MOSFET. In the example in Figure 6, the current supply circuit 74 is a P-channel MOSFET. The current supply circuit 74, which is a P-channel MOSFET, has a predetermined constant voltage (V) at its gate. tau A current is applied, the source is connected to the power supply potential, and the drain is connected to the drain of the charge adjustment circuit 72. Then, this current supply circuit 74 is supplied with a weight current (I) by the spike input circuit 64. W When ) flows, the potential V of the second terminal 66b of the first capacitor 66 is C Therefore, the first current (I tau ) changes. Also, the weight current (I W If the first capacitor 66 is in a discharged state and the current does not flow, the first current (I tau This causes the first capacitor 66 to be charged. This results in a spike current I S There is a delay in the change in intensity.

[0076] The current control circuit 76 controls the weight current (I) via the spike input circuit 64. W When a current flows through the control voltage (V cont ) Control current (I cont ) is supplied to node A. For example, the current control circuit 76 is a MOSFET. In the example in Figure 6, the current control circuit 76 is an N-channel MOSFET. The current control circuit 76, which is an N-channel MOSFET, supplies a control voltage (V) to its gate. contA current is applied, the drain is connected to the power supply potential, and the source is connected to node A. Then, the current control circuit 76 is controlled by the spike input circuit 64 to apply a weight current (I W When ) flows, the control current (I cont ) is supplied to node A via the charge adjustment circuit 72.

[0077] In such a propagation circuit 60, the capacitor voltage (V C ) is a spike signal (S in The synaptic current (I) begins to decrease from the first time step when the voltage changes from the second voltage to the first voltage. S ) is the capacitor voltage (V C A current of a certain amount is passed through, corresponding to the value of the current.

[0078] Here, the capacitor voltage (V C The time rate of change of ) is the weight current (I W ) and control current (I cont It changes depending on the magnitude of the weight current (I W ) is the amount of current corresponding to the synaptic weight (W) held by the synaptic weight holding circuit 56. Control current (I cont ) is the amount of current corresponding to the control signal supplied from the first neuron circuit 32. Therefore, the first synaptic circuit 30 with such a configuration has a synaptic current (I) corresponding to the synaptic weight (W) and the control signal. s It can output (this).

[0079] Figure 7 shows a second example of the configuration of the first synaptic circuit 30. For example, when the control signal is digital data, the first synaptic circuit 30 may have the configuration shown in Figure 7. When the control signal is digital data, for example, the first synaptic circuit 30 has a synaptic current source 82, a synaptic output circuit 84, and a current control controller 86.

[0080] The synaptic current source 82 is a variable current source. The synaptic current source 82 controls the amount of synaptic current (I) according to the control by the current control controller 86. s Outputs ).

[0081] The synaptic output circuit 84 receives the spike signal (S) output from the first neuron circuit 32. in The synaptic output circuit 84 receives the spike signal (S) output from the first neuron circuit 32. in ) In accordance with the synaptic current output from the synaptic current source 82, the synaptic current (I s The synaptic output circuit 84 switches whether or not to output the spike signal (S) to the outside. in If ) is the first voltage, the synaptic current (I s ) outputs to the outside, and spike signal (S in If ) is the second voltage, the synaptic current (I s ) Prevent it from being output externally.

[0082] For example, the synaptic output circuit 84 is a MOSFET. In the example in Figure 7, the output amplifier circuit 68 is an N-channel MOSFET. The synaptic output circuit 84, which is an N-channel MOSFET, receives a spike signal (S) at its gate. in A current (S) is applied, the drain is connected to the synaptic current source 82, and the source is connected to the output terminal. The synaptic output circuit 84, which is a MOSFET, receives the spike signal (S) in ) Depending on the synaptic current (I s Switches whether or not to play ).

[0083] The current control controller 86 is configured, for example, by a digital circuit. The current control controller 86 acquires control signals, which are digital data, from the first neuron circuit 32. The synaptic weights are also set externally for the current control controller 86. Then, the current control controller 86 outputs synaptic current (I) from the synaptic current source 82 according to the synaptic weights and control signals. s It controls the amount of current in the )

[0084] The first synaptic circuit 30 with this configuration generates a synaptic current (I) corresponding to the synaptic weight (W) and the control signal. s It can output (this).

[0085] Figure 8 shows the membrane potential V in the first embodiment. mem The figure also shows an example of a synaptic current waveform.

[0086] Synaptic circuits 20 other than the first synaptic circuit 30 do not receive the control signal. Therefore, synaptic circuits 20 other than the first synaptic circuit 30 do not receive the membrane potential V. mem Threshold potential V th Regardless of the magnitude of the excess component exceeding the threshold, the amplitude of the synaptic current does not change.

[0087] In response, the first neuron circuit 32 according to the first embodiment provides a control signal to the subsequent first synaptic circuit 30 that represents a comparison voltage based on the excess component at the time of firing of the spike signal. The first synaptic circuit 30, upon receiving the control signal, changes the amplitude of the synaptic current according to the magnitude of the control signal. For example, the first synaptic circuit 30 controls the membrane potential V mem Threshold potential V th When the excess component exceeding the threshold is large, it outputs a synaptic current with a larger amplitude than when the excess component is small.

[0088] As described above, the neural network device 10 according to the first embodiment can reflect the excess component at firing in the first neuron circuit 32 in the amount of synaptic current output from the first synaptic circuit 30 downstream of the first neuron circuit 32. As a result, the neural network device 10 according to the first embodiment can realize a highly accurate spiking-type neural network with less loss of transmitted information and less information transmission delay.

[0089] (Second Embodiment) Next, the neural network device 10 according to the second embodiment will be described. The neural network device 10 according to the second embodiment has substantially the same functions and configuration as the first embodiment, so components with substantially the same functions and configurations are given the same reference numerals as in the first embodiment, and detailed descriptions will be omitted except for differences.

[0090] Figure 9 is a diagram showing the connection relationships of the peripheral circuits of the first neuron circuit 32 in the second embodiment.

[0091] The first neuron circuit 32 according to the second embodiment outputs a spike signal but does not output a control signal. However, the first neuron circuit 32 changes the time width of the voltage pulse of the spike signal according to the control signal. For example, the first neuron circuit 32 outputs a voltage pulse of the membrane potential V mem Threshold potential V th The larger the excess component exceeding the threshold, the longer the duration of the voltage pulse in the spike signal.

