A STDP neural network unsupervised online learning circuit and method

Abstract

The present invention relates to the field of artificial intelligence technology, and in particular to an STDP neural network unsupervised online learning circuit and method. The circuit comprises a front neuron, a back neuron, a photoelectric memristor, a long-term inhibition module, and a long-term enhancement module. The front neuron and the back neuron are both used to emit pulses. The long-term inhibition module is used to reduce the weight of the photoelectric memristor. The long-term enhancement module is used to increase the weight of the photoelectric memristor. The photoelectric memristor is used to perform pulse transmission and weight adjustment. The method comprises: comparing the emission order of the emission pulses of the front neuron and the back neuron, and adjusting the synaptic weight of the photoelectric memristor. By introducing light regulation, the problem that traditional circuits cannot simultaneously transmit pulses and adjust the memristor weights is solved. The complexity of the circuit structure and control signals is reduced, which is conducive to improving circuit integration and reducing power consumption.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an STDP neural network unsupervised online learning circuit and method. Background Art

[0002] Most existing computers are based on the von Neumann architecture, which separates storage and computation. As processor performance increases, the speed at which data can be transferred between these units becomes a bottleneck, known as the von Neumann bottleneck. In biological neural systems, however, information processing and storage are combined, allowing computation to proceed in parallel. Spiking neural networks (SNNs), a neural network structure modeled after biological neural networks, hold promise for overcoming the limitations of the von Neumann architecture.

[0003] SNN training can generally be categorized into supervised and unsupervised learning. Supervised learning is typically based on the backpropagation algorithm, which still requires significant computational resources. Unsupervised learning, however, based on spike-time-dependent plasticity (STDP), offers the advantages of low power consumption and computational requirements. Furthermore, STDP, a learning rule discovered in biological neural systems, offers excellent biological interpretability and computational parallelism.

[0004] In existing SNN circuits, the Leaky-Integrate and Fire (LIF) neuron model is widely adopted due to its simple structure and ease of implementation. The pulse waveform emitted by this neuron is usually also relatively simple, which means that this waveform can only transmit pulses but cannot simultaneously adjust synaptic weights according to the STDP rule. As a result, SNN circuits with this structure require various additional control units to realize the transmission of pulses in the synaptic array and the adjustment of synaptic weights in different time periods. This not only greatly increases the complexity of the circuit and the complexity of the control signals, but also increases the circuit area and power consumption. Summary of the Invention

[0005] The purpose of the present invention is to provide an STDP neural network unsupervised online learning circuit and method, aiming to solve the technical problem in the prior art that various additional control units are required to realize the transmission of pulses in the synaptic array and the adjustment of synaptic weights in different time periods, which not only greatly increases the complexity of the circuit and the complexity of the control signal, but also increases the area and power consumption of the circuit.

[0006] To achieve the above objectives, the present invention adopts an STDP neural network unsupervised online learning circuit, including a front neuron, a rear neuron, a photoelectric memristor, a long-term inhibition module and a long-term enhancement module, wherein the front neuron and the rear neuron are respectively connected to the photoelectric memristor, the long-term inhibition module is respectively connected to the front neuron and the rear neuron, and the long-term enhancement module is respectively connected to the front neuron and the rear neuron;

[0007] The front neuron and the rear neuron are both used to emit pulses;

[0008] The long-term suppression module is used to reduce the weight of the photoelectric memristor;

[0009] The long-term enhancement module is used to increase the weight of the photoelectric memristor;

[0010] The photoelectric memristor is used for pulse transmission and weight adjustment.

[0011] The long-time enhancement module includes an electronically controlled switch S1, an electronically controlled switch S2, a resistor R1, a resistor R2, a capacitor C1, an NMOS transistor N1, a light-emitting diode D1, and a node A. The electronically controlled switch S1 is connected to the front neuron, the electronically controlled switch S2 is connected to the back neuron, the resistor R1 is respectively connected to the electronically controlled switch S1, the resistor R2, the capacitor C1, and the NMOS transistor N1, the electronically controlled switch S2 is respectively connected to the NMOS transistor N1 and the light-emitting diode D1, and the node A is provided between the resistor R1, the resistor R2, the capacitor C1, and the NMOS transistor N1.

