Memristor synaptic circuit, artificial neuron and neural network system formed by memristor synaptic circuit and artificial neuron
By introducing an excitation/inhibition module of memristor synaptic circuit into the artificial neural network, the pulse signal transmission is dynamically controlled, which solves the problem of lack of excitation and inhibition mechanisms in the existing technology and improves the recognition rate and computational efficiency.
Patent Information
- Application Number
- CN202411088138.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-09
- Publication Date
- 2024-11-29
AI Technical Summary
Existing artificial neural networks lack incentive and inhibition mechanisms, which makes them unable to perfectly adapt to unknown environments and limits their computational speed.
By employing a memristor synaptic circuit and controlling the transmission of pulse signals through an excitation/inhibition module, dynamic adjustment of synaptic weights is achieved, simulating the excitation and inhibition mechanisms of biological neurons.
It improves the recognition rate of artificial neural networks, reduces storage resource consumption, and enables fast and efficient pulse signal processing.
Smart Images

Figure CN119047528A_ABST
Abstract
Description
Technical Field
[0001] The present application belongs to the field of artificial intelligence technology, and in particular relates to a memristor synaptic circuit, an artificial neuron, and a neural network system formed thereby. Background Art
[0002] The human brain consists of hundreds of billions of neurons that are extensively interconnected to form complex neural networks. Neurons provide intelligence capabilities related to learning and memory. Neurons use their synapses to exchange signals with thousands of other neurons in a neural network. Therefore, neurons are the structural and functional basic unit for data transmission across neural networks. A synapse refers to the junction between two neurons where the axon of the first neuron and the dendrite of the second neuron come close to each other in order to exchange signals. A single neuron is often connected to thousands of other neurons through synapses.
[0003] With the popularization of Internet technology, all aspects of human life have undergone great changes, and a large amount of data is generated anytime and anywhere. The use and processing of these data has given rise to the development of artificial intelligence. At the current stage, various big data models, neural network algorithms, and deep learning methods are emerging in an endless stream. Existing computer devices use the traditional von Neumann computing architecture with separate memory and arithmetic units. A large amount of data is moved back and forth on the bus, making it impossible to further improve the computing speed, thus showing obvious deficiencies in today's artificial intelligence era that relies on big data computing.
[0004] In order to solve the contradiction between data and computing, people have proposed some new computing paradigms, including quantum computing, in-memory computing, and brain-like computing. Among them, brain-like computing draws on the computing characteristics of the human brain and processes a large amount of data efficiently and in parallel by simulating neuron units. Deep learning technology based on artificial neural networks has achieved rapid development in recent years and has great use value in many fields such as smart security, smart home, smart medical care, and smart transportation. At present, most of the bottom layers of artificial neural networks use neurons based on multiplication and addition. After the input signal is weighted and summed, it passes through a nonlinear activation layer to generate an output signal. The bottom artificial neurons of existing artificial neural networks do not have an excitation and inhibition mechanism, which makes it impossible to perfectly realize the intelligent system so that it can adapt itself to an unknown environment. This invention comes from this. Summary of the invention
[0005] The embodiments of the present application provide a memristor synaptic circuit, an artificial neuron, and a neural network system formed thereby, which can realize the excitation and inhibition of artificial neurons, quickly screen out pulse signals, further reduce the consumption of storage resources, and can modify the synaptic weight of the pulse signal through the excitation signal and the inhibition signal. It has good bionics and can obtain a relatively satisfactory recognition rate.
[0006] In a first aspect, an embodiment of the present application provides a memristor synaptic circuit, characterized in that it includes:
[0007] A signal transmission module, used for transmitting a pulse signal, the pulse signal enters the memristor from the signal input end of the memristor and is output from the signal output end of the memristor;
[0008] The excitation / inhibition module is connected to the two ends of the memristor and is used to apply the same or opposite charges to the two ends of the memristor according to the synaptic excitation signal or the synaptic inhibition signal; wherein, when the synaptic excitation signal is connected, the memristor's memristor value decreases; when the synaptic inhibition signal is connected, the memristor's memristor value increases. The pulse signal transmission direction can be the current output direction of the pulse signal.
[0009] In some possible embodiments, the signal transmission module includes a memristor, a first transistor and a second transistor for controlling the transmission of a pulse signal; the first end of the first transistor is connected to the pulse signal, the second end of the first transistor is connected to the signal input end of the memristor, and its control end is connected to the input control signal; the first end of the second transistor is connected to the signal output end of the memristor, the second end of the second transistor is connected to the signal output end, and its control end is connected to the output control signal. The first end may be a source, and the second end may be a drain.
[0010] In some possible embodiments, the output control signal and the input control signal are the same control signal; the first transistor and the second transistor are both N-type MOS transistors or both P-type MOS transistors.
[0011] In some possible embodiments, the excitation / inhibition module includes a constant first voltage terminal and a constant second voltage terminal, the constant first voltage terminal is connected to the signal input terminal of the memristor through an excitation signal transistor, the signal output terminal of the memristor is connected to the constant second voltage terminal through an excitation signal transistor, and a synaptic excitation signal is applied to both ends of the memristor; the constant second voltage terminal is connected to the signal input terminal of the memristor through an inhibition signal transistor, the signal output terminal of the memristor is connected to the constant first voltage terminal through an inhibition signal transistor, and a synaptic inhibition signal is applied to both ends of the memristor; the constant first voltage terminal has a higher voltage than the constant second voltage terminal.
[0012] In some possible embodiments, the control end of the inhibition signal transistor is connected to an inhibition control signal. When the inhibition control signal is applied, the conduction of the inhibition signal transistor is controlled, and a synaptic inhibition signal is applied across the memristor; the control end of the excitation signal transistor is connected to an excitation control signal. When the excitation control signal is applied, a synaptic excitation signal is applied across the memristor.
[0013] In some possible embodiments, the signal transmission module further includes a unidirectional control submodule connected to the signal output end of the memristor, and configured to control the forward bias output or the reverse bias output of the signal transmission module.
[0014] In some possible embodiments, the memristor synapse circuit is a forward memristor synapse circuit, the unidirectional control submodule is a forward biased diode, the P end of the forward biased diode is connected to the signal output end of the memristor, and the N end is connected to the output.
[0015] In some possible embodiments, the memristor synapse circuit is a reverse memristor synapse circuit, the unidirectional control submodule is a reverse biased diode, the N end of the reverse biased diode is connected to the signal output end of the memristor, and the P end is connected to the output.
