Deep neural network circuit based on adjustable memristor neurons and control method thereof

Abstract

The invention provides a deep neural network circuit based on adjustable memristor neurons and a control method thereof, and relates to the technical field of memristor application, the circuit comprises a multilayer neural network circuit and a control circuit; each layer of neural network circuit comprises a synaptic memristor array circuit and a neuron array circuit; the synaptic memristor array circuit comprises a plurality of rows and columns of synaptic memristors, and the synaptic memristors adjust signal transmission between the two threshold-adjustable neuron memristors according to stored weights in a neural network; the control circuit outputs a control signal to the neuron array circuit according to the current task mode; the neuron array circuit comprises a plurality of threshold-adjustable neuron memristors, and the threshold-adjustable neuron memristors respond to control signals and adjust the size of excitation thresholds of the neuron memristors and the size of discharge currents during excitation so as to achieve the current task mode. According to the scheme provided by the invention, the hardware cost is saved while the multi-modal task is realized.

Description

Technical Field

[0001] The present invention relates to the field of memristor application technology, and in particular to a deep neural network circuit based on adjustable memristor neurons and a control method thereof. Background Art

[0002] At the device level, as the development of Moore's Law slows down, the problem of the storage wall between the processor and memory in the von Neumann architecture becomes increasingly prominent, and the development of traditional computers gradually slows down. However, artificial intelligence (AI), big data, and cloud computing all require computers with powerful parallel computing capabilities. Neuromorphic computing simulates the behavior and connection of neurons in the human brain and is an efficient parallel computing method. Traditional circuits require a large number of capacitors, resistors, and transistors to realize neuronal behavior.

[0003] In some exemplary technologies, most biological neuron behaviors, such as action potential spikes, threshold-driven spikes, cumulative firing, and refractory period, are realized through volatile memristors. For bionic neurons, its stability and adjustability are important factors affecting performance. Currently, there is no work that can take into account both the single device integration and the instant adjustability of neuron behavior, so studying new memristor models will help improve the performance of memristors in neural networks.

[0004] At the neural network level, early fusion and late fusion are two strategies in multimodal learning. Early fusion refers to combining information from different modalities in the early stages of data processing for joint processing and analysis. The core idea of ​​this method is to merge data before or during the feature extraction stage, allowing the model to simultaneously learn the relationship between the various modalities and the features within each modality. Early fusion can improve the utilization of input features, thereby improving accuracy, and compared to late fusion, it does not require a preprocessing neural network for feature extraction for each modality input, so it greatly reduces hardware overhead. However, due to the large differences between the input features of different modalities, as the number of modalities increases, this will cause training difficulties, and the network will not be able to decipher the relationship between heterogeneous features.

[0005] To address this problem, a late fusion neural network was proposed. For each modality of data, an independent neural network is used to process it, and the information of different modalities is merged in the final stage of the model. This approach brings about the problem of increased hardware costs.

[0006] Therefore, how to save hardware costs while achieving multimodal tasks has become a technical problem that needs to be solved urgently. Summary of the invention

[0007] The present invention provides a deep neural network circuit based on adjustable memristor neurons and a control method thereof, which are used to solve the defect of high hardware cost in the prior art when realizing multimodal tasks, and save hardware costs while realizing multimodal tasks.

[0008] The present invention provides a deep neural network circuit based on adjustable memristor neurons, including a multi-layer neural network circuit and a control circuit; Each layer of neural network circuit includes a synaptic memristor array circuit and a neuron array circuit; The synaptic memristor array circuit includes synaptic memristors in multiple rows and columns, wherein the synaptic memristors store weights in a neural network, and the synaptic memristors are connected to two adjacent adjustable threshold neuron memristors, and the synaptic memristors are used to adjust the signal transmission between the two adjustable threshold neuron memristors according to the weights stored in the neural network; The control circuit is connected to the neuron array circuit, and the control circuit is used to output a control signal to the neuron array circuit according to the current task mode, wherein the control signal represents the current task mode; The neuron array circuit includes a plurality of adjustable threshold neuron memristors, wherein the adjustable threshold neuron memristors are used to adjust the magnitude of the excitation threshold of the adjustable threshold neuron memristors and the magnitude of the discharge current during excitation in response to the control signal, so as to realize the current task mode; the adjustable threshold neuron memristors are also used to generate a discharge current and output an activation signal when the input signal of the adjustable threshold neuron memristor is greater than the excitation threshold.