[0012] The resistor R2 and the capacitor C1 are connected to the DC voltage port V th1 ; where V th1 is the threshold voltage of the NMOS transistor N1.

[0013] The long-time suppression module includes an electronically controlled switch S3, an electronically controlled switch S4, a resistor R3, a resistor R4, a capacitor C2, an NMOS transistor N2, a light-emitting diode D2, and a node B. The electronically controlled switch S4 is connected to the front neuron, the electronically controlled switch S3 is connected to the back neuron, the electronically controlled switch S3 is respectively connected to the resistor R3, the resistor R4, the capacitor C2, and the NMOS transistor N2, the electronically controlled switch S4 is respectively connected to the NMOS transistor N2 and the light-emitting diode D2, and the node B is provided between the electronically controlled switch S3, the resistor R4, the capacitor C2, and the NMOS transistor N2.

[0014] The resistor R3 is connected to the DC voltage port V th2; where V th2 is the threshold voltage of the NMOS transistor N2.

[0015] The present invention also provides an STDP neural network unsupervised online learning method, comprising: comparing the emission order of the emission pulses of the front neuron and the rear neuron, and adjusting the synaptic weight of the photoelectric memristor.

[0016] Wherein, in the step of comparing the emission order of the emission pulses of the front neuron and the rear neuron and adjusting the synaptic weight of the photoelectric memristor:

[0017] If the front neuron emits a pulse earlier than the rear neuron, the synaptic weight of the photoelectric memristor increases, and the smaller the time difference between the two pulses, the greater the increase in the weight.

[0018] Wherein, in the step of the front neuron emitting a pulse earlier than the rear neuron:

[0019] When the front neuron emits a pulse, the electronically controlled switch S1 is briefly closed, and the voltage at the node A rises;

[0020] When the front neuron does not emit a pulse, the node A discharges through the resistor R2 and the voltage decreases exponentially;

[0021] When the post-neuron emits a pulse, the electronically controlled switch S2 is closed, and the voltage at the node A is converted into a current flowing through the light-emitting diode D1 through the transconductance of the NMOS transistor N1.

[0022] Wherein, in the step of comparing the emission order of the emission pulses of the front neuron and the rear neuron and adjusting the synaptic weight of the photoelectric memristor:

[0023] If the rear neuron emits a pulse earlier than the front neuron, the synaptic weight of the photoelectric memristor decreases, and the smaller the time difference between the two pulses, the greater the weight reduction.

[0024] Wherein, in the step of the rear neuron emitting a pulse earlier than the front neuron:

[0025] When the post-neuron emits a pulse, the electronically controlled switch S3 is briefly closed, and the voltage at the node B decreases;

[0026] When the post-neuron does not transmit a pulse, the node B is charged through the resistor R4, and the voltage difference between the power supply and the node B decreases exponentially;

[0027] When the front neuron emits a pulse, the electronically controlled switch S4 is closed, and the voltage at the node B is converted into a current flowing through the light emitting diode D2 through the transconductance of the NMOS transistor N2.

[0028] The present invention discloses an STDP neural network unsupervised online learning circuit and method, wherein both the front neuron and the back neuron are used to emit pulses; the long-term suppression module is used to reduce the weight of the photoelectric memristor; the long-term enhancement module is used to increase the weight of the photoelectric memristor; and the photoelectric memristor is used to perform pulse transmission and weight adjustment. Compared with traditional circuits, the present invention solves the problem that traditional circuits cannot simultaneously transmit pulses and adjust memristor weights by introducing optical regulation, under the premise that the pulse shapes of the front neuron and the back neuron are simple. The circuit structure and the complexity of the control signal are reduced, which is conducive to improving circuit integration and reducing power consumption. In addition, compared with traditional pure electric signal regulation, optical signals have the advantages of fast operation speed, high bandwidth, low crosstalk, and no Joule heat. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0030] Figure 1 It is a circuit principle diagram of the STDP neural network unsupervised online learning circuit of the present invention.

[0031] Figure 2 This is a schematic diagram of STDP learning based on trace implementation of the present invention.

[0032] Figure 3 It is a schematic diagram of the long-term enhancement module part in the STDP learning of the present invention.