[0016] In a second aspect, an embodiment of the present application provides an artificial neuron, characterized in that the artificial neuron includes an input unit, a storage and calculation unit, a comparison unit, and an output unit;
[0017] The input unit comprises a plurality of memristor synaptic circuits according to any one of claims 1 to 5, and is used to obtain a pulse output signal transmitted on each memristor synaptic circuit;
[0018] The storage and calculation unit includes a storage module and a calculation module, wherein the storage module is connected to each memristor synaptic circuit of the input unit and is used to store the pulse output signal of the memristor synaptic circuit; the calculation module is used to calculate the pulse output signals of all memristor synaptic circuits on the input unit stored in the storage module to obtain a total output signal;
[0019] A comparison unit, used to determine the control conduction of the pulse signal in the output unit according to the comparison between the total output signal and the preset signal;
[0020] The output unit includes a plurality of the aforementioned memristor synaptic circuits and is used to control the forward bias output or reverse bias output of the pulse signal.
[0021] In some possible embodiments, the input unit includes n forward memristor synaptic circuits and m reverse memristor synaptic circuits, wherein n and m are integers and are not zero at the same time.
[0022] In some possible embodiments, when the artificial neuron is a forward artificial neuron, the output unit is p forward memristor synaptic circuits; when the artificial neuron is a negative artificial neuron, the output unit is q reverse memristor synaptic circuits, where p and q are natural numbers.
[0023] In a third aspect, an embodiment of the present application provides a neural network system, characterized in that it includes a central module and a signal transmission link connected to the central module,
[0024] The central module is used to send a neuron excitation signal or a neuron inhibition signal to each artificial neuron in the signal transmission link according to the transmission result of the pulse signal after being output through the signal transmission link;
[0025] The signal transmission link is formed by cascading x levels of artificial neurons in sequence, wherein the output unit of the first level artificial neuron outputs a pulse signal, which is used as the pulse signal input of the input unit of the second level artificial neuron, and the two are cascaded in sequence, and the output unit of the last level artificial neuron outputs a pulse signal to the central module, wherein x is a natural number; wherein the neuron excitation signal or the neuron inhibition signal can enhance or inhibit the signal transmission of the artificial neuron.
[0026] In some possible embodiments, there are y signal transmission links, any one of which is formed by cascading x levels of artificial neurons in sequence, wherein the output unit of the first-level artificial neuron outputs a pulse signal, which is used as the pulse signal input of the input unit of the second-level artificial neuron, and the links are cascaded in sequence, and the output unit of the last-level artificial neuron outputs a pulse signal to the central module, wherein x and y are natural numbers.
[0027] The present application provides a memristor synaptic circuit, an artificial neuron, and a neural network system constructed thereby, which performs gain or inhibition regulation on the pulse signal of the synapse through an excitation signal and an inhibition signal, thereby realizing fast and efficient parallel processing of a large amount of pulse signal data. The neural network system constructed thereby can provide considerable intelligent analysis and data processing performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] In order to more clearly illustrate the technical solution of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.
[0029] Figure 1 A schematic diagram of the structure of a neural network system provided in an embodiment of the present application;
[0030] Figure 2 A schematic diagram of the structure of a signal transmission link in a neural network system provided in an embodiment of the present application.
[0031] Figure 3 A schematic diagram of the structure of an artificial neuron in a signal transmission link provided in an embodiment of the present application;
[0032] Figure 4A schematic diagram of the structure of an artificial neuron in a signal transmission link provided by another specific embodiment of the present application;
[0033] Figure 5 A schematic diagram of the structure of a memristor synaptic circuit in an artificial neuron provided in a specific embodiment of the present application;
[0034] Figure 6 A schematic diagram of the implementation structure of a memristor synaptic circuit in an artificial neuron provided in another specific embodiment of the present application;
[0035] Figure 7 A schematic diagram of the implementation structure of a forward memristor synaptic circuit in an artificial neuron provided in a specific embodiment of the present application;
[0036] Figure 8 A schematic diagram of the implementation structure of a reverse memristor synaptic circuit in an artificial neuron provided in a specific embodiment of the present application;
[0037] Fig. 9 A schematic diagram of the implementation structure of the storage and calculation unit in the artificial neuron provided in a specific embodiment of the present application;
[0038] Fig.10 A working timing diagram of a storage and calculation unit in an artificial neuron provided in a specific embodiment of the present application;
[0039] Fig.11 A schematic diagram of the implementation structure of a comparison unit in an artificial neuron provided in a specific embodiment of the present application;
[0040] Fig.12 A schematic diagram of the implementation structure of the output unit in the artificial neuron provided in a specific embodiment of the present application;
[0041] Fig.13 A specific implementation circuit diagram of an artificial neuron provided in a specific embodiment of the present application;
[0042] Fig.14 for Fig.13 Working timing diagram of the circuit;
[0043] Fig.15 A specific implementation circuit diagram of an artificial neuron provided in another specific embodiment of the present application;
[0044] Fig.16 for Fig.15 Working timing diagram of the circuit;
[0045] Fig.17 A block diagram of the electronic device architecture of a neural network chip based on a neural network system according to an embodiment of the present invention is provided. DETAILED DESCRIPTION
[0046] The features and exemplary embodiments of various aspects of the present application will be described in detail below. In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain the present application, rather than to limit the present application. For those skilled in the art, the present application can be implemented without the need for some of these specific details. The following description of the embodiments is only to provide a better understanding of the present application by illustrating the examples of the present application.
[0047] It should be understood that in the description of the specification and claims of this application, if there are terms such as "first", "second", etc., they are only used to distinguish the objects described, and do not have any order or technical meaning. Therefore, the objects defined as "first", "second", etc. may explicitly or implicitly include one or more of the objects. In addition, "one" or "a" and similar words do not indicate a quantitative limitation, but indicate the existence of at least one, and "multiple" means no less than two.
[0048] The human brain can learn and memorize large amounts of information by transmitting and processing various signals through a neural network formed by connecting a large number of neurons to each other. The large number of interconnections between neurons in the human brain is directly related to the massively parallel nature of biological computing, so various attempts have been made to efficiently process large amounts of information by simulating artificial neural networks. The behavior of biological neurons can be simulated by mathematical models. Artificial neural networks can perform operations based on received input data and generate output data based on the results of the operations.
[0049] An artificial neural network may include multiple artificial neurons. Artificial neurons may be referred to as "neurons", "processing elements (PE)", "units" or similar terms. The research on artificial neurons in the prior art is still in its infancy, and the simulation of the functions of artificial neurons is also very mechanical, requiring more peripheral circuits to simulate the functions.