[0009] According to a deep neural network circuit based on an adjustable memristor neuron provided by the present invention, the control signal is a gate voltage; the adjustable threshold neuron memristor comprises: a first top electrode, a first insulating layer, a first bottom electrode comprising a two-dimensional material, a second insulating layer, a gate and a first substrate; The gate is located above the first substrate, the gate is connected to the control circuit, and the gate is used to adjust the magnitude of the excitation threshold of the adjustable threshold neuron memristor and the magnitude of the discharge current during excitation in response to the control signal; The second insulating layer is located above the gate and between the first bottom electrode; The first bottom electrode is located above the second insulating layer; The first insulating layer is located between the first bottom electrode and the first top electrode; The first top electrode is located above the first insulating layer.

[0010] According to a deep neural network circuit based on adjustable memristor neurons provided by the present invention, the synaptic memristor includes: a second top electrode, a third insulating layer, a second bottom electrode and a second substrate; The second bottom electrode is located above the second substrate; The third insulating layer is located above the second bottom electrode; The second top electrode is located above the third insulating layer.

[0011] According to a deep neural network circuit based on adjustable memristor neurons provided by the present invention, the synaptic memristor is used to adjust the signal transmission between two adjustable threshold neuron memristors according to the weights stored in the neural network, specifically used for: The synaptic memristor implements a multiplication algorithm of the synaptic memristor input value and the weight stored in the synaptic memristor by calculating the product of the voltage value of the signal received by the second top electrode and the conductance value of the synaptic memristor.

[0012] According to a deep neural network circuit based on adjustable memristor neurons provided by the present invention, the first insulating layer and the second insulating layer are metal oxide materials of 5-30 nanometers; The first top electrode and the first bottom electrode include conductive metals such as Ti, Pt or Ag.

[0013] According to a deep neural network circuit based on adjustable memristor neurons provided by the present invention, the two-dimensional material is MoS2.

[0014] According to a deep neural network circuit based on adjustable memristor neurons provided by the present invention, it is characterized in that the third insulating layer is a metal oxide material of 5-30 nanometers; The second top electrode and the second bottom electrode include conductive metal Ti\Pt or Ag.

[0015] According to a deep neural network circuit based on adjustable memristor neurons provided by the present invention, the control circuit is also connected to the first layer of the neural network circuit, and the control circuit is also used to apply an electrical signal corresponding to the data to be processed to the first layer of the neural network circuit.

[0016] The present invention also provides a control method for a deep neural network circuit based on adjustable memristor neurons, which is applied to a control circuit in a deep neural network circuit based on adjustable memristor neurons; The deep neural network circuit based on adjustable memristor neurons includes a multi-layer neural network circuit and a control circuit; each layer of the neural network circuit includes a synaptic memristor array circuit and a neuron array circuit; the synaptic memristor array circuit includes multiple rows and columns of synaptic memristors, and the synaptic memristors store weights in the neural network; the neuron array circuit includes multiple adjustable threshold neuron memristors; The method comprises: The control circuit outputs a control signal to the neuron array circuit according to the current task mode, wherein the control signal represents the current task mode.

[0017] According to a control method of a deep neural network circuit based on adjustable memristor neurons provided by the present invention, the method further includes: The control circuit applies an electrical signal corresponding to the data to be processed to the first layer of the neural network circuit.