[0033] Figure 4 It is a schematic diagram of the long-term inhibition module part in the STDP learning of the present invention.

[0034] Figure 5 This is a relationship diagram between the weight change and the time difference between the previous and next neuron pulses when only light regulation is considered in the present invention.

[0035] Figure 6 This is a diagram showing the relationship between the weight change and the time difference between the preceding and following neuron pulses when light control and electrical control are considered in the present invention.

[0036] Figure 7 It is a schematic diagram of the overall structure of the SNN of the present invention.

[0037] Figure 8 This is a circuit diagram of a long-term enhancement module in the SNN of the present invention.

[0038] Figure 9 This is the circuit schematic diagram of a long-term inhibition module in the SNN of the present invention. DETAILED DESCRIPTION

[0039] Exemplary embodiments are described in detail herein, with examples illustrated in the accompanying drawings. When the following description refers to the drawings, identical numerals in different drawings represent identical or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all embodiments consistent with this application.

[0040] The terms used in this application are for the purpose of describing specific embodiments only and are not intended to limit this application. As used in this application and the appended claims, the singular forms "a," "an," "the," and "the" are intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0041] It should be understood that although the terms first, second, third, etc. may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from each other. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0042] See also Figures 1 to 9 The present invention provides an STDP neural network unsupervised online learning circuit, comprising a front neuron, a rear neuron, a photoelectric memristor, a long-term inhibition module, and a long-term enhancement module, wherein the front neuron and the rear neuron are respectively connected to the photoelectric memristor, the long-term inhibition module is respectively connected to the front neuron and the rear neuron, and the long-term enhancement module is respectively connected to the front neuron and the rear neuron;

[0043] The front neuron and the rear neuron are both used to emit pulses;

[0044] The long-term suppression module is used to reduce the weight of the photoelectric memristor;

[0045] The long-term enhancement module is used to increase the weight of the photoelectric memristor;

[0046] The photoelectric memristor is used for pulse transmission and weight adjustment.

[0047] In this embodiment, the front neuron and the rear neuron are both used to emit pulses; the long-term suppression module reduces the weight of the photoelectric memristor; the long-term enhancement module increases the weight of the photoelectric memristor; the photoelectric memristor performs pulse transmission and weight adjustment; compared with traditional circuits, the present invention solves the problem that traditional circuits cannot transmit pulses and adjust memristor weights at the same time by introducing optical regulation, under the premise that the pulse shapes of the front neuron and the rear neuron are simple; it reduces the complexity of the circuit structure and the control signal, which is conducive to improving circuit integration and reducing power consumption; in addition, compared with traditional pure electrical signal regulation, optical signals also have the advantages of fast operation speed, high bandwidth, low crosstalk, and no Joule heat.

[0048] Furthermore, the long-time enhancement module includes an electronically controlled switch S1, an electronically controlled switch S2, a resistor R1, a resistor R2, a capacitor C1, an NMOS transistor N1, a light-emitting diode D1 and a node A, the electronically controlled switch S1 is connected to the front neuron, the electronically controlled switch S2 is connected to the back neuron, the resistor R1 is respectively connected to the electronically controlled switch S1, the resistor R2, the capacitor C1 and the NMOS transistor N1, the electronically controlled switch S2 is respectively connected to the NMOS transistor N1 and the light-emitting diode D1, and the node A is provided between the resistor R1, the resistor R2, the capacitor C1 and the NMOS transistor N1.

[0049] Furthermore, the resistor R2 and the capacitor C1 are connected to the DC voltage port V th1 ; where V th1 is the threshold voltage of the NMOS transistor N1.

[0050] Furthermore, the long-time suppression module includes an electronically controlled switch S3, an electronically controlled switch S4, a resistor R3, a resistor R4, a capacitor C2, an NMOS transistor N2, a light-emitting diode D2 and a node B. The electronically controlled switch S4 is connected to the front neuron, the electronically controlled switch S3 is connected to the back neuron, the electronically controlled switch S3 is respectively connected to the resistor R3, the resistor R4, the capacitor C2 and the NMOS transistor N2, the electronically controlled switch S4 is respectively connected to the NMOS transistor N2 and the light-emitting diode D2, and the node B is provided between the electronically controlled switch S3, the resistor R4, the capacitor C2 and the NMOS transistor N2.