[0092] Each of the at least one first synaptic circuit 30, upon receiving a spike signal from the first neuron circuit 32, outputs a synaptic current whose current amount corresponds to the time width of the voltage pulse of the acquired control signal and the synaptic weight. For example, each of the at least one first synaptic circuit 30 increases the time width for which it outputs the synaptic current as the time width of the voltage pulse of the spike signal increases. This allows each of the at least one first synaptic circuit 30 to increase the amount of synaptic current as the excess component at the time of firing of the first neuron circuit 32 increases.

[0093] Figure 10 shows the configuration of the first neuron circuit 32 according to the second embodiment, along with the first synaptic circuit 30 connected to the first neuron circuit 32.

[0094] In the second embodiment, the control signal output circuit 48 provides the control signal to the firing circuit 44 instead of outputting it to the first synapse circuit 30.

[0095] The ignition circuit 44 outputs a spike signal, which is a voltage pulse with a time width corresponding to the control signal, when the judgment signal obtained from the comparison circuit 42 changes from a second value to a first value. For example, the ignition circuit 44 increases the time width of the voltage pulse of the spike signal as the comparison voltage represented by the control signal increases.

[0096] The first neuronal circuit 32 with this configuration has a membrane potential V mem Threshold potential V th It can output a spike signal with a pulse width corresponding to the excess component that exceeds a certain threshold.

[0097] Figure 11 shows the membrane potential V in the second embodiment. mem The figure also shows an example of a synaptic current waveform.

[0098] Synaptic circuits 20 other than the first synaptic circuit 30 receive spike signals with a fixed voltage pulse duration. Therefore, synaptic circuits 20 other than the first synaptic circuit 30 receive a membrane potential V mem Threshold potential V th Regardless of the magnitude of the excess component exceeding a certain threshold, the duration for generating synaptic currents remains unchanged.

[0099] In contrast, the first neuron circuit 32 according to the second embodiment supplies the subsequent first synaptic circuit 30 with spike signals of voltage pulses of different time widths depending on the excess component. The first synaptic circuit 30 then changes the duration for which it generates synaptic current according to the pulse width of the spike signal. For example, the first synaptic circuit 30 uses the membrane potential V mem Threshold potential V th When the excess component exceeding the threshold is large, the synaptic current of the predetermined amplitude is output for a longer period of time than when the excess component is small.

[0100] As described above, the neural network device 10 according to the second embodiment can reflect the excess component at firing in the first neuron circuit 32 in the amount of synaptic current output from the first synaptic circuit 30 downstream of the first neuron circuit 32. As a result, the neural network device 10 according to the second embodiment can realize a highly accurate spiking-type neural network with less loss of transmitted information and less information transmission delay.

[0101] (Third embodiment) Next, the neural network device 10 according to the third embodiment will be described. The neural network device 10 according to the third embodiment has substantially the same functions and configuration as the first embodiment, so components with substantially the same functions and configurations are given the same reference numerals as in the first embodiment, and detailed descriptions will be omitted except for differences.

[0102] Figure 12 is a diagram showing the connection relationships of the peripheral circuits of the first neuron circuit 32 in the third embodiment.

[0103] The first neuron circuit 32 according to the third embodiment outputs an intensity control signal along with a spike signal. The intensity control signal is similar to the control signal according to the third embodiment, with respect to the membrane potential V mem Threshold potential V th This represents the comparison voltage based on the excess component that exceeds a certain threshold. The intensity control signal may be an analog voltage or digital data.

[0104] The first neuron circuit 32 according to the third embodiment changes the time width of the voltage pulse of the spike signal according to the control signal. For example, the first neuron circuit 32 changes the membrane potential V mem Threshold potential V th The larger the excess component exceeding the threshold, the longer the duration of the voltage pulse in the spike signal.

[0105] Furthermore, each of the at least one first synaptic circuits 30 that acquires a spike signal from the first neuron circuit 32 acquires an intensity control signal from the first neuron circuit 32 instead of a control signal.

[0106] Each of at least one first synaptic circuit 30, upon receiving a spike signal from the first neuron circuit 32, outputs a synaptic current corresponding to the time width of the voltage pulse of the acquired spike signal, a control signal, and a current amount corresponding to the synaptic weight.

[0107] For example, each of the at least one first synaptic circuits 30 will increase the amplitude of the output synaptic current as the comparison voltage represented by the control signal increases, provided that the time width of the voltage pulse of the spike signal is the same. Also, for example, each of the at least one first synaptic circuits 30 will increase the duration for which the synaptic current flows as the time width of the voltage pulse of the spike signal increases, provided that the control signal is the same. As a result, each of the at least one first synaptic circuits 30 can increase the amount of output synaptic current as the excess component at the time of firing of the first neuron circuit 32 increases.

[0108] Figure 13 shows the configuration of the first neuron circuit 32 according to the third embodiment, along with the first synaptic circuit 30 connected to the first neuron circuit 32.

[0109] The control signal output circuit 48 according to the third embodiment outputs an intensity control signal and a time control signal as control signals.

[0110] Each of the intensity control signal and the time control signal represents a comparison voltage. The intensity control signal and the time control signal may be signals of different formats. For example, one of the intensity control signal and the time control signal may be an analog voltage and the other may be digital data. Alternatively, the intensity control signal and the time control signal may be signals of the same format but different values. Furthermore, the intensity control signal and the time control signal may be signals of the same format and value. In this case, the intensity control signal and the time control signal may be output from the same part of the control signal output circuit 48.

[0111] The control signal output circuit 48 provides an intensity control signal to each of the at least one first synaptic circuits 30 connected to the first neuron circuit 32. The control signal output circuit 48 also provides a time control signal to the firing circuit 44.

[0112] The ignition circuit 44 outputs a spike signal, which is a voltage pulse with a time width corresponding to the time control signal, when the judgment signal obtained from the comparison circuit 42 changes from a second value to a first value. For example, the ignition circuit 44 increases the time width of the voltage pulse of the spike signal as the comparison voltage represented by the time control signal increases.

[0113] The first neuronal circuit 32 with this configuration has a membrane potential V mem Threshold potential V th It can output a spike signal with a pulse width corresponding to the excess component that exceeds a certain threshold, as well as an intensity control signal representing the excess component.

[0114] Figure 14 shows the membrane potential V in the third embodiment. mem The figure also shows an example of a synaptic current waveform.