[0018] The present invention provides a deep neural network circuit based on adjustable memristor neurons and a control method thereof. The deep neural network circuit based on adjustable memristor neurons includes: a multi-layer neural network circuit and a control circuit; each layer of the neural network circuit includes a synaptic memristor array circuit and a neuron array circuit; the synaptic memristor array circuit includes a plurality of rows and columns of synaptic memristors, the synaptic memristors store weights in the neural network, and the synaptic memristors adjust the signal transmission between two adjustable threshold neuron memristors according to the stored weights in the neural network; the control circuit is used to output a control signal to the neuron array circuit according to the current task mode, and the control signal represents the current task mode; the neuron array circuit includes a plurality of adjustable threshold neuron memristors, and the adjustable threshold neuron memristors are used to adjust the magnitude of the excitation threshold of the adjustable threshold neuron memristors and the magnitude of the discharge current during excitation in response to the control signal to achieve the current task mode. In the solution of the present invention, different multimodal tasks can be adapted by adjusting the excitation threshold of the adjustable threshold neuron memristor and the discharge current during excitation, that is, multimodal tasks can be achieved through a single deep neural network circuit, which can reduce the hardware cost of the neural network and save hardware costs compared to the traditional late-fusion deep neural network. Furthermore, the synaptic memristor has the characteristics of storage and computing in one, which reduces data handling, so that its characteristics can be used to complete the task of deep neural network reasoning based on very little hardware, saving the power consumption of the deep neural network circuit. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the present invention or 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 some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0020] Figure 1 It is a structural schematic diagram of a deep neural network circuit based on adjustable memristor neurons provided by the present invention.

[0021] Figure 2 It is a schematic diagram of the structure of the adjustable threshold neuron memristor provided by the present invention.

[0022] Figure 3 This is a light microscope image of the adjustable threshold neuron memristor provided by the present invention.

[0023] Figure 4 It is a box-type statistical diagram of the relationship between gate voltage and discharge current of a large number of neuron memristors provided by the present invention.

[0024] Figure 5 It is a flow chart of a control method of a deep neural network circuit based on adjustable memristor neurons provided by the present invention. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] The technical solution of the present invention is used to solve the above technical problems. The technical solution of the present application and how the technical solution of the present application solves the above technical problems are described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. Figure 1-Figure 4 The present invention describes a deep neural network circuit based on tunable memristor neurons.

[0027] Figure 1 is a schematic diagram of the structure of a deep neural network circuit based on adjustable memristor neurons provided by the present invention, such as Figure 1 As shown, the deep neural network circuit based on adjustable memristor neurons includes a multi-layer neural network circuit 10 and a control circuit 20.

[0028] like Figure 1As shown, each layer of the neural network circuit 10 includes a synaptic memristor array circuit 101 and a neuron array circuit 102 .

[0029] Among them, the synaptic memristor array circuit 101 includes multiple rows and columns of synaptic memristors, the synaptic memristors store weights in the neural network, the synaptic memristors connect two adjacent adjustable threshold neuron memristors, and the synaptic memristors are used to adjust the signal transmission between the two neuron memristors according to the weights stored in the neural network.

[0030] The control circuit 20 is connected to the neuron array circuit 102 , and is used to output a control signal to the neuron array circuit 102 according to the current task mode, wherein the control signal represents the current task mode.

[0031] Among them, the neuron array circuit 102 includes a plurality of adjustable threshold neuron memristors, which are used to respond to control signals, adjust the size of the excitation threshold of the adjustable threshold neuron memristor and the size of the discharge current during excitation to achieve the current task mode; the adjustable threshold neuron memristor is also used to generate a discharge current and output an activation signal when the input signal of the adjustable threshold neuron memristor is greater than the excitation threshold.

[0032] It can be understood that the multi-layer neural network circuit 10 is a deep neural network topology composed of multiple synaptic memristors and multiple adjustable threshold neuron memristors. The deep neural network topology is used to combine these two memristors, and the first half of each layer of the neural network circuit 10 is a synaptic memristor array circuit 101 with multiple rows and columns, and the second half is a plurality of neuron array circuits 102 with adjustable gate voltages.

[0033] The synaptic memristor array circuit 101 includes multiple rows and columns of synaptic memristors, and the synaptic memristors store weights in the neural network. Specifically, the synaptic memristor is an electronic device that simulates the function of synapses in biological nervous systems, that is, transmitting and processing signals between neurons. In this embodiment, the synaptic memristors store weights in the neural network, which determine how input signals are propagated and combined in the network.