[0050] like Figure 1As shown, an embodiment of the present invention provides a neural network system, including a central module 2 and a signal transmission link connected to the central module, wherein the central module is used to send a neuron excitation signal or a neuron inhibition signal to each artificial neuron in the signal transmission link according to the transmission result of the pulse signal after being output through the signal transmission link; the signal transmission link is formed by cascading x levels of artificial neurons 1 in sequence, wherein the output unit of the first-level artificial neuron 11 outputs a pulse signal as the pulse signal input of the second-level artificial neuron input unit 12, cascading in sequence, and the output unit of the last-level artificial neuron outputs a pulse signal to the central module, wherein x is a natural number. The neural network system of this embodiment can independently perform artificial neural network calculations, for example, it can perform artificial neural network calculations for facial recognition, it can perform artificial neural network calculations for cryptanalysis, it can perform artificial neural network calculations for trend inference, and it can perform artificial neural network calculations for pattern recognition.
[0051] Although the robot system composed of artificial neural networks can accomplish some tasks similar to humans, it is still a product of computer systems and is essentially different from human systems. Because of this, robots do not have a perfect emotional system and are not as sensitive as biological organisms. Humans process information through neurons and synapses, and a large number of synapses and neurons work together to complete complex thinking and instructions. Therefore, simulating human neurons and synapses and using them in robot systems is of great significance to the development of robot systems. The neural network system of the present invention can perform data analysis tasks such as speech recognition, image recognition, and image classification.
[0052] The following is a detailed description of each component of the neural network system.
[0053] like Figure 2 As shown, the most basic component of the signal transmission link in the specific embodiment of the present invention is an artificial neuron. The single signal transmission link is formed by cascading x levels of artificial neurons 1 in sequence. The first level artificial neuron receives the neural pulse signal, and the artificial neuron processes the neural pulse signal and then transmits it to the central module.
[0054] like Figure 3 and Figure 4As shown, the present invention develops an artificial neuron based on the knowledge of neurons, and the artificial neuron includes an input unit, a storage and calculation unit, a comparison unit, and an output unit; wherein the input unit includes a plurality of memristor synaptic circuits, which are used to obtain the pulse signal transmitted on each memristor synaptic circuit; the storage and calculation unit includes a storage module and a calculation module, and the storage module is connected to each memristor synaptic circuit, which is used to store the output signal stored by the memristor synaptic circuit; the calculation module is used to calculate the output signals of all memristor synaptic circuits on the input unit to obtain the total output signal; the comparison unit is used to judge the control conduction of the pulse signal in the output unit according to the comparison between the total output signal and the preset signal; the output unit includes a plurality of memristor synaptic circuits, which are used to control the forward bias output or reverse bias output of the pulse signal. The artificial neuron provided in this embodiment can be connected to multiple memristor synaptic circuits as input units, and can include a positive memristor synaptic circuit or a negative memristor synaptic circuit, so that different input signals can be received. The output unit can have all outputs be positive memristor synaptic circuit outputs or negative memristor synaptic circuit outputs as required.
[0055] In the neural network system of the present invention, the artificial neuron directly outputs a pulse signal to the next artificial neuron. The artificial neuron has a function of storing pulse electrical signals and calculating. The input unit of each artificial neuron can receive the stored pulse electrical signal through the corresponding second direction and calculate the result. The output unit compares the calculation result of the input unit with the preset value provided by the central module to determine whether to output a pulse signal - a spike. For example, when the cumulative value of the received calculation result is greater than or equal to a predetermined threshold (e.g., C_Vin2), each artificial neuron can output a spike.
[0056] The artificial neuron of the present invention designs a novel synaptic circuit based on a memristor, which simulates the structure and function of human synapses and has a signal conduction and excitation inhibition mechanism. In this way, the memristor synaptic circuit in the signal transmission process can be fed back in time according to the calculation results, and the excitation inhibition signal can be adjusted, so that the weight parameters of the memristor synaptic circuit can be quickly adjusted, which is beneficial to practical use. Figure 5 As shown, an embodiment of the present invention provides a memristor synapse circuit, including: a signal transmission module, used to transmit a pulse signal, the pulse signal enters the memristor from the signal input end of the memristor, and is output from the signal output end of the memristor; an excitation / inhibition module, connected to the two ends of the memristor, used to apply opposite charges to the two ends of the memristor according to a synaptic excitation signal or a synaptic inhibition signal; wherein, when the synaptic excitation signal is connected, the memristor's memristor value decreases; when the synaptic inhibition signal is connected, the memristor's memristor value increases.
[0057] Each memristor synaptic circuit may be a circuit that simulates a synapse between neurons. The memristor in the memristor synaptic circuit has a memory property and can store weights corresponding to the connection strength between neurons. Each memristor synaptic circuit may include a memory element for storing weights, or may be connected to a memory element that already has weights. In an example embodiment, such a memory element may be a memristor or include a memristor.
[0058] The present invention utilizes the resistive storage characteristics of the memristor. When a charge is applied at both ends of the memristor according to an excitation signal, the memristor changes from a high-resistance state (HRS) to a low-resistance state (LRS). In this way, applying an electrical excitation signal to the device will cause the resistance to decrease and the current flowing through the device to increase. On the contrary, when an opposite charge is applied at both ends of the memristor according to an inhibition signal, the memristor is reset from a low-resistance state (LRS) to a high-resistance state (HRS), which will cause the resistance to increase and the current flowing through the device to decrease. The process of the memristor changing from a high-resistance state (HRS) to a low-resistance state (LRS) is a "SET" process; conversely, the process from LRS to HRS is a "RESET" process.
[0059] By generating an excitation signal and an inhibition signal in the excitation / inhibition module, the memristor's memristance value changes, which is used to characterize the weight change of the artificial simulated synapse. Synaptic plasticity is the neurobiological basis of brain memory and learning, the basic characteristic of artificial neural circuits, and the primary function that artificial electronic synaptic devices need to achieve. The increase and decrease in the conductivity of the memristor correspond to the increase and decrease in the synaptic weight, respectively. The neural network system can achieve different data processing by changing the stimulation pulse signal (such as the shape, frequency, duration and other parameters of the voltage to simulate the characteristics of the neural stimulation signal corresponding to different synaptic functions), and the neural network system can also adjust the parameters of the acceptance of the results.
[0060] In the technical solution of the present invention, the reward and punishment mechanism of the incentive inhibition module is implemented independently and separated from the transmission signal. The incentive signal and the inhibition signal can be controlled independently, and the direct intervention and parameter adjustment of the reward and punishment mechanism by the neural network system can be realized to avoid interference in the signal transmission process.
[0061] like Figure 6The figure shows a typical circuit structure diagram of the memristor synaptic circuit of the present invention. In the preferred technical solution of the present invention, the signal transmission module includes a memristor MR_R, a first transistor T1 and a second transistor T2; the first end of the first transistor is connected to the pulse signal MR_S+ (MR_S-), the second end of the first transistor is connected to the signal input end of the memristor, and its control end is connected to the input control signal; the first end of the second transistor is connected to the signal output end of the memristor, the second end of the second transistor is connected to the signal output end MR_S- (MR_S+), and its control end is connected to the output control signal. The present invention uses a signal transmission module to implement separately, and the functions of the signal transmission module and the excitation suppression module are independently implemented to avoid interference of the excitation suppression module to the signal transmission module. The signal transmission is controlled by a transistor in the signal transmission module, and the mature process of the transistor can be integrated into the integrated process of the memristor to achieve rapid intervention and have strong anti-interference ability.