[0115] Synaptic circuits 20 other than the first synaptic circuit 30 do not receive intensity control signals and receive spike signals with a fixed voltage pulse duration. Therefore, synaptic circuits 20 other than the first synaptic circuit 30 receive a membrane potential V mem Threshold potential V th Regardless of the magnitude of the excess component exceeding the threshold, neither the amplitude of the synaptic current nor the duration for generating the synaptic current remains unchanged.

[0116] In response to this, the first neuron circuit 32 according to the third embodiment provides an intensity control signal to the subsequent first synaptic circuit 30, corresponding to the excess component at the time of firing of the spike signal. Upon receiving the intensity control signal, the first synaptic circuit 30 changes the amplitude of the synaptic current according to the magnitude of the intensity control signal. For example, the first synaptic circuit 30 controls the membrane potential V mem Threshold potential V th When the excess component exceeding the threshold is large, it outputs a synaptic current with a larger amplitude than when the excess component is small.

[0117] Furthermore, the first neuron circuit 32 according to the third embodiment supplies the subsequent first synaptic circuit 30 with spike signals of voltage pulses of different time widths depending on the excess component. The first synaptic circuit 30 then changes the duration for which it generates synaptic current according to the pulse width of the spike signal. For example, the first synaptic circuit 30 uses the membrane potential V mem Threshold potential V th When the excess component exceeding the threshold is large, the synaptic current is output for a longer period of time than when the excess component is small.

[0118] As described above, the neural network device 10 according to the third embodiment can reflect the excess component at firing in the first neuron circuit 32 in the amount of synaptic current output from the first synaptic circuit 30 downstream of the first neuron circuit 32. As a result, the neural network device 10 according to the third embodiment can realize a highly accurate spiking-type neural network with less loss of transmitted information and less information transmission delay.

[0119] (Variation of the first neuron circuit 32) Next, modifications of the first neuron circuit 32 according to the first to third embodiments will be described. While the modifications described below are modifications of the first embodiment, similar modifications may be applied to the second and third embodiments as well.

[0120] Figure 15 shows the configuration of the first neuron circuit 32 according to the first modified example, along with the first synaptic circuit 30 connected to the first neuron circuit 32.

[0121] The control signal output circuit 48 in the first modified example has a membrane potential V mem This is obtained as the comparison voltage. Threshold potential V th This is a fixed potential. Therefore, the membrane potential V mem is the membrane potential V mem Threshold potential V th It is proportional to the excess component that exceeds the membrane potential V. Therefore, the control signal output circuit 48 is proportional to the membrane potential V. memEven when the voltage is obtained as the comparison voltage, the membrane potential V mem Threshold potential V th It is possible to generate a control signal that represents a proportional voltage based on the excess component that exceeds a certain threshold.

[0122] Figure 16 shows the configuration of the first neuron circuit 32 according to the second modified example, along with the first synaptic circuit 30 connected to the first neuron circuit 32.

[0123] The control signal output circuit 48 in the second modified example has a membrane potential V mem From the threshold potential V th The difference voltage obtained by subtracting is acquired from the comparator circuit 42 as the comparison voltage. The comparator included in the comparator circuit 42 internally generates a voltage that represents the difference between the voltage at the non-inverting input terminal and the voltage at the inverting input terminal. The voltage that represents the difference between the voltage at the non-inverting input terminal and the voltage at the inverting input terminal is the membrane potential V mem From the threshold potential V th This is the difference voltage after subtracting the voltage at the non-inverting input terminal. Therefore, the control signal output circuit 48 obtains a voltage from inside the comparator included in the comparison circuit 42 that represents the difference between the voltage at the non-inverting input terminal and the voltage at the inverting input terminal, thereby determining the membrane potential V mem Threshold potential V th It is possible to generate a control signal that represents a proportional voltage based on the excess component that exceeds a certain threshold.

[0124] Figure 17 shows the configuration of the first neuron circuit 32 according to the third modified example, along with the first synaptic circuit 30 connected to the first neuron circuit 32.

[0125] The first neuron circuit 32 according to the third modified example further includes a control circuit 92. The control circuit 92 controls the timing at which the charge control circuit 46 charges and discharges the charge stored in the charge holding circuit 40. The control circuit 92 according to this modified example controls the release timing at which the charge stored in the charge holding circuit 40 is released in accordance with the timing at which a spike signal is output from the firing circuit 44.

[0126] For example, the control circuit 92 in this modified example turns on the charge control circuit 46 to start releasing charge after a predetermined time has elapsed since the ignition circuit 44 outputs a spike signal, and turns off the charge control circuit 46 to stop releasing charge after a predetermined time has elapsed since the start of charge release. As a result, the first neuron circuit 32 in the third modified example reacts to the membrane potential V generated from the charge holding circuit 40 after the ignition circuit 44 outputs a spike signal. mem It can be returned to its initial potential.

[0127] Figure 18 shows the configuration of the first neuron circuit 32 according to the fourth modified example, along with the first synaptic circuit 30 connected to the first neuron circuit 32.

[0128] The first neuron circuit 32 according to the fourth modification also further includes a control circuit 92, similar to the third modification. The control circuit 92 according to this modification controls the timing of charging and discharging the charge stored in the charge holding circuit 40 in accordance with the timing at which at least one first synaptic circuit 30 connected to the first neuron circuit 32 acquires a spike signal.

[0129] For example, the control circuit 92 in this modified example receives a feedback signal indicating the timing of spike signal acquisition from at least one of the first synaptic circuits 30. Then, for example, the control circuit 92 in this modified example turns on the charge control circuit 46 to start releasing charge after a predetermined time after receiving the feedback signal, and turns off the charge control circuit 46 to stop releasing charge after a predetermined time after the timing of the start of charge release. As a result, the first neuron circuit 32 in the fourth modified example receives the membrane potential V generated from the charge holding circuit 40 after at least one first synaptic circuit 30 has acquired a spike signal. mem It can be returned to its initial potential.

[0130] Figure 19 shows the configuration of the first neuron circuit 32 according to the fifth modified example, along with the first synaptic circuit 30 connected to the first neuron circuit 32.

[0131] The first neuron circuit 32 according to the fifth modification also further includes a control circuit 92, similar to the third modification. The control circuit 92 according to this modification controls the timing of charging and discharging the charge stored in the charge holding circuit 40 according to the timing when the determination signal output from the comparison circuit 42 changes from a second value to a first value. For example, in the control circuit 92 according to this modification, when the comparison circuit 42 is at the membrane potential V mem The threshold potential V th From a state where it is determined that it is not larger, the membrane potential V mem The threshold potential V th The timing of releasing the charge accumulated in the charge holding circuit 40 is controlled according to the timing at which the state changes to one that is determined to be larger.