[0034] Exemplarily, the synaptic memristor may be a non-volatile synaptic memristor, which has the ability to maintain its resistance state even after the power is turned off. This property is similar to the long-term potentiation (LTP) and long-term depression (LTD) of synapses in biological nervous systems, where the strength of a synapse can continuously change based on previous activity history.

[0035] In the present invention, the synaptic memristor has the characteristics of storage and calculation, that is, the synaptic memristor integrates storage and calculation functions, allowing calculation to be performed while storing weights, without the need to frequently read weights. The weights are stored in the synaptic memristor and participate in the calculation, thereby improving the calculation efficiency and reducing power consumption. This method provides new possibilities for realizing efficient and low-power neural networks.

[0036] The following examples illustrate the role and application process of the weights stored in the synaptic memristor. In one example, the synaptic memristor can weight the input signal according to the stored weights to reflect the importance of different input features. In the neural network, each input signal is multiplied by its corresponding weight. This operation simulates the weighted effect of synapses in biological neural systems. In another example, the weights determine how different input features are combined to activate neurons. By adjusting the weights, the network can learn how to extract complex features from the input data. In other examples, during the training process, the weights are adjusted to minimize the difference between the network output and the target output. This process, which is implemented through backpropagation algorithms and other optimization techniques, enables the network to learn and adapt to the training data. In other examples, the weights stored in the synaptic memristor are the conductance values ​​of the synaptic memristors. The weights stored in different synaptic memristors in the synaptic memristor array circuit 101 are dynamically adjusted to adapt to the changing input data and task requirements.

[0037] It is understandable that the deep neural network circuit provided in the present invention supports multiple task modalities. Exemplarily, the task modality can be a visual modality, involving the processing of image or video data, such as object recognition, scene understanding, etc. The task modality can also be an auditory modality, involving the processing of sound or voice data, such as speech recognition, music analysis, etc. The task modality can also be a tactile modality, involving the processing of tactile feedback or force data, such as robot tactile perception. The task modality can also be a text modality, involving the processing of text data, such as natural language processing, sentiment analysis, etc. The task modality can also be a biomedical modality, involving the processing of biomedical signals, such as electroencephalogram (EEG), electrocardiogram (ECG) data analysis, etc. The task modality can also be an environmental modality, involving the processing of environmental data, such as sensor data such as temperature, humidity, and air pressure.

[0038] Specifically, the function of the control circuit 20 is to provide the necessary control signals to the neuron array circuit 102 to adjust the size of the excitation threshold of the adjustable threshold neuron memristor and the size of the discharge current during excitation to adapt it to the current task mode. In practical applications, the control circuit 20 can identify the current task mode. The control circuit 20 generates a corresponding control signal according to the current task mode, and the control signal characterizes the current task mode, and the control signal includes the parameters required to adjust the excitation threshold of the adjustable threshold neuron memristor and the discharge current during excitation. As an example, the gate voltage of the adjustable threshold neuron memristor can be adjusted to adjust the excitation threshold and the discharge current during excitation, and the control signal includes the gate voltage corresponding to each adjustable threshold neuron memristor.

[0039] In this embodiment, there is no specific limitation on the implementation of the control circuit 20. For example, the control circuit 20 may be a microcontroller unit (MCU), a digital signal processor (DSP), a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), an integrated chip (IC), a server, etc.

[0040] Among them, the adjustable threshold neuron memristor is an electronic device that combines the physical properties of a memristor with the dynamic behavior of a biological neuron, in particular, an adjustable activation threshold. This memristor is designed to simulate the activation characteristics of biological neurons, where the activation threshold of the neuron can be dynamically adjusted in response to different input signals and network states. For example, the adjustable threshold neuron memristor can be a volatile neuron memristor with an adjustable threshold.

[0041] In the present invention, by adjusting the size of the excitation threshold of the adjustable threshold neuron memristor and the size of the discharge current during excitation, different task modes can be achieved, the attention mechanism of the brain can be simulated, and a neural network that can efficiently process multimodal tasks can be obtained using fewer synapses.