[0062] In a specific implementation manner of the present invention, the output control signal and the input control signal can be the same control signal MR_S; through the same signal of the output control signal and the input control signal, the signal transmission module can be uniformly set to the same type of transistors, which is convenient for panel-level integration or chip-level integration. In a preferred technical solution, the first transistor T1 and the second transistor T2 are both N-type MOS tubes (NMOS tubes) or both P-type MOS tubes (PMOS tubes).
[0063] like Figure 6 As shown, the excitation / inhibition module includes a constant first voltage terminal MR_H and a constant second voltage terminal MR_L, wherein the constant first voltage terminal MR_H is connected to the signal input terminal of the memristor through an excitation signal transistor TA, and the signal output terminal of the memristor is connected to the constant second voltage terminal through an excitation signal transistor TA, and an excitation signal is applied to both ends of the memristor; the constant second voltage terminal MR_L is connected to the signal input terminal of the memristor through an inhibition signal transistor TN, and the signal output terminal of the memristor MR_R is connected to the constant first voltage terminal through an inhibition signal transistor TN, and an inhibition signal is applied to both ends of the memristor. The constant first voltage terminal MR_H can be a constant high voltage, and the constant second voltage terminal MR_L can be a constant low voltage.
[0064] In a preferred technical solution of the present invention, the control end of the inhibition signal transistor is connected to the inhibition control signal MR_N. When the inhibition control signal MR_N is applied, the conduction of the inhibition signal transistor is controlled, and the inhibition signal is applied at both ends of the memristor; the control end of the excitation signal transistor is connected to the excitation control signal MR_A. When the excitation control signal MR_A is applied, the excitation signal is applied at both ends of the memristor.
[0065] The technical solution of the present invention simplifies the structural design of artificial neurons, integrates memristor synaptic circuits inside artificial neurons, and implements both input units and output units through memristor synaptic circuits. This eliminates the need to construct pre-synaptic and post-synaptic environments, making artificial neurons more flexible and providing greater flexibility for neural network algorithms.
[0066] Each memristor synapse circuit may be implemented with a memory element such as a memristor element. The memristor element may store weights therein through a memristor-based design and perform multiplication (i.e., an AND operation) at the intersection. The memristor element of each memristor synapse circuit according to the present example embodiment may be implemented as a resistor element such as a phase change random access memory (PRAM) using a phase change material or a resistive random access memory (RRAM) using a variable resistance material such as a composite metal oxide. The material constituting the resistor element has a variable resistance value that changes according to the magnitude and / or direction of a current or voltage, and has a non-volatile property of maintaining its resistance even when the current or voltage is cut off.
[0067] The memristor in each memristor synaptic circuit may include a control gate electrode, a gate insulating film, polysilicon oxide, a resistive material or a phase change material, and a pillar oxide. The memristor element may be in the form of a pillar, with the pillar oxide being at the center of the memristor element and being surrounded by a plurality of layers in the order of the resistive material, the polysilicon oxide, the gate insulating film, and the control gate electrode. The control gate electrode may be an electrode pad. The control gate electrode may be made of platinum (Pt), ruthenium (Ru), iridium (Ir), silver (Ag), aluminum (Al), titanium (Ti), tantalum (Ta), tungsten (W), silicon (Si), copper (Cu), nickel (Ni), cobalt (Co), a conductive nitride thereof (e.g., TiN), or a combination thereof (e.g., Ti / TiN). In an exemplary embodiment, the control gate electrode may be made of TiN, which is well compatible with CMOS-based Si processes.
[0068] The memristor may include at least one of materials having a characteristic (i.e., a variable resistance characteristic), wherein the resistance of the memristor can be selectively changed by a current flowing through the memristor element. According to some example embodiments, the resistive material may be a transition metal oxide film, for example, any one of TiO2, NiO, HfO2, Al2O3, ZrO2, and ZnO films or a combination thereof. In another embodiment, the resistive material may be formed of a material capable of variably switching resistance, such as a chalcogen-based compound or a perovskite-based compound. A material having a large ratio between a resistance value in a low resistance state and a resistance value in a high resistance state and having a small driving voltage for reducing power consumption may be used.
[0069] The resistive material and the control gate electrode can be a single layer, but can have a multilayer structure including a barrier film suitable for other embodiments. The current can flow through the channel layer (i.e., polysilicon oxide), and when the current flows through the resistive material, the resistance value changes according to the degree of the current flowing through the resistive material. Then, the output current (or output signal) can be determined from the resistance value that changes according to the input voltage (or input signal). In the present application, by controlling the opening and closing of the excitation transistor and the suppression transistor, the current flowing through the memristor element is changed, thereby changing the high resistance state or low resistance state of the memristor.
[0070] In another specific embodiment of the present invention, Figure 7 and Figure 8 As shown, the signal transmission module further includes a unidirectional control submodule connected to the signal output end of the memristor, which is used to control the forward bias output or reverse bias output of the signal transmission module.
[0071] Specific as Figure 7 As shown, the memristor synaptic circuit in the scheme is a forward memristor synaptic circuit, the unidirectional control submodule is a forward biased diode, the P end of the forward biased diode is connected to the signal output end of the memristor, and the N end is connected to the output; when the excitation signal transistor TA is turned on, the constant first voltage end MR_H is connected to the signal input end of the memristor through the excitation signal transistor TA, the signal output end of the memristor is connected to the constant second voltage end through the excitation signal transistor TA, and the excitation signal is applied to both ends of the memristor, the excitation signal reduces the resistance of the memristor element, and enhances the signal transmission effect of the forward memristor synaptic circuit; when the inhibition signal transistor TN is turned on, the constant second voltage end MR_L is connected to the signal input end of the memristor through the inhibition signal transistor TN, the signal output end of the memristor MR_R is connected to the constant first voltage end through the inhibition signal transistor TN, and the inhibition signal is applied to both ends of the memristor, the resistance of the memristor element is increased, and the signal transmission effect of the forward memristor synaptic circuit is inhibited.