[0132] For example, the control circuit 92 in this modified example receives a determination signal from the comparison circuit 42. Then, for example, the control circuit 92 in this modified example turns on the charge control circuit 46 to start releasing charge after a predetermined time has elapsed since the determination signal changed from the second value to the first value, and turns off the charge control circuit 46 to stop releasing charge after a predetermined time has elapsed since the start of charge release. As a result, the first neuron circuit 32 in the fifth modified example determines that the comparison circuit 42 has a membrane potential V mem The threshold potential V th After it is determined that the value has increased, the membrane potential V is generated from the charge holding circuit 40. mem It can be returned to its initial potential.

[0133] Figure 20 shows the configuration of the first neuron circuit 32 according to the sixth modified example, along with the first synaptic circuit 30 connected to the first neuron circuit 32.

[0134] The first neuron circuit 32 in the sixth modified example also further includes a control circuit 92, similar to the third modified example. Furthermore, the control circuit 92 in this modified example receives a control signal from the control signal output circuit 48. Then, the control circuit 92 in this modified example changes the time for the charge control circuit 46 to charge and discharge the charge stored in the charge holding circuit 40 according to the control signal. For example, the control circuit 92 in this modified example controls the membrane potential Vmem Threshold potential V th The larger the excess component that exceeds the limit, the longer the release time. For example, the control circuit 92 in this modified example controls the membrane potential V mem The larger the value, or the membrane potential V mem From the threshold potential V th The larger the difference after subtracting [the value], the longer the release time.

[0135] For example, the control circuit 92 in this modified example turns on the charge control circuit 46 and starts releasing charge after a predetermined time has elapsed since the ignition circuit 44 outputs a spike signal. Then, the control circuit 92 in this modified example turns off the charge control circuit 46 and stops releasing charge after a time elapsed according to the control signal from the moment the release of charge started. As a result, the first neuron circuit 32 in the sixth modified example can adjust the time for which the charge control circuit 46 releases charge according to the amount of charge accumulated in the charge holding circuit 40. Therefore, the first neuron circuit 32 in the sixth modified example can control the membrane potential V generated from the charge holding circuit 40. mem By returning to the initial potential, charge accumulation can be initiated in a shorter time.

[0136] Furthermore, in the control circuits 92 of the fourth and fifth modified examples, similar to the sixth modified example, the release time for which the charge control circuit 46 releases the charge accumulated in the charge holding circuit 40 may be changed in accordance with the control signal.

[0137] Figure 21 shows the configuration of the first neuron circuit 32 according to the seventh modified example, along with the first synaptic circuit 30 connected to the first neuron circuit 32.

[0138] The first neuron circuit 32 according to the seventh modified example further includes a threshold changing circuit 94. The threshold changing circuit 94 adjusts the threshold potential V according to the comparison voltage. th Change it.

[0139] For example, the threshold changing circuit 94 uses the membrane potential V as the comparison voltage. mem From the threshold potential V thThe differential voltage obtained by subtracting is then set by the threshold changing circuit 94. mem From the threshold potential V th The difference voltage obtained by subtracting is greater than 0, i.e., the membrane potential V mem The threshold potential V th If it is greater, the threshold potential V th Lower the membrane potential V mem The threshold potential V th If it is smaller, the threshold potential V th The potential may be returned to a preset potential. Also, the threshold changing circuit 94 controls the membrane potential V mem From the threshold potential V th The voltage obtained by subtracting the difference voltage is amplified and the threshold potential V th It may also be applied to.

[0140] In this seventh modified example, the first neuronal circuit 32 has a membrane potential V mem and threshold potential V th When the difference between the two is small, it is possible to prevent a chattering state in which the value of the judgment signal switches back and forth in a short time. The threshold change circuit 94 acquires a control signal from the control signal output circuit 48 and sets the threshold potential V based on the control signal. th You may change it.

[0141] Figure 22 shows the configuration of the first neuron circuit 32 according to the eighth modified example, together with the first synaptic circuit 30 connected to the first neuron circuit 32.

[0142] The first neuron circuit 32 according to the eighth modified example further includes a cutoff switch 96. The cutoff switch 96 switches between an ON state, which supplies synaptic current from the input circuit 38 to the charge holding circuit 40, and an OFF state, which cuts off the supply of synaptic current from the input circuit 38 to the charge holding circuit 40. The cutoff switch 96 is implemented, for example, by a MOSFET. Alternatively, the cutoff switch 96 may be implemented by a current mirror circuit, a potentiometer, a resistive memory element, a phase-change memory element, a magnetoresistive memory, or a flip-flop circuit.

[0143] In the eighth modified example, the cutoff switch 96 acquires a determination signal from the comparison circuit 42. The cutoff switch 96 switches from the ON state to the OFF state in response to the determination signal changing from the second value to the first value. Then, after a predetermined time has elapsed since the cutoff switch 96 switched from the ON state to the OFF state, it switches back from the OFF state to the ON state.

[0144] As a result, the first neuron circuit 32 in the eighth modified example has a membrane potential V mem The threshold potential V th After exceeding a certain threshold, the supply of further synaptic current to the charge retention circuit 40 can be stopped. In addition, the first neuron circuit 32 according to the eighth modified example can be configured to have a refractory period during which the first neuron circuit 32 does not respond after firing a spike signal.

[0145] Figure 23 shows the configuration of the first neuron circuit 32 according to the ninth modified example, along with the first synaptic circuit 30 connected to the first neuron circuit 32.

[0146] The first neuron circuit 32 in the ninth modification also further includes a cutoff switch 96, similar to the eighth modification. In the ninth modification, the cutoff switch 96 acquires a comparison voltage or a control signal. The cutoff switch 96 switches from the ON state to the OFF state when the comparison signal exceeds a predetermined voltage. Then, after a predetermined time has elapsed from the moment the cutoff switch 96 switches from the OFF state to the ON state.

[0147] As a result, the first neuron circuit 32 according to the ninth modified example has a membrane potential V mem The threshold potential V th After exceeding this value, the supply of further synaptic current to the charge retention circuit 40 is stopped, and the membrane potential V mem Threshold potential V th This makes it possible to prevent an increase in excess components that exceed a certain threshold. Furthermore, the first neuron circuit 32 according to the ninth modified example can be configured to have a refractory period during which the first neuron circuit 32 does not respond after firing a spike signal.