[0042] Among them, the excitation threshold refers to the critical point of activation of the adjustable threshold neuron memristor. In biological neurons, the excitation threshold usually refers to the minimum voltage value reached by the cell membrane potential. When this value is exceeded, the neuron will generate an action potential (or "nerve impulse", "spike"). In the adjustable threshold neuron memristor, the excitation threshold can refer to the minimum voltage or current required to switch the memristor from a high resistance state to a low resistance state. It can be understood that the excitation threshold determines whether the adjustable threshold neuron memristor is activated under a given input signal. Only when the voltage or current corresponding to the input signal exceeds this threshold, the adjustable threshold neuron memristor will change its state and produce an output.

[0043] Specifically, by adjusting the firing threshold, the response sensitivity of the neural network to the input signal can be controlled. A lower threshold makes it easier for the neuron to be activated, while a higher threshold requires a stronger input signal to activate the neuron. The adjustable threshold neuron memristor simulates the dynamic threshold characteristics of biological neurons, where the activation threshold can be adjusted based on the previous activity of the neuron and the needs of the network. By adjusting the firing threshold, the network can flexibly switch between different input signals and task requirements without redesign or retraining. Therefore, the adjustment of the firing threshold is the key to realizing efficient, flexible and adaptable neuromorphic computing systems. By precisely controlling the firing threshold, the neuron memristor can better simulate the behavior of biological neurons and provide powerful computing power for various artificial intelligence applications.

[0044] In practical applications, the magnitude of the discharge current during excitation can reflect the activation strength of the adjustable threshold neuron memristor. The magnitude of the discharge current will affect the activation state of subsequent neuron memristors or circuits. A larger discharge current may lead to stronger signal transmission, thereby propagating farther in the network or affecting more neurons.

[0045] In the present invention, there is no specific limitation on the adjustment method of the excitation threshold size and discharge current size corresponding to the adjustable threshold neuron memristor. In one example, the gate voltage of the adjustable threshold neuron memristor is adjusted to achieve the adjustment of the excitation threshold size and discharge current size corresponding to the adjustable threshold neuron memristor. In another example, by changing the material properties used in the adjustable threshold neuron memristor, such as the doping level or the conductivity of the material, its conductivity state and threshold voltage can be changed. In another example, by temperature regulation, the change in temperature can affect the electron mobility of the adjustable threshold neuron memristor material, thereby changing its conductivity and threshold voltage. In another example, the operating point of the neuron array circuit 102 is dynamically adjusted by a software algorithm, thereby indirectly adjusting the excitation threshold size and discharge current size corresponding to the adjustable threshold neuron memristor. It can be understood that for different adjustment methods, the control signal output by the above-mentioned control circuit 20 is in a corresponding form, which will not be described in detail here.

[0046] It can be understood that adjusting the excitation threshold and discharge current of the adjustable threshold neuron memristor can realize the current task mode, and this adjustment mechanism can adapt the deep neural network circuit to different input characteristics and processing requirements. On the one hand, different task modes (such as vision, hearing, touch, etc.) have different data characteristics and statistical characteristics. By adjusting the excitation threshold and discharge current of the adjustable threshold neuron memristor, the deep neural network circuit can better respond to these specific characteristics. On the other hand, the deep neural network circuit needs to dynamically adapt to different input data and task requirements. The adjustable threshold neuron memristor provides a mechanism that allows the network to adjust its parameters at runtime to best handle the current task mode. On the other hand, by adjusting the excitation threshold and discharge current of the adjustable threshold neuron memristor, the activation mode of the neuron can be changed. For example, for task modes that require high sensitivity, the threshold can be lowered to make the neuron more easily activated; while for task modes that need to suppress noise or unimportant information, the threshold can be increased to reduce false activation. On the other hand, by adjusting the excitation threshold and discharge current of the adjustable threshold neuron memristor, biological neurons can be simulated, which have variable thresholds and discharge characteristics, which are essential for the brain to process different types of information. The adjustable threshold neuron memristor mimics these biological characteristics, allowing artificial neural networks to process multimodal data more naturally. Therefore, the scheme in this application can adapt to different task modalities and improve the flexibility, adaptability and performance of the network.

[0047] It should be noted that the deep neural network circuit of the present invention has a high degree of computational parallelism, and the computation time does not increase with the complexity of the input signal, which can save a lot of time compared to traditional transistor circuits.