[0072] On the contrary, Figure 8As shown, in the preferred technical solution of the present invention, the memristor synaptic circuit is a reverse memristor synaptic circuit, the unidirectional control submodule is a reverse biased diode, the N end of the reverse biased diode is connected to the signal output end of the memristor, and the P end is connected to the output. When the excitation signal transistor TA is turned on, the constant first voltage end MR_H is connected to the signal output end of the memristor through the excitation signal transistor TA, and the signal output end of the memristor is connected to the constant second voltage end through the excitation signal transistor TA, and the excitation signal is applied to both ends of the memristor. Since the excitation signal is opposite to the polarity of the memristor element, the excitation signal increases the resistance of the memristor element, thereby enhancing the signal transmission effect of the reverse memristor synaptic circuit; when the inhibition signal transistor TN is turned on, the constant second voltage end MR_L is connected to the signal output end of the memristor through the inhibition signal transistor TN, and the signal input end of the memristor MR_R is connected to the constant first voltage end through the inhibition signal transistor TN, and the inhibition signal is applied to both ends of the memristor, and the inhibition signal reduces the resistance of the memristor element, thereby inhibiting the signal transmission effect of the reverse memristor synaptic circuit.
[0073] The neuron excitation signal is the excitation signal of the entire artificial neuron, including the excitation of the forward memristor synaptic circuit and the inhibition of the reverse memristor synaptic circuit. Each neuron can include multiple memristor synaptic circuits, which can have a forward memristor synaptic circuit or a reverse memristor synaptic circuit. The synaptic excitation signal is formed by the control signal of the excitation transistor, the constant first voltage terminal and the constant second voltage terminal in a single memristor synaptic circuit. Regardless of the forward memristor synaptic circuit or the reverse memristor synaptic circuit, the overall effect of the neuron excitation signal is to reduce the resistance of the memristor. The neuron excitation signal is the sum of the excitation signals of all memristor synaptic circuits of the input unit in the artificial neuron; the neuron inhibition signal is the sum of the excitation signals of all memristor synaptic circuits of the input unit in the artificial neuron. The neuron inhibition signal and the neuron inhibition signal can be connected by one signal to enhance or weaken the signal of the entire artificial neuron.
[0074] Therefore, the neuron excitation signal can enhance the signal transmission of the artificial neuron, including reducing the resistance of the forward memristor synaptic circuit and increasing the resistance of the reverse memristor synaptic circuit; while the neuron inhibition signal can inhibit the signal transmission of the artificial neuron, including increasing the resistance of the forward memristor synapse and reducing the resistance of the reverse memristor synapse. Figure 7 and Figure 8 The forward memristor synaptic circuit and the reverse memristor synaptic circuit shown in FIG. 5 can construct a forward artificial neuron when the output units are all forward memristor synaptic circuits; and can construct a negative artificial neuron when the output units are all reverse memristor synaptic circuits.
[0075] In a preferred technical solution of the present invention, the input unit of the artificial neuron includes n forward memristor synaptic circuits and m reverse memristor synaptic circuits, wherein n and m are integers and are not zero at the same time. As an example, Figures 9 to 12 The figure shows four memristor synaptic circuits, two forward memristor synaptic circuits and two reverse memristor synaptic circuits, which receive the clock signal or neural pulse signal output by the output unit of the previous artificial neuron. When a neuron excitation signal is applied, both the forward memristor synaptic circuit and the reverse memristor synaptic circuit can enhance signal transmission; when a neuron inhibition signal is applied, both the forward memristor synaptic circuit and the reverse memristor synaptic circuit can inhibit signal transmission.
[0076] The storage and calculation unit in the artificial neuron in the embodiment of the present invention includes a storage module 21 and a calculation module 22. The storage module 21 is connected to each memristor synaptic circuit and is used to store the output signal stored by the memristor synaptic circuit; the calculation module is used to calculate the output signals of all memristor synaptic circuits on the input unit to obtain a total output signal.
[0077] like Fig. 9 As shown, it is a typical example of a storage calculation unit. Taking four memristor synaptic circuits as an example, the storage calculation unit electrically connects the pulse output signals of the four memristor synaptic circuits as input terminals S_V1, S_V2, S_V3, and S_V4. The storage module in the storage calculation unit includes storage capacitors C1, C2, C3, and C4 connected to the input terminals S_V1, S_V2, S_V3, and S_V4, respectively, for storing the pulse signals (pulse voltage values) obtained by the input terminals S_V1, S_V2, S_V3, and S_V4; the other end of each storage capacitor C1, C2, C3, and C4 is grounded respectively. The pulse signal input terminals S_V1, S_V2, S_V3, and S_V4 of the storage module are respectively controlled to be turned on by storage transistors TC1, TC2, TC3, and TC4; the storage transistor is a switching transistor, whose control terminal is connected to the S_AVE control signal to control the conduction, and its source and drain are respectively connected to each pulse signal input terminal and the storage capacitor.
[0078] like Fig. 9 and Fig.10As shown, the storage calculation unit also includes a calculation module 22. Taking the four memristor synaptic circuits as an example, each pulse signal input terminal S_V1, S_V2, S_V3, S_V4 controls the output S_DOUT signal through the third transistor T3P\T3N. The third transistor T3P\T3N is a switch transistor, and its control terminal is respectively connected to the S_RP\S_RN control signal to control the on and off of the third transistor S_RP\S_RN, and its source or drain is respectively controlled to connect to each pulse signal input terminal or one end of the storage capacitor C1, C2, C3, C4. After the storage capacitors C1, C2, C3, C4 store the pulse signals transmitted by the pulse signal input terminals S_V1, S_V2, S_V3, S_V4, the signal superposition can be performed through the calculation module, so that the signal superposition of the pulse signal input terminals S_V1, S_V2, S_V3, S_V4 is realized, that is, the addition operation of adding the results of each multiplication in the receiving neuron is realized. The pulse signals stored in each storage capacitor C1, C2, C3, and C4 are respectively connected to the source or drain of the third transistor T3P\T3N through the fourth transistor T41\T42\T43\T44\T45\T46\T47\T48 controlled by the S_Q1, S_Q2, S_Q3, and S_Q4 control signals.
[0079] The storage calculation unit also includes a zeroing module 23, which includes a zeroing transistor connected to the total output signal. The zeroing transistor is a switching transistor, and the other end of the zeroing transistor is grounded. When the zeroing transistor is turned on, the zeroing operation of the total output signal is realized. Fig.11 As shown, the zero-clearing transistor TS is a switch transistor, whose control end is connected to the S_CK control signal to control the conduction and shutdown of the zero-clearing transistor, and the other end is grounded. When the zero-clearing transistor is turned on, the zero-clearing operation of the total output signal C_Vin1 is realized to prevent the influence on the subsequent processing of the calculation unit. In the embodiment of the present invention, zero-clearing can be performed when the calculation result (C_Vin1 signal) is greater than the preset signal C_Vin2 signal to prevent the influence of the accumulated charge in the artificial neuron.