[0148] Figure 24 shows the configuration of the first neuron circuit 32 according to the 10th modified example, along with the first synaptic circuit 30 connected to the first neuron circuit 32.

[0149] The first neuron circuit 32 according to the 10th modified example further includes a leakage current circuit 98. The leakage current circuit 98 reduces the charge accumulated in the charge retention circuit 40 over time. The leakage current circuit 98 is connected in parallel between two terminals of the charge retention circuit 40, and the membrane potential V of the charge retention circuit 40 mem A leakage current flows from the generation terminal to the ground terminal, causing the charge accumulated in the charge retention circuit 40 to leak out. Therefore, when no synaptic current is supplied, the membrane potential V mem It will be lowered over time.

[0150] For example, the leakage current circuit 98 is a resistive element mounted on the semiconductor device. The resistive element is connected between the first terminal 34 and the ground terminal. The magnitude of the leakage current flowing through the leakage current circuit 98 is determined by the resistance value of the resistive element and the film potential V generated from the charge holding circuit 40. mem This is determined by the following: The resistive element has a relatively large resistance value, for example, 100 MΩ or more, and releases the charge accumulated in the charge holding circuit 40 over a sufficiently long period of time. Alternatively, the resistive element may be composed of a MOSFET. In this case, the leakage current value is determined by the gate voltage of the MOSFET.

[0151] The first neuron circuit 32 according to this 10th modification can operate in a manner that mimics a neuron following a leakage model. The first neuron circuit 32 according to this 10th modification may further include the cutoff switch 96 shown in the 8th or 9th modification. In this case, the cutoff switch 96 may, when in the off state, cut off the supply of synaptic current to the leakage current circuit 98. Alternatively, the cutoff switch 96 may, when in the off state, cut off the supply of synaptic current to the charge holding circuit 40, but continue to supply synaptic current to the leakage current circuit 98. Furthermore, the first neuron circuit 32 may have a cutoff switch 96 between the leakage current circuit 98 and the charge holding circuit 40, and when in the off state, cut off the supply from the charge holding circuit 40 to the leakage current circuit 98.

[0152] (Example of the hierarchical configuration of neural network device 10) Next, we will describe examples of the hierarchical configuration of the neural network device 10. The neural network device 10 may have the configurations shown in each of the following hierarchical configuration examples.

[0153] Figure 25 shows a neural network device 10 in the first example of a hierarchical configuration.

[0154] For example, in the neural network device 10, each of the multiple neuron circuits 22 contained in any h layers of layer 12-h (where h is an integer greater than or equal to 1) includes forward connections via synaptic circuits 20 with all of the multiple neuron circuits 22 contained in the next (h+1) layers of layer 12-(h+1) following the h layers of layer 12-h. Thus, the neural network device 10 may include layers 12 that perform forward propagation with full connectivity.

[0155] Furthermore, in the neural network device 10, all of the multiple neuron circuits 22 may have the same configuration as the first neuron circuit 32, or some of the multiple neuron circuits 22 may have the same configuration as the first neuron circuit 32.

[0156] Figure 26 shows a neural network device 10 in a second example of a hierarchical configuration.

[0157] For example, in the neural network device 10, some of the multiple neuronal circuits 22 included in the h-stage layer 12-h do not necessarily have a synaptic connection 20 with the neuronal circuits 22 included in the (h+1)-stage layer 12-(h+1).

[0158] Figure 27 shows a neural network device 10 in the third example of a hierarchical configuration.

[0159] For example, in the neural network device 10, some of the multiple neuronal circuits 22 included in the h-stage layer 12-h may be connected to a synaptic circuit 20 that propagates synaptic currents backward to its own input.

[0160] Figure 28 shows a neural network device 10 in the fourth hierarchical configuration example.

[0161] For example, in the neural network device 10, some of the multiple neuronal circuits 22 included in the h-stage layer 12-h may be connected via synaptic circuits 20 to the neuronal circuits 22 included in the (hj)-stage layer 12-(hj) (where j is an integer of 1 or more) preceding the h-stage layer 12-h.

[0162] Figure 29 shows a neural network device 10 in the fifth layer configuration example.

[0163] For example, the neural network device 10 may include multiple partial neural networks 110 (110-1, 110-2). In this case, each of the multiple partial neural networks 110 (110-1, 110-2) has one or more neuron circuits 22 in its final stage. Furthermore, all of the multiple partial neural networks 110 (110-1, 110-2), except for the first partial neural network 110-1, are connected to the other partial neural networks 110 via one or more synaptic circuits 20.

[0164] Furthermore, each of the multiple partial neural networks 110 (110-1, 110-2) may have one of the configurations shown in the first to fourth hierarchical configuration examples.

[0165] Figure 30 shows the connection relationships around the first neuron circuit 32 in the sixth example of a hierarchical configuration. For example, the first neuron circuit 32 included in the neural network device 10 may be a first synaptic circuit 30 in which all of the multiple synaptic circuits 20 connected to the subsequent stages are first synaptic circuits 30.

[0166] Figure 31 shows the connection relationships around the first neuron circuit 32 in the seventh example of a hierarchical configuration. For example, the first neuron circuit 32 included in the neural network device 10 may be part of a plurality of synaptic circuits 20 connected to the next stage, with the remaining part being synaptic circuits 20.

[0167] Figure 32 shows the connectivity relationships around the first neuron circuit 32 in the eighth example of a hierarchical configuration. For example, the first neuron circuit 32 included in the neural network device 10 may be connected to a synaptic circuit 20 that propagates synaptic currents backward to its own input. In this case, the synaptic circuit 20 that propagates synaptic currents backward to its own input may be the first synaptic circuit 30.

[0168] Figure 33 shows the connection relationships around the first neuron circuit 32 in the ninth example of hierarchical configuration. Furthermore, for example, if the first neuron circuit 32 included in the neural network device 10 is connected to a synaptic circuit 20 that propagates synaptic currents backward to its own input, the synaptic circuit 20 that propagates synaptic currents backward to its own input does not have to be the first synaptic circuit 30.

[0169] Although embodiments of the present invention have been described above, these embodiments are presented as examples only and are not intended to limit the scope of the invention. These novel embodiments can be implemented in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These novel embodiments and their variations are included in the scope and spirit of the invention, as well as in the claims of the invention and its equivalents.