[0048] In the solution of this embodiment, the excitation threshold of the adjustable threshold neuron memristor and the discharge current during excitation can be adapted to different multimodal tasks, that is, a single deep neural network circuit can realize multimodal tasks, which can reduce the hardware cost of the neural network and save hardware costs compared to the traditional late-fusion deep neural network. Furthermore, the synaptic memristor has the characteristics of storage and computing in one, which reduces data handling, so that its characteristics can be used to complete the task of deep neural network reasoning based on very little hardware, saving the power consumption of the deep neural network circuit.

[0049] In addition, in a possible implementation, the control circuit 20 is also connected to the first-layer neural network circuit 10, and the control circuit 20 is also used to apply an electrical signal corresponding to the data to be processed to the first-layer neural network circuit 10.

[0050] In practical applications, the control circuit 20 applies the electrical signal of the data to be processed to the first layer neural network circuit 10, and outputs the control signal for the modality of the current task, adjusts the excitation threshold of the adjustable threshold neuron memristor and the discharge current during excitation, and realizes the current task modality. The electrical signal of the data to be processed is processed by the weight of the multi-layer neural network circuit 10 to obtain the inference result corresponding to the data to be processed.

[0051] Optionally, in one embodiment, the control signal is a gate voltage; the adjustable threshold neuron memristor comprises: a first top electrode, a first insulating layer, a first bottom electrode comprising a two-dimensional material, a second insulating layer, a gate and a first substrate.

[0052] Specifically, the gate is located above the first substrate, the gate is connected to the control circuit, and the gate is used to adjust the magnitude of the excitation threshold of the adjustable threshold neuron memristor and the magnitude of the discharge current during excitation in response to the control signal.

[0053] The second insulating layer is located above the gate and between the first bottom electrode.

[0054] The first bottom electrode is located above the second insulating layer.

[0055] The first insulating layer is located between the first bottom electrode and the first top electrode.

[0056] The first top electrode is located above the first insulating layer.

[0057] In this embodiment, the adjustable threshold neuron memristor is a three-port device, and its gate can adjust the magnitude of the excitation threshold of the memristor and the magnitude of the discharge current during excitation.

[0058] On the one hand, adjusting the gate voltage can adjust the size of the excitation threshold. Specifically, the change in gate voltage will affect the distribution of charge in the adjustable threshold neuron memristor, thereby changing its conductance state. This change in conductance state will directly affect the voltage required for the memristor to switch from a high resistance state (OFF) to a low resistance state (ON), that is, the threshold voltage. By adjusting the gate voltage, this threshold voltage can be increased or decreased, thereby achieving control over the neuron activation threshold.

[0059] On the other hand, adjusting the gate voltage can adjust the size of the discharge current during excitation. In practical applications, when the conductance state of the adjustable threshold neuron memristor changes, the discharge current generated during excitation will also change. Specifically, if the gate voltage increases the conductance of the adjustable threshold neuron memristor, then at the same input voltage, the discharge current through the adjustable threshold neuron memristor will also increase. Conversely, if the gate voltage decreases the conductance, the discharge current will also decrease. This regulation mechanism allows the adjustable threshold neuron memristor to simulate the discharge behavior of biological neurons, where the size of the discharge current can reflect the activation strength of the neuron.

[0060] Specifically, the second insulating layer is used to electrically isolate the gate and the first bottom electrode, prevent short circuit, and allow the gate voltage to affect the charge distribution in the first bottom electrode. The first bottom electrode is used to receive and integrate input signals and transmit output signals when the adjustable threshold neuron memristor is activated. The first insulating layer is used to electrically isolate the first bottom electrode and the first top electrode, prevent short circuit, and provide a medium for the electric field between the first top electrode and the first bottom electrode. The first top electrode is used to apply a voltage signal and adjust the conductance state of the adjustable threshold neuron memristor.

[0061] The adjustable threshold neuron memristor of the present invention integrates the functions of neuron leakage, integration, firing, etc., and adds an additional gate for adjusting the excitability of the neuron, that is, the threshold value and the discharge current. The present invention uses the gate-controlled threshold to simulate the attention mechanism, and achieves the recognition accuracy of each mode close to that of late fusion on the basis of reducing the number of synapses, which means a reduction in hardware cost and circuit power consumption.