[0080] like Fig.11As shown, in the embodiment of the present invention, the comparison unit includes an activation module 31, a comparison module 32 and an output module 33, the comparison module 32 includes a sixth transistor T6 connected to the total output signal C_Vin1 of the storage calculation unit to control conduction and a seventh transistor T7 connected to the preset signal C_Vin2 to control conduction, the sources of the sixth transistor T6 and the seventh transistor are respectively connected to the power supply voltage signal; the sources of the sixth transistor T6 and the seventh transistor are grounded through the twelfth transistor T12. The output module 33 includes a comparison signal output terminal C_X connected to the source of the sixth transistor T6, a comparison signal control terminal C_Y connected to the source of the seventh transistor T7, and the output signal of the comparison signal output terminal is connected to the output unit. In the preferred technical solution, the output signal of the comparison signal output terminal is connected to the control terminal of the fifth transistor T5 to control the conduction of the fifth transistor T5.
[0081] Specifically, the activation module 31 includes an eighth transistor T81\T82 and a ninth transistor T91\T92, wherein the control end of the eighth transistor T81\T82 is connected to the clock signal, and the source of the sixth transistor and the seventh transistor are enabled, respectively, and a power supply voltage signal is provided; the ninth transistor T91\T92 is enabled for the comparison signal control end and the comparison signal output end, respectively, and a power supply voltage signal is provided. A tenth transistor T10 is arranged between the comparison signal control end and the source of the sixth transistor, and the control end of the tenth transistor T10 is connected to the comparison signal output end; an eleventh transistor T11 is arranged between the comparison signal output end and the source of the seventh transistor, and the control end of the eleventh transistor T11 is electrically connected to the comparison signal control end.
[0082] like Fig.11As shown, the eighth transistor T81\T82 and the ninth transistor T91\T92 are both PMOS transistors. When the clock signal C_CK is a high-level signal, the eighth transistor T81\T82 and the ninth transistor T91\T92 are turned off. When the clock signal C_CK is a low-level signal, the eighth transistor T81\T82 and the ninth transistor T91\T92 are turned on. The power supply voltage signal C_VDD is transmitted to the comparison signal control terminal C_Y and the comparison signal output terminal C_X through the PMOS transistors, the eighth transistor T81\T82 and the ninth transistor T91\T92. The voltage signal at the comparison signal control terminal C_X can control the conduction of the PMOS transistor C_M6 and the NMOS transistor tenth transistor T10; the voltage signal at the comparison signal output terminal C_Y can control the conduction of the PMOS transistor C_M5 and the NMOS transistor eleventh transistor T11. The clock signal C_CK also controls the conduction of the twelfth transistor T12, and the twelfth transistor T12 can be an NMOS transistor. When the voltage of C_VIN1 is greater than the voltage of C_VIN2, the sixth transistor T6 controlled by the voltage signal of C_VIN1 leaks current faster than the seventh transistor T7 controlled by the voltage signal of C_VIN2, causing the voltage at the comparison signal output terminal to drop rapidly, and the tenth transistor T10 connected to the comparison signal output terminal is turned off. The voltage signal at the comparison signal control terminal maintains a stable high voltage, and the low-level signal output from the comparison signal output terminal controls the fifth transistor T5 of the output unit to turn on, and the periodic clock signal in the output unit generates pulses to output pulse signals to each memristor synapse circuit.
[0083] like Fig.12 As shown, the output unit also includes a pulse signal generating module 41, and the pulse signal generating module includes a periodic clock signal controlled by a fifth transistor T5, and the periodic clock signal is selected from a periodic high-level pulse signal G_CK_H or a periodic low-level pulse signal G_CK_L. The control end of the fifth transistor is connected to the output signal of the comparison unit, the source is connected to the periodic clock signal, and the drain is respectively connected to the signal input ends of several memristor synapse circuits. The periodic clock signal is connected through the control of the comparison unit, and can control the periodic clock signal to enter the memristor synapse circuit, and generate a neural pulse signal of the next level of artificial neuron after being output by the memristor synapse circuit. Fig.13 and Fig.15 As shown, the output unit includes a plurality of memristor synaptic circuits for controlling the forward bias output or reverse bias output of the pulse signal. In a preferred technical solution of the present invention, when the artificial neuron is a forward artificial neuron, the output unit is p forward memristor synaptic circuits; when the artificial neuron is a negative artificial neuron, the output unit is q reverse memristor synaptic circuits, wherein p and q are natural numbers.
[0084] Fig.13 A forward artificial neuron comprising an output unit of four memristor synaptic circuits is shown; Fig.14 The timing of the forward artificial neuron is provided, and the four memristor synaptic circuits are all memristor synaptic circuits with forward bias output, that is, the forward artificial neuron. The pulse signal input end of each memristor synaptic circuit is connected to the clock pulse signal control through the fifth transistor G_M1. The fifth transistor is a switching transistor, and its control end is connected to the C_Y end of the switch unit, and the conduction and closing of the fifth transistor are controlled according to the voltage of the C_Y end of the switch unit. Fig.15 An output unit including four memristor synaptic circuits is shown, and the four memristor synaptic circuits are all memristor synaptic circuits with reverse bias outputs, that is, reverse artificial neurons; Fig.16 The timing of the reverse artificial neuron is provided.
[0085] The pulse signal input terminals of the multiple memristor synaptic circuits are controlled and connected to multiple clock pulse signals through the fifth transistor G_M1; in the preferred technical solution, the multiple clock pulse signals share one pulse signal for outputting a high-level pulse signal G_CK_H or a low-level pulse signal G_CK_L, so that the pulse signal input terminals of the multiple memristor synaptic circuits are controlled and connected to one pulse signal through the fifth transistor G_M1. The periodic clock signal G_CK_H in the output unit is used to form the neural pulse signal of the next level MR_S+; and the periodic clock signal G_CK_L is used to form the neural pulse signal of the next level MR_S-.
[0086] When the periodic clock signal uses a high-level pulse signal G_CK_H, a memristor synaptic circuit with a forward biased output can be selected for output, and the output pulse signal can be used as the input of the neural pulse signal of the memristor synaptic circuit with a forward biased output of the next-level artificial neuron; when the periodic clock signal uses a low-level pulse signal G_CK_L, a memristor synaptic circuit with a reverse biased output can be selected for output, and the output pulse signal can be used as the input of the neural pulse signal of the memristor synaptic circuit with a reverse biased output of the next-level artificial neuron.
[0087] The artificial neurons of the neural network system are described in detail above. The core part of the neural network system is the structural composition of the artificial neurons. The artificial neurons described in level x are cascaded in sequence to form a signal transmission link, which is transmitted to the central module through the signal transmission link to realize the specific computing function of the neural network. The artificial neurons of the present invention are combined with input units, storage and calculation units, comparison units and output units, and the whole can realize the neuron function. A plurality of memristor synaptic circuits are arranged in the input unit and the output unit, and the memristor synaptic circuit has a reward and punishment mechanism, as well as signal transmission, and the reward and punishment mechanism are independently designed, which can facilitate the flexible implementation of subsequent algorithms.