[0051] Furthermore, the resistor R3 is connected to the DC voltage port V th2 ; where V th2 is the threshold voltage of the NMOS transistor N2.

[0052] In this embodiment, if Figure 1As shown, the long-term suppression module is the LTD module, the long-term enhancement module is the LTP module, and the photoelectric memristor is M, wherein the photoelectric memristor is optically written and electrically erased, that is, when exposed to light, the conductivity of the memristor increases, and the greater the light intensity, the greater the conductivity increase; when a reset pulse voltage is applied to the memristor, the conductivity of the memristor decreases. The overall circuit structure is shown in FIG. Figure 1 As shown, the electronically controlled switches S1 and S4 are controlled by the front neuron. When the front neuron emits a pulse, the electronically controlled switches S1 and S4 are closed, and are disconnected at other times. The electronically controlled switches S2 and S3 are controlled by the back neuron. When the back neuron emits a pulse, the electronically controlled switches S2 and S3 are closed, and are disconnected at other times.

[0053] The STDP neural network unsupervised online learning circuit implements STDP learning based on the trace method. The schematic diagram of implementing STDP learning by the trace method is as follows: Figure 2 As shown. When the front neuron emits a pulse, the front neuron trace rises to a certain value, and exponentially decays for the rest of the time; when the rear neuron emits a pulse, the rear neuron trace rises to a certain value, and exponentially decays for the rest of the time. In addition, when the front neuron emits a pulse, for example, at time t2 in the figure, the weight will decrease according to the value of the rear neuron trace. The larger the rear neuron trace value, the more the weight will decrease; when the rear neuron emits a pulse, for example, at time t1 in the figure, the weight will increase according to the value of the front neuron trace. The larger the front neuron trace value, the more the weight will increase.

[0054] In STDP learning, if the front neuron emits a pulse earlier than the back neuron, the synaptic weight increases, and the smaller the time difference between the two pulses, the greater the increase in weight. This is the long-term potentiation module LTP part in STDP learning, such as Figure 3 As shown; if the front neuron emits a pulse later than the back neuron, the synaptic weight decreases, and the smaller the time difference between the two pulses, the greater the weight reduction. This is the long-term inhibition part of STDP learning, as shown Figure 4 shown.

[0055] 1. STDP unsupervised online learning circuit based on photoelectric memristor

[0056] For the long-term enhancement module, when the front neuron emits a pulse, the electronically controlled switch S1 is closed, the capacitor C1 is charged, and the voltage of the node A rises by a certain value. When the front neuron has no pulse, the node A is discharged through the loop formed by the capacitor C1 and the resistor R2, and the voltage of the node A decreases exponentially. The above part realizes the change of the front neuron trace by the change of the voltage of the node A. When the rear neuron emits a pulse, the electronically controlled switch S2 is closed, and the voltage of the node A is converted into a current passing through the light-emitting diode D1 through the transconductance of the NMOS transistor N1. The larger the voltage of the node A, the greater the current passing through the light-emitting diode D1, the greater the brightness of the light-emitting diode D1, and the greater the increase in the weight of the photoelectric memristor, thereby realizing the regulation of the increase in the weight of the photoelectric memristor.

[0057] In the long-term inhibition (LTD) module, when the post-neuron emits a pulse, the electronically controlled switch S3 closes, the capacitor C2 discharges, and the node voltage B decreases by a certain value, that is, the difference between the power supply and the node B voltage increases by a certain value. When the post-neuron does not pulse, the node B is charged through the loop formed by the capacitor C2 and the resistor R4, and the difference between the power supply and the node B voltage decreases exponentially. This part realizes the change of the post-neuron trace based on the difference between the power supply and the node B voltage. When the front neuron emits a pulse, the electrically controlled switch S4 is closed, and the voltage of the node B is converted into a current passing through the light-emitting diode D2 through the transconductance of the NMOS transistor N2. The larger the voltage of the node B, the larger the current passing through the light-emitting diode D2, the greater the brightness of the light-emitting diode D2, and the greater the increase in the weight of the photoelectric memristor. At the same time, since the front neuron emits a pulse, a pulse voltage is applied to the photoelectric memristor, and the photoelectric memristor is regulated by the electrical pulse, and the weight is reduced by a certain value. By adjusting the parameters so that this weight reduction is slightly higher than the increase in the weight of the photoelectric memristor by light regulation when the node B voltage is equal to the power supply, the photoelectric memristor weight reduction regulation can be achieved.