[0062] In this embodiment, by adjusting the gate voltage, the activation threshold and discharge current of the adjustable threshold neuron memristor can be changed, thereby achieving precise control of the behavior of the neural network and improving the flexibility, efficiency and reliability of the deep neural network circuit.

[0063] In addition, in one embodiment, the first insulating layer and the second insulating layer are made of a metal oxide material with a thickness of 5-30 nanometers; the first top electrode and the first bottom electrode include conductive metals such as Ti\Pt or Ag.

[0064] In addition, in one embodiment, the above-mentioned two-dimensional material is MoS2.

[0065] Specifically, Figure 2 Schematic diagram of the structure of the adjustable threshold neuron memristor provided by the present invention. Figure 2 The specific implementation methods and functions of the adjustable threshold neuron memristor in the embodiments of the present invention are described respectively.

[0066] Reference Figure 2 The present invention proposes an adjustable threshold bionic neuron device based on a gate-controlled memristor. A gate is added to the top electrode, insulating layer and bottom electrode of the traditional memristor, and the bottom electrode is modified to a two-dimensional material / bottom electrode. The ultra-thin and electric field-penetrating properties of the two-dimensional material ensure that the gate can control the growth of conductive filaments in the insulating layer, and the conductivity of the two-dimensional material ensures very good electrical connectivity between the bottom electrode and it. When charges accumulate in the top electrode and the voltage increases, conductive filaments will grow in the insulating layer. The voltage of the back gate can affect the thickness and number of the conductive silver filaments. When the silver filaments connect the top electrode and the bottom electrode, they will be instantly turned on, forming a current pulse. The silver filaments are quickly cut off and wait for the next pulse.

[0067] The preparation process of the adjustable threshold bionic neuron memristor of the present invention comprises the following steps: first, Ti / Pt with a thickness of 2nm / 35nm is selected as the gate metal 206 by photolithography and electron beam evaporation (EVP) technology. Then, a 10nm thick HfO2 layer 205 is deposited by atomic layer deposition at a temperature of 200°C as a gate dielectric. Then, MoS2204 grown by metal organic chemical vapor deposition (MOCVD) on a sapphire substrate is wet transferred and patterned by photolithography on the HfO2 substrate. Then, Ti / Pt203 with a thickness of 5nm / 35nm is made by photolithography and electron beam evaporation (EVP) technology as the contact metal of MoS2, that is, the bottom electrode metal. Next, a 10nm thick HfO2 layer 202 is deposited by atomic layer deposition at a temperature of 200°C. Finally, 40nm Ag201 is deposited as the top electrode metal of the memristor.

[0068] In addition, in one embodiment, the synaptic memristor includes: a second top electrode, a third insulating layer, a second bottom electrode and a second substrate.

[0069] Wherein, the second bottom electrode is located above the second substrate.

[0070] The third insulating layer is located above the second bottom electrode.

[0071] The second top electrode is located above the third insulating layer.

[0072] Specifically, the second top electrode is the upper contact point of the synaptic memristor structure, which is used to apply a voltage signal or receive a signal from the output of the upper network layer. In a neural network, the top electrode can be regarded as an input terminal, which is responsible for transmitting input data to the synaptic memristor. The third insulating layer is located between the second top electrode and the second bottom electrode, and its main function is to electrically isolate the top electrode and the bottom electrode to prevent a short circuit between them. The material and thickness of the third insulating layer will affect the conductivity state of the synaptic memristor and its response characteristics to voltage changes. The second bottom electrode is the lower contact point of the synaptic memristor structure, which is in contact with the two-dimensional material layer or formed by the two-dimensional material. The second bottom electrode plays the role of collecting or transmitting current in the synaptic memristor, and participates in forming the conductivity channel of the memristor together with the second top electrode. The second bottom electrode is also responsible for transmitting the weight-adjusted signal to subsequent neurons or other circuits. The second substrate is the base of the entire synaptic memristor structure, providing physical support for other components. The second substrate is usually made of insulating material to ensure electrical isolation and