[0170] Furthermore, the following additional information is disclosed regarding the above embodiments. <Note>

[0171] (Note 1) Multiple synaptic circuits, each with its own assigned synaptic weight, Each of the multiple neuronal circuits outputs a spike signal, which is a voltage pulse. Equipped with, Each of the plurality of synaptic circuits acquires the spike signal output from any one of the plurality of neuronal circuits, and when it acquires the spike signal, it outputs a synaptic current of a current amount corresponding to the set synaptic weight. The first neuron circuit among the aforementioned plurality of neuron circuits is An input circuit that acquires the synaptic current from each of at least one of the plurality of synaptic circuits, A charge-holding circuit that accumulates a charge corresponding to the synaptic current acquired by the input circuit and generates a membrane potential corresponding to the accumulated charge, A comparison circuit generates a determination signal which becomes a first value when the sign of the difference voltage between a preset threshold potential and the membrane potential is a first sign, and becomes a second value when the sign of the difference voltage is a second sign different from the first sign. A firing circuit that outputs the spike signal when the determination signal changes from the second value to the first value, A charge control circuit that, after outputting the spike signal, charges and discharges the charge stored in the charge holding circuit so that the sign of the differential voltage becomes the second sign, A control signal output circuit that outputs a control signal representing a comparison voltage based on the excess component that exceeds the threshold potential at the membrane potential, It has, Each of the multiple synaptic circuits, at least one first synaptic circuit that acquires the spike signal from the first neuron circuit, When the spike signal is obtained from the first neuron circuit, the control signal and the synaptic current corresponding to the synaptic weight are output. A neural network device.

[0172] (Note 2) Multiple synaptic circuits, each with its own assigned synaptic weight, Each of the multiple neuronal circuits outputs a spike signal, which is a voltage pulse. Equipped with, Each of the plurality of synaptic circuits acquires the spike signal output from any one of the plurality of neuronal circuits, and when it acquires the spike signal, it outputs a synaptic current of a current amount corresponding to the set synaptic weight. The first neuron circuit among the aforementioned plurality of neuron circuits is An input circuit that acquires the synaptic current from each of at least one of the plurality of synaptic circuits, A charge-holding circuit that accumulates a charge corresponding to the synaptic current acquired by the input circuit and generates a membrane potential corresponding to the accumulated charge, A comparison circuit generates a determination signal which becomes a first value when the sign of the difference voltage between a preset threshold potential and the membrane potential is a first sign, and becomes a second value when the sign of the difference voltage is a second sign different from the first sign. A firing circuit that outputs the spike signal when the determination signal changes from the second value to the first value, A charge control circuit that, after outputting the spike signal, charges and discharges the charge stored in the charge holding circuit so that the sign of the differential voltage becomes the second sign, A control signal output circuit that outputs a control signal representing a comparison voltage based on an excess component exceeding the threshold potential in the membrane potential. It has The firing circuit outputs the spike signal of a voltage pulse with a time width corresponding to the control signal. Each of at least one first synapse circuit among the plurality of synapse circuits that acquires the spike signal from the first neuron circuit When acquiring the spike signal from the first neuron circuit, outputs the synaptic current of a current amount corresponding to the time width of the voltage pulse of the spike signal and the synaptic weight. Neural network device.

[0173] (Appendix 3) The control signal output circuit outputs an intensity control signal and a time control signal as the control signal. The firing circuit changes the time width for outputting the spike signal according to the time control signal. Each of the at least one first synapse circuit When acquiring the spike signal from the first neuron circuit, outputs the synaptic current of a current amount corresponding to the time width of the spike signal, the intensity control signal, and the synaptic weight. The neural network device according to Appendix 2.

[0174] (Appendix 4) Each of the at least one first synapse circuit When acquiring the spike signal from the first neuron circuit, outputs the synaptic current of a current amount corresponding to the time width of the spike signal, the control signal, and the synaptic weight. The neural network device according to Appendix 2.

[0175] (Appendix 5) The control signal output circuit acquires the membrane potential as the comparison voltage. The neural network device according to any one of Appendices 1 to 4.

[0176] (Note 6) The control signal output circuit obtains the difference voltage obtained by subtracting the threshold potential from the membrane potential from the comparison circuit as the comparison voltage. A neural network device as described in any one of the appendices 1 through 4.

[0177] (Note 7) The first neuron circuit further includes a control circuit that controls the timing at which the charge control circuit charges and discharges the charge stored in the charge holding circuit in accordance with the timing at which the spike signal is output from the firing circuit. A neural network device as described in any one of the appendices 1 through 6.

[0178] (Note 8) The first neuron circuit further includes a control circuit that controls the timing at which the charge control circuit charges and discharges the charge stored in the charge holding circuit, in accordance with the timing at which the at least one first synaptic circuit acquires the spike signal. A neural network device as described in any one of the appendices 1 through 6.

[0179] (Note 9) The first neuron circuit further includes a control circuit that controls the timing at which the charge control circuit charges and discharges the charge stored in the charge holding circuit in accordance with the timing at which the determination signal changes from the second value to the first value. A neural network device as described in any one of the appendices 1 through 6.

[0180] (Note 10) The control circuit changes the time it takes for the charge control circuit to charge and discharge the charge stored in the charge holding circuit, in accordance with the control signal. A neural network device as described in any one of the appendices 7 through 9.

[0181] (Note 11) The first neuron circuit further includes a threshold changing circuit that changes the threshold potential according to the comparison voltage. A neural network device as described in any one of the appendices 1 through 10.

[0182] (Note 12) The first neuron circuit further includes a cutoff switch that switches between an ON state, which supplies the synaptic current from the input circuit to the charge holding circuit, and an OFF state, which cuts off the supply of the synaptic current from the input circuit to the charge holding circuit. The cutoff switch switches from the ON state to the OFF state in response to the determination signal changing from the second value to the first value. A neural network device as described in any one of the appendices 1 through 11.

[0183] (Note 13) The first neuron circuit further includes a cutoff switch that switches between an ON state, which supplies the synaptic current from the input circuit to the charge holding circuit, and an OFF state, which cuts off the supply of the synaptic current from the input circuit to the charge holding circuit. The cutoff switch switches from the ON state to the OFF state when the comparison voltage exceeds a predetermined value. A neural network device as described in any one of the appendices 1 through 11.

[0184] (Note 14) The first neuron circuit further includes a leakage current circuit that reduces the charge accumulated in the charge holding circuit over time. A neural network device as described in any one of the appendices 1 through 13.