[0058] When the above circuit only considers light control, the relationship between the change in the weight of the photoelectric memristor and the time difference between the previous and next pulses is as follows: Figure 5 When both optical control and electrical control are considered, the relationship between the change in the weight of the photoelectric memristor and the time difference between the front neuron and the back neuron is as follows: Figure 6 shown. Figure 5 、 Figure 6 The horizontal axis Δt represents the time when the front neuron pulse is emitted minus the time when the back neuron pulse is emitted, and the vertical axis represents the change in weight. Figure 6It can be seen that the weight change of the photoelectric memristor complies with the STDP rule. In addition, every time the front neuron emits a pulse, the pulse signal can be transmitted to the back neuron through the photoelectric memristor. This process does not conflict with the weight adjustment process of the photoelectric memristor, and the two can be carried out simultaneously.

[0059] 2. STDP neural network unsupervised online learning circuit based on photoelectric memristor

[0060] When the above circuits form a fully connected SNN network, the schematic diagram is as follows Figure 7 As shown. Pre1, Pre2, ..., Pre_n represent n front neurons; Post1, Post2, ..., Post_m represent m back neurons. Each front neuron is connected to each back neuron via the photoelectric memristor, forming a fully connected SNN network.

[0061] In the above network, each of the pre-neurons will have a corresponding long-term enhancement module. The long-term enhancement module corresponding to the pre-neuron Pre_x is taken as an example to illustrate the structure of the long-term enhancement module in the network. Figure 8 As shown. Whenever the front neuron Pre_x emits a pulse, the electronically controlled switch S10 is closed to adjust the trace value of the front neuron. When the back neuron Post1 emits a pulse, the electronically controlled switch S11 is closed, and the light-emitting diode D11 is turned on to adjust the weight of the photoelectric memristor between Pre_x and Post1; when the back neuron Post2 emits a pulse, the electronically controlled switch S12 is closed, and the light-emitting diode D12 is turned on to adjust the weight of the photoelectric memristor between Pre_x and Post2; and so on; when the back neuron Post_m emits a pulse, the electronically controlled switch S1m is closed, and the light-emitting diode D1m is turned on to adjust the weight of the photoelectric memristor between Pre_x and Post_m.

[0062] In the above network, each post-neuron has a corresponding long-term inhibition module LTD module. Taking the long-term inhibition module LTD module corresponding to the post-neuron Post_y as an example, the long-term inhibition module LTD module structure in the network is described. Figure 9As shown. Whenever the post-neuron Post_y emits a pulse, the electronically controlled switch S20 is closed to adjust the post-neuron trace value. When the pre-neuron Pre1 emits a pulse, the electronically controlled switch S21 is closed, and the light-emitting diode D21 is turned on to adjust the weight of the photoelectric memristor between Pre1 and Post_y; when the pre-neuron Pre2 emits a pulse, the electronically controlled switch S22 is closed, and the light-emitting diode D22 is turned on to adjust the weight of the photoelectric memristor between Pre2 and Post_y; and so on; when the pre-neuron Pre_n emits a pulse, the electronically controlled switch S2n is closed, and the light-emitting diode D2n is turned on to adjust the weight of the photoelectric memristor between Pre_n and Post_y.

[0063] The present invention also provides an STDP neural network unsupervised online learning method, comprising: comparing the emission order of the emission pulses of the front neuron and the rear neuron, and adjusting the synaptic weight of the photoelectric memristor.

[0064] Furthermore, in the step of comparing the emission order of the emission pulses of the front neuron and the rear neuron and adjusting the synaptic weight of the photoelectric memristor:

[0065] If the front neuron emits a pulse earlier than the rear neuron, the synaptic weight of the photoelectric memristor increases, and the smaller the time difference between the two pulses, the greater the increase in the weight.