[0088] The core of the artificial neuron of the present invention is a memristor synaptic circuit. Arranging all the memristor synaptic circuits in the neural network system in a first direction (such as a horizontal direction) and a second direction (such as a vertical direction) can form a synaptic array. The first direction and the second direction intersect each other. The first direction can be the direction in which the pulse signal input line extends, and the second direction can be the direction in which the pulse signal output line extends. The memristor synaptic circuit is arranged in the intersection area of the pulse signal input line and the pulse signal output line. It is generally understood that the first direction can be a row direction and the second direction can be a column direction, so that the neural network system can be arranged and integrated into a neural network chip. According to some example embodiments, the synaptic array is implemented as an n×m array to achieve high integration of the neural network chip.
[0089] In the preferred technical solution of the present invention, there are y signal transmission links, any one of which is formed by sequentially cascading x levels of artificial neurons, wherein the output unit of the first-level artificial neuron outputs a pulse signal, which is used as the pulse signal input of the second-level artificial neuron input unit, and the two are sequentially cascaded, and the output unit of the last-level artificial neuron outputs a pulse signal to the central module, wherein x and y are natural numbers. In the technical solution of the present invention, the number of artificial neurons on each signal transmission link can be equal or unequal, and the specific signal transmission link can be matched according to different algorithms based on the specific information of the pulse signal input. In a usual manner, artificial neurons between multiple signal transmission links can be shared, and the shared artificial neurons can receive pulse signals output by artificial neurons of different signal transmission links.
[0090] The operation of the artificial neural network can include a multiplication operation of multiplying the output value of the neuron in the previous layer by the weight of the synapse and an addition operation of adding the results of each multiplication in the receiving neuron. The present invention directly implements the addition and multiplication operations of the matrix through artificial neurons. The central module can be a specific algorithm module or a deeper neural network system.
[0091] Fig.17The block diagram of the electronic device of the neural network chip of the neural network system of the embodiment of the present invention is shown. The neural network system in the embodiment of the present invention can be solidified into a chip, and the neural network chip based on the neural network system can be used as a device with artificial neuron reasoning and data processing functions for some data processing, such as Fig.15 As shown, in addition to the neural network chip, other common components (for example, central processing unit (CPU), graphics processing unit (GPU), application processor (AP), sensors, communication modules, etc.) and so on.
[0092] Such an electronic device may be included in various types of electronic devices such as personal computers (PCs), server devices, mobile devices, and embedded devices. Such a device may correspond to a hardware component included in, for example, a smart phone, a tablet device, an augmented reality (AR) device, an Internet of Things (IoT) device, an autonomous vehicle, a robot, a medical device, etc., which performs, for example, speech recognition, image recognition, and image classification using a neural network. Therefore, the device may correspond to a dedicated hardware accelerator installed on the above-mentioned electronic device, or to a hardware accelerator such as a neural processor (NPU), a tensor processor (TPU), or a neural engine (Neural Engine) (a neural engine is a dedicated module for driving a neural network), etc.).
[0093] The neural network chip can be used to control the overall function of the neural network system used to drive the device.
[0094] The memory may be hardware that stores various neuromorphic data processed in the neural network chip. The memory may include a random access memory (RAM) (such as a dynamic random access memory (DRAM) and a static random access memory (SRAM)), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a CD-ROM, a Blu-ray or another optical disk storage device, a hard disk drive (HDD), a solid-state drive (SSD), or a flash memory. The device may analyze input data in real time based on a neural network to extract valid information, based on the extracted information. For example, the device may be applied to robotic devices (such as drones and advanced driver assistance systems (ADAS)), smart TVs, smart phones, medical devices, mobile devices, image display devices, measuring devices, and IoT devices, etc. In addition, the device may be installed on at least one of various other types of electronic devices.
[0095] The electronic device may include a processor, RAM, a neural network chip, a memory, a sensor, and a communication module. It may also include an input / output module, a security module, a power control device, etc. The processor may control the overall operation of the electronic device. The processor may include one processor core (single core) or multiple processor cores (multi-core). The processor may process or execute programs and / or data stored in the memory. In some example embodiments, the processor may control the functions of the neural network chip by executing a program stored in the memory. The processor may be implemented as a CPU, a GPU, an AP, etc.
[0096] RAM can temporarily store programs, data or instructions. For example, programs and / or data stored in the memory can be temporarily stored in RAM according to the control or boot code of the processor. RAM can be implemented as dynamic RAM (DRAM), static RAM (SRAM), etc.
[0097] The neural network chip can perform calculations of the neural network based on the received input data, and generate an information signal based on the calculation results. The neural network may include, but is not limited to, CNN, RNN, FNN, deep belief network, restricted Boltzmann machine, etc. According to an example embodiment, the neural network chip is a hardware accelerator dedicated to the neural network or a device including a hardware accelerator, and may include a neuromorphic processor according to the above-mentioned example embodiment.
[0098] The information signal may include one of various types of recognition signals (such as a voice recognition signal, an object recognition signal, an image recognition signal, and a biometric information recognition signal). For example, the neural network chip may receive frame data included in a video stream as input data, and generate a recognition signal of an object included in an image represented by the frame data from the frame data. In other example embodiments, the neural network chip may receive various types of input data according to the type or function of the electronic device on which the electronic device is installed, and may generate a recognition signal according to the input data.
[0099] The memory is a storage location for storing data, and can store an operating system (OS), various programs, and various data. In one embodiment, the memory can store intermediate results generated during the computational processing of the neural network chip.
[0100] The memory may be a DRAM, etc. The memory may include at least one of a volatile memory or a non-volatile memory. Non-volatile memories include ROM, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), flash memory, phase change random access memory (PRAM), magnetoresistive RAM (MRAM), resistive RAM (RRAM), ferroelectric RAM (FRAM), etc. Volatile memories include DRAM, SRAM, SDRAM, PRAM, MRAM, RRAM, FeRAM, etc. In one embodiment, the memory may include at least one of a HDD, SSD, CF, SD, micro SD, mini SD, xD, and a memory stick.
[0101] The sensor may collect peripheral information of the electronic device on which the electronic device is mounted. The sensor may sense or receive a signal (e.g., a video signal, an audio signal, a magnetic signal, a bio-signal, a touch signal, etc.) from the outside of the electronic device and convert the sensed or received signal into data. For example, the sensor may include at least one of various types of sensing devices (such as a microphone, an imaging device, an image sensor, a light detection and ranging (LIDAR) sensor, an ultrasonic sensor, an infrared sensor, a biosensor, and a touch sensor).