[0185] (Note 15) A signal processing method in a neural network device, The neural network device is Multiple synaptic circuits, each with its own assigned synaptic weight, A plurality of neuron circuits each outputting a spike signal that is a voltage pulse, comprising: Each of the plurality of synaptic circuits acquires the spike signal output from any one of the plurality of neuron circuits, and when the spike signal is acquired, outputs a synaptic current having a current amount corresponding to the set synaptic weight, A first neuron circuit among the plurality of neuron circuits, An input circuit that acquires the synaptic current from each of at least one of the plurality of synaptic circuits, A charge holding circuit that accumulates a charge corresponding to the synaptic current acquired by the input circuit and generates a membrane potential corresponding to the accumulated charge, A comparison circuit that generates a determination signal that becomes a first value when the sign of the difference voltage between a preset threshold potential and the membrane potential is a first sign, and becomes a second value when the sign of the difference voltage is a second sign different from the first sign, A firing circuit that outputs the spike signal when the determination signal changes from the second value to the first value, A charge control circuit that charges and discharges the charge accumulated in the charge holding circuit so that the sign of the difference voltage becomes the second sign after the spike signal is output, A control signal output circuit that outputs a control signal representing a comparison voltage based on an excess component exceeding the threshold potential in the membrane potential, having: Each of at least one first synaptic circuit that acquires the spike signal from the first neuron circuit among the plurality of synaptic circuits, when acquiring the spike signal from the first neuron circuit, outputs the synaptic current having a current amount corresponding to the control signal and the synaptic weight Signal processing method.

[0186] (Appendix 16) A signal processing method in a neural network device, The neural network device is Multiple synaptic circuits, each with its own assigned synaptic weight, Each of the multiple neuronal circuits outputs a spike signal, which is a voltage pulse. Equipped with, Each of the plurality of synaptic circuits acquires the spike signal output from any one of the plurality of neuronal circuits, and when it acquires the spike signal, it outputs a synaptic current of a current amount corresponding to the set synaptic weight. The first neuron circuit among the aforementioned plurality of neuron circuits is An input circuit that acquires the synaptic current from each of at least one of the plurality of synaptic circuits, A charge-holding circuit that accumulates a charge corresponding to the synaptic current acquired by the input circuit and generates a membrane potential corresponding to the accumulated charge, A comparison circuit generates a determination signal which becomes a first value when the sign of the difference voltage between a preset threshold potential and the membrane potential is a first sign, and becomes a second value when the sign of the difference voltage is a second sign different from the first sign. A firing circuit that outputs the spike signal when the determination signal changes from the second value to the first value, A charge control circuit that, after outputting the spike signal, charges and discharges the charge stored in the charge holding circuit so that the sign of the differential voltage becomes the second sign, A control signal output circuit that outputs a control signal representing a comparison voltage based on the excess component that exceeds the threshold potential at the membrane potential, It has, The ignition circuit outputs the spike signal, which is a voltage pulse with a time width corresponding to the control signal. Each of the multiple synaptic circuits, at least one first synaptic circuit that acquires the spike signal from the first neuron circuit, When the spike signal is obtained from the first neuron circuit, the synaptic current is output, which corresponds to the time width of the voltage pulse of the spike signal and the current amount corresponding to the synaptic weight. Signal processing method. [Explanation of Symbols]

[0187] 10 Neural Network Devices 20 Synaptic Circuits 22 Neuron Circuits 30. First synaptic circuit 32. First Neuron Circuit 38 Input Circuits 40 Charge retention circuit 42 Comparison circuit 44 Firing Circuit 46 Charge Control Circuit 48 Control signal output circuit 92 Control circuits 94. Threshold change circuit 96. Disconnection switch 98 Leakage Current Circuit

Claims

1. Multiple synaptic circuits, each with its own assigned synaptic weight, Each of the multiple neuronal circuits outputs a spike signal, which is a voltage pulse. Equipped with, Each of the plurality of synaptic circuits acquires the spike signal output from any one of the plurality of neuronal circuits, and when it acquires the spike signal, it outputs a synaptic current of a current amount corresponding to the set synaptic weight. The first neuron circuit among the plurality of neuron circuits is An input circuit that acquires the synaptic current from each of at least one of the plurality of synaptic circuits, A charge-holding circuit that accumulates a charge corresponding to the synaptic current acquired by the input circuit and generates a membrane potential corresponding to the accumulated charge, A comparison circuit generates a determination signal which becomes a first value when the sign of the difference voltage between a preset threshold potential and the membrane potential is a first sign, and becomes a second value when the sign of the difference voltage is a second sign different from the first sign. A firing circuit that outputs the spike signal when the determination signal changes from the second value to the first value, A charge control circuit that, after outputting the spike signal, charges and discharges the charge stored in the charge holding circuit so that the sign of the differential voltage becomes the second sign, A control signal output circuit that outputs a control signal representing a comparison voltage based on the excess component that exceeds the threshold potential at the membrane potential, It has, Of the plurality of synaptic circuits, each of the at least one first synaptic circuit that acquires the spike signal from the first neuron circuit is: When the spike signal is obtained from the first neuron circuit, the control signal and the synaptic current corresponding to the synaptic weight are output. A neural network device.

2. Multiple synaptic circuits, each with its own assigned synaptic weight, Each of the multiple neuronal circuits outputs a spike signal, which is a voltage pulse. Equipped with, Each of the plurality of synaptic circuits acquires the spike signal output from any one of the plurality of neuronal circuits, and when it acquires the spike signal, it outputs a synaptic current of a current amount corresponding to the set synaptic weight. The first neuron circuit among the plurality of neuron circuits is An input circuit that acquires the synaptic current from each of at least one of the plurality of synaptic circuits, A charge-holding circuit that accumulates a charge corresponding to the synaptic current acquired by the input circuit and generates a membrane potential corresponding to the accumulated charge, A comparison circuit generates a determination signal which becomes a first value when the sign of the difference voltage between a preset threshold potential and the membrane potential is a first sign, and becomes a second value when the sign of the difference voltage is a second sign different from the first sign. A firing circuit that outputs the spike signal when the determination signal changes from the second value to the first value, A charge control circuit that, after outputting the spike signal, charges and discharges the charge stored in the charge holding circuit so that the sign of the differential voltage becomes the second sign, A control signal output circuit that outputs a control signal representing a comparison voltage based on the excess component that exceeds the threshold potential at the membrane potential, It has, The ignition circuit outputs the spike signal, which is a voltage pulse with a time width corresponding to the control signal. Of the plurality of synaptic circuits, each of the at least one first synaptic circuit that acquires the spike signal from the first neuron circuit is: When the spike signal is obtained from the first neuron circuit, the synaptic current is output, which corresponds to the time width of the voltage pulse of the spike signal and the current amount corresponding to the synaptic weight. A neural network device.