[0066] Furthermore, in the step of the front neuron emitting a pulse earlier than the rear neuron:

[0067] When the front neuron emits a pulse, the electronically controlled switch S1 is briefly closed, and the voltage at the node A rises;

[0068] When the front neuron does not emit a pulse, the node A discharges through the resistor R2 and the voltage decreases exponentially;

[0069] When the post-neuron emits a pulse, the electronically controlled switch S2 is closed, and the voltage at the node A is converted into a current flowing through the light-emitting diode D1 through the transconductance of the NMOS transistor N1.

[0070] Furthermore, in the step of comparing the emission order of the emission pulses of the front neuron and the rear neuron and adjusting the synaptic weight of the photoelectric memristor:

[0071] If the rear neuron emits a pulse earlier than the front neuron, the synaptic weight of the photoelectric memristor decreases, and the smaller the time difference between the two pulses, the greater the weight reduction.

[0072] Furthermore, in the step of the rear neuron emitting a pulse earlier than the front neuron:

[0073] When the post-neuron emits a pulse, the electronically controlled switch S3 is briefly closed, and the voltage at the node B decreases;

[0074] When the post-neuron does not transmit a pulse, the node B is charged through the resistor R4, and the voltage difference between the power supply and the node B decreases exponentially;

[0075] When the front neuron emits a pulse, the electronically controlled switch S4 is closed, and the voltage at the node B is converted into a current flowing through the light emitting diode D2 through the transconductance of the NMOS transistor N2.

[0076] In this embodiment, in STDP learning, if the front neuron emits a pulse earlier than the rear neuron, the synaptic weight of the photoelectric memristor increases, and the smaller the time difference between the two pulses, the greater the increase in weight; in the long-term enhancement module, when the front neuron emits a pulse, the electrically controlled switch S1 is briefly closed, and the voltage of the node A rises; when the front neuron does not emit a pulse, the node A is discharged through the resistor R2, and the voltage index decreases. The node A voltage represents the change of the front neuron trace in the STDP rule implemented by trace; when the rear neuron emits a pulse, the electrically controlled switch S2 is closed, and the node A voltage is converted into a current flowing through the light-emitting diode D1 through the transconductance of the NMOS transistor N1. The light-emitting diode D1 emits different light intensities according to the different current intensities, which in turn causes the weight of the photoelectric memristor to increase by different amounts, thereby realizing the long-term enhancement part in STDP learning.

[0077] In STDP learning, if the rear neuron emits a pulse earlier than the front neuron, the synaptic weight of the photoelectromechanical resistor decreases, and the smaller the time difference between the two pulses, the greater the weight reduction; in the long-term inhibition module, when the rear neuron emits a pulse, the electrically controlled switch S3 is briefly closed, and the voltage of the node B decreases; when the rear neuron does not emit a pulse, the node B is charged through the resistor R4, and the voltage difference between the power supply and the node B decreases exponentially. The voltage difference between the power supply and the node B represents the change of the rear neuron trace in the STDP rule implemented by trace; when When the front neuron emits a pulse, the electronically controlled switch S4 is closed, and the voltage at node B is converted into a current flowing through the light-emitting diode D2 through the transconductance of the NMOS transistor N2. The light intensity emitted by the light-emitting diode D2 varies depending on the current intensity, thereby causing the weight of the photoelectric memristor to increase by a different amount. At the same time, since the electrical pulse emitted by the front neuron passes through the memristor, the conductance of the photoelectric memristor decreases under the action of the electrical signal while transmitting the pulse signal to the rear neuron. Under the joint action of the optical signal and the electrical signal, the long-term inhibition part of STDP learning can be achieved.

[0078] The above disclosure is merely one or more preferred embodiments of the present invention, and certainly cannot be used to limit the scope of the present invention. A person skilled in the art can understand that all or part of the processes of the above embodiments and equivalent changes made in accordance with the claims of the present invention still fall within the scope of the invention.