[0102] The sensor may provide the converted data as input data to the neural network chip. For example, the sensor may include an image sensor that generates a video stream by photographing the external environment of the electronic device, and sequentially provides continuous data frames of the video stream to the neural network chip as input data. In other example embodiments, the sensor may provide various types of data to the neural network chip.
[0103] The communication module may be provided with various wired or wireless interfaces capable of communicating with an external device. For example, the communication module may include a communication interface capable of accessing a wired local area network (LAN), a wireless local area network (WLAN) (such as wireless fidelity (Wi-Fi)), a wireless personal area network (WPAN) (such as Bluetooth), a wireless universal serial bus (wireless USB), Zigbee, near field communication (NFC), radio frequency identification (RFID), power line communication (PLC) or a mobile cellular network (such as third generation (3G), fourth generation (4G) and long term evolution (LTE)).
[0104] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.
[0105] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.
[0106] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
Claims
1. A memristor synaptic circuit, characterized in that: include: A signal transmission module, used for transmitting a pulse signal, the pulse signal enters the memristor from the signal input end of the memristor and is output from the signal output end of the memristor; An excitation / inhibition module is connected to the two ends of the memristor and is used to apply opposite charges to the two ends of the memristor according to a synaptic excitation signal or a synaptic inhibition signal; wherein when the synaptic excitation signal is connected, the memristor's memristor value decreases; when the synaptic inhibition signal is connected, the memristor's memristor value increases.
2. The memristor synaptic circuit according to claim 1, characterized in that: The signal transmission module includes a memristor, a first transistor and a second transistor for controlling the transmission of a pulse signal; the first end of the first transistor is connected to the pulse signal, the second end of the first transistor is connected to the signal input end of the memristor, and the control end thereof is connected to the input control signal; the first end of the second transistor is connected to the signal output end of the memristor, the second end of the second transistor is connected to the signal output end, and the control end thereof is connected to the output control signal; Preferably, the output control signal and the input control signal are the same control signal; The first transistor and the second transistor are both N-type MOS transistors or both P-type MOS transistors.
3. The memristor synaptic circuit according to claim 1, characterized in that: The excitation / inhibition module includes a constant first voltage terminal and a constant second voltage terminal, wherein the constant first voltage terminal is connected to the signal input terminal of the memristor through an excitation signal transistor, the signal output terminal of the memristor is connected to the constant second voltage terminal through an excitation signal transistor, and a synaptic excitation signal is applied to both ends of the memristor; the constant second voltage terminal is connected to the signal input terminal of the memristor through an inhibition signal transistor, the signal output terminal of the memristor is connected to the constant first voltage terminal through an inhibition signal transistor, and a synaptic inhibition signal is applied to both ends of the memristor; the constant first voltage terminal has a higher voltage than the constant second voltage terminal.
4. The memristor synaptic circuit according to claim 1, characterized in that: The control end of the inhibition signal transistor is connected to the inhibition control signal. When the inhibition control signal is applied, the conduction of the inhibition signal transistor is controlled, and a synaptic inhibition signal is applied at both ends of the memristor; the control end of the excitation signal transistor is connected to the excitation control signal. When the excitation control signal is applied, a synaptic excitation signal is applied at both ends of the memristor.
5. The memristor synaptic circuit according to claim 1, characterized in that: The signal transmission module further comprises a unidirectional control submodule connected to the signal output end of the memristor, and used for controlling the forward bias output or reverse bias output of the signal transmission module; Preferably, the memristor synaptic circuit is a forward memristor synaptic circuit, the unidirectional control submodule is a forward biased diode, the P end of the forward biased diode is connected to the signal output end of the memristor, and the N end is connected to the output; Preferably, the memristor synapse circuit is a reverse memristor synapse circuit, the unidirectional control submodule is a reverse biased diode, the N end of the reverse biased diode is connected to the signal output end of the memristor, and the P end is connected to the output.
6. An artificial neuron, characterized in that: The artificial neuron includes an input unit, a storage and calculation unit, a comparison unit, and an output unit; The input unit comprises a plurality of memristor synaptic circuits according to any one of claims 1 to 5, and is used to obtain a pulse output signal transmitted on each memristor synaptic circuit; The storage and calculation unit includes a storage module and a calculation module, wherein the storage module is connected to each memristor synaptic circuit of the input unit and is used to store the pulse output signal of the memristor synaptic circuit; the calculation module is used to calculate the pulse output signals of all memristor synaptic circuits on the input unit stored in the storage module to obtain a total output signal; A comparison unit, used to determine the control conduction of the pulse signal in the output unit according to the comparison between the total output signal and the preset signal; An output unit comprises a plurality of memristor synaptic circuits according to any one of claims 1 to 5, and is used to control the forward bias output or reverse bias output of a pulse signal.
7. The artificial neuron according to claim 6, characterized in that The input unit includes n forward memristor synaptic circuits and m reverse memristor synaptic circuits, wherein n and m are integers and are not zero at the same time; Preferably, the pulse output signal of the memristor synaptic circuit includes a forward bias pulse output signal and a reverse bias pulse output signal; Preferably, the storage module is used to store the forward pulse output signals and reverse bias pulse output signals of all memristor synaptic circuits of the input unit.
8. The artificial neuron according to claim 6, characterized in that: When the artificial neuron is a positive artificial neuron, the output unit is p positive memristor synaptic circuits; when the artificial neuron is a negative artificial neuron, the output unit is q reverse memristor synaptic circuits, wherein p and q are natural numbers.
9. A neural network system, characterized in that: It includes a central module and a signal transmission link connected to the central module. The central module is used to send a neuron excitation signal or a neuron inhibition signal to each artificial neuron in the signal transmission link according to the transmission result of the pulse signal after being output through the signal transmission link; The signal transmission link is formed by cascading x levels of artificial neurons described in any one of claims 6 to 8 in sequence, wherein the output unit of the first-level artificial neuron outputs a pulse signal, which is used as the pulse signal input of the input unit of the second-level artificial neuron, and the two are cascaded in sequence, and the output unit of the last-level artificial neuron outputs a pulse signal to the central module, wherein x is a natural number; wherein the neuron excitation signal or the neuron inhibition signal can enhance or inhibit the signal transmission of the artificial neuron.
10. The neural network system according to claim 9, characterized in that There are y signal transmission links, any one of which is formed by cascading x levels of artificial neurons described in any one of claims 6 to 8 in sequence, wherein the output unit of the first-level artificial neuron outputs a pulse signal, which is used as the pulse signal input of the input unit of the second-level artificial neuron, and the links are cascaded in sequence, and the output unit of the last-level artificial neuron outputs a pulse signal to the central module, wherein x and y are natural numbers.
Citation Information
Cited By
Speech recognition method based on memristor and electronic equipment
CN122337184A