3. The control signal output circuit outputs an intensity control signal and a time control signal as the control signals. The ignition circuit changes the time width for outputting the spike signal according to the time control signal, Each of the at least one first synaptic circuit is, When the spike signal is acquired from the first neuron circuit, the synaptic current is output, which corresponds to the time width of the spike signal, the intensity control signal, and the current amount corresponding to the synaptic weight. The neural network device according to claim 2.

4. Each of the at least one first synaptic circuit is, When the spike signal is obtained from the first neuron circuit, the synaptic current is output, which corresponds to the time width of the spike signal, the control signal, and the synaptic weight. The neural network device according to claim 2.

5. The control signal output circuit acquires the membrane potential as the comparison voltage. A neural network device according to any one of claims 1 to 4.

6. The control signal output circuit obtains the difference voltage obtained by subtracting the threshold potential from the membrane potential from the comparison circuit as the comparison voltage. A neural network device according to any one of claims 1 to 4.

7. The first neuron circuit further includes a control circuit that controls the timing at which the charge control circuit charges and discharges the charge stored in the charge holding circuit, in accordance with the timing at which the spike signal is output from the firing circuit. A neural network device according to any one of claims 1 to 4.

8. The first neuron circuit further includes a control circuit that controls the timing at which the charge control circuit charges and discharges the charge stored in the charge holding circuit in accordance with the timing at which the at least one first synaptic circuit acquires the spike signal. A neural network device according to any one of claims 1 to 4.

9. The first neuron circuit further includes a control circuit that controls the timing at which the charge control circuit charges and discharges the charge stored in the charge holding circuit in accordance with the timing at which the determination signal changes from the second value to the first value. A neural network device according to any one of claims 1 to 4.

10. The control circuit changes the time it takes for the charge control circuit to charge and discharge the charge stored in the charge holding circuit, in accordance with the control signal. The neural network device according to claim 7.

11. The first neuron circuit further includes a threshold changing circuit that changes the threshold potential according to the comparison voltage. A neural network device according to any one of claims 1 to 4.

12. The first neuron circuit further includes a cutoff switch that switches between an ON state, which supplies the synaptic current from the input circuit to the charge holding circuit, and an OFF state, which cuts off the supply of the synaptic current from the input circuit to the charge holding circuit. The cutoff switch switches from the ON state to the OFF state in response to the determination signal changing from the second value to the first value. A neural network device according to any one of claims 1 to 4.

13. The first neuron circuit further includes a cutoff switch that switches between an ON state, which supplies the synaptic current from the input circuit to the charge holding circuit, and an OFF state, which cuts off the supply of the synaptic current from the input circuit to the charge holding circuit. The cutoff switch switches from the ON state to the OFF state when the comparison voltage exceeds a predetermined value. A neural network device according to any one of claims 1 to 4.

14. The first neuron circuit further includes a leakage current circuit that reduces the charge accumulated in the charge holding circuit over time. A neural network device according to any one of claims 1 to 4.

15. A signal processing method in a neural network device, The neural network device is Multiple synaptic circuits, each with its own assigned synaptic weight, Each of the multiple neuronal circuits outputs a spike signal, which is a voltage pulse. Equipped with, Each of the plurality of synaptic circuits acquires the spike signal output from any one of the plurality of neuronal circuits, and when it acquires the spike signal, it outputs a synaptic current of a current amount corresponding to the set synaptic weight. The first neuron circuit among the plurality of neuron circuits is An input circuit that acquires the synaptic current from each of at least one of the plurality of synaptic circuits, A charge-holding circuit that accumulates a charge corresponding to the synaptic current acquired by the input circuit and generates a membrane potential corresponding to the accumulated charge, A comparison circuit generates a determination signal which becomes a first value when the sign of the difference voltage between a preset threshold potential and the membrane potential is a first sign, and becomes a second value when the sign of the difference voltage is a second sign different from the first sign. A firing circuit that outputs the spike signal when the determination signal changes from the second value to the first value, A charge control circuit that, after outputting the spike signal, charges and discharges the charge stored in the charge holding circuit so that the sign of the differential voltage becomes the second sign, A control signal output circuit that outputs a control signal representing a comparison voltage based on the excess component that exceeds the threshold potential at the membrane potential, It has, Of the plurality of synaptic circuits, each of the at least one first synaptic circuit that acquires the spike signal from the first neuron circuit, When the spike signal is obtained from the first neuron circuit, the control signal and the synaptic current corresponding to the synaptic weight are output. Signal processing method.

16. A signal processing method in a neural network device, The neural network device is Multiple synaptic circuits, each with its own assigned synaptic weight, Each of the multiple neuronal circuits outputs a spike signal, which is a voltage pulse. Equipped with, Each of the plurality of synaptic circuits acquires the spike signal output from any one of the plurality of neuronal circuits, and when it acquires the spike signal, it outputs a synaptic current of a current amount corresponding to the set synaptic weight. The first neuron circuit among the plurality of neuron circuits is An input circuit that acquires the synaptic current from each of at least one of the plurality of synaptic circuits, A charge-holding circuit that accumulates a charge corresponding to the synaptic current acquired by the input circuit and generates a membrane potential corresponding to the accumulated charge, A comparison circuit generates a determination signal which becomes a first value when the sign of the difference voltage between a preset threshold potential and the membrane potential is a first sign, and becomes a second value when the sign of the difference voltage is a second sign different from the first sign. A firing circuit that outputs the spike signal when the determination signal changes from the second value to the first value, A charge control circuit that, after outputting the spike signal, charges and discharges the charge stored in the charge holding circuit so that the sign of the differential voltage becomes the second sign, A control signal output circuit that outputs a control signal representing a comparison voltage based on the excess component that exceeds the threshold potential at the membrane potential, It has, The ignition circuit outputs the spike signal, which is a voltage pulse with a time width corresponding to the control signal. Of the plurality of synaptic circuits, each of the at least one first synaptic circuit that acquires the spike signal from the first neuron circuit, When the spike signal is obtained from the first neuron circuit, the synaptic current is output, which corresponds to the time width of the voltage pulse of the spike signal and the current amount corresponding to the synaptic weight. Signal processing method.

Images