Claims

1. An STDP neural network unsupervised online learning circuit, characterized in that It includes a front neuron, a rear neuron, a photoelectric memristor, a long-term inhibition module and a long-term enhancement module, wherein the front neuron and the rear neuron are respectively connected to the photoelectric memristor, the long-term inhibition module is respectively connected to the front neuron and the rear neuron, and the long-term enhancement module is respectively connected to the front neuron and the rear neuron; The front neuron and the rear neuron are both used to emit pulses; The long-term suppression module is used to reduce the weight of the photoelectric memristor; The long-term enhancement module is used to increase the weight of the photoelectric memristor; The photoelectric memristor is used for pulse transmission and weight adjustment; The long-time enhancement module includes an electronically controlled switch S1, an electronically controlled switch S2, a resistor R1, a resistor R2, a capacitor C1, an NMOS transistor N1, a light-emitting diode D1, and a node A, wherein the electronically controlled switch S1 is connected to the front neuron, the electronically controlled switch S2 is connected to the back neuron, the resistor R1 is respectively connected to the electronically controlled switch S1, the resistor R2, the capacitor C1, and the NMOS transistor N1, the electronically controlled switch S2 is respectively connected to the NMOS transistor N1 and the light-emitting diode D1, and the node A is provided between the resistor R1, the resistor R2, the capacitor C1, and the NMOS transistor N1; The resistor R2 and the capacitor C1 are connected to the DC voltage port V th1 ; where V th1 is the threshold voltage of the NMOS transistor N1.

2. The STDP neural network unsupervised online learning circuit according to claim 1, wherein The long-time suppression module includes an electronically controlled switch S3, an electronically controlled switch S4, a resistor R3, a resistor R4, a capacitor C2, an NMOS transistor N2, a light-emitting diode D2, and a node B. The electronically controlled switch S4 is connected to the front neuron, the electronically controlled switch S3 is connected to the back neuron, the electronically controlled switch S3 is respectively connected to the resistor R3, the resistor R4, the capacitor C2, and the NMOS transistor N2, the electronically controlled switch S4 is respectively connected to the NMOS transistor N2 and the light-emitting diode D2, and the node B is provided between the electronically controlled switch S3, the resistor R4, the capacitor C2, and the NMOS transistor N2.

3. The STDP neural network unsupervised online learning circuit as claimed in claim 2, characterized in that The resistor R3 is connected to the DC voltage port V th2 ; where V th2 is the threshold voltage of the NMOS transistor N2.

4. A STDP neural network unsupervised online learning method, applied to the STDP neural network unsupervised online learning circuit according to claim 3, characterized in that: The emission order of the emission pulses of the front neuron and the rear neuron is compared to adjust the synaptic weight of the photoelectric memristor.

5. The STDP neural network unsupervised online learning method according to claim 4, wherein In the step of comparing the emission order of the emission pulses of the front neuron and the rear neuron and adjusting the synaptic weight of the photoelectric memristor: If the front neuron emits a pulse earlier than the rear neuron, the synaptic weight of the photoelectric memristor increases, and the smaller the time difference between the two pulses, the greater the increase in the weight.

6. The STDP neural network unsupervised online learning method according to claim 4, wherein: In the step of the front neuron firing a pulse earlier than the rear neuron: When the front neuron emits a pulse, the electronically controlled switch S1 is briefly closed, and the voltage at the node A rises; When the front neuron does not emit a pulse, the node A discharges through the resistor R2 and the voltage decreases exponentially; When the post-neuron emits a pulse, the electronically controlled switch S2 is closed, and the voltage at the node A is converted into a current flowing through the light-emitting diode D1 through the transconductance of the NMOS transistor N1.

7. The STDP neural network unsupervised online learning method according to claim 5, wherein: In the step of comparing the emission order of the emission pulses of the front neuron and the rear neuron and adjusting the synaptic weight of the photoelectric memristor: If the rear neuron emits a pulse earlier than the front neuron, the synaptic weight of the photoelectric memristor decreases, and the smaller the time difference between the two pulses, the greater the weight reduction.

8. The STDP neural network unsupervised online learning method according to claim 7, wherein: In the step of firing a pulse earlier than the firing of a pulse by the rear neuron: When the post-neuron emits a pulse, the electronically controlled switch S3 is briefly closed, and the voltage at the node B decreases; When the post-neuron does not transmit a pulse, the node B is charged through the resistor R4, and the voltage difference between the power supply and the node B decreases exponentially; When the front neuron emits a pulse, the electronically controlled switch S4 is closed, and the voltage at the node B is converted into a current flowing through the light emitting diode D2 through the transconductance of the NMOS transistor N2.

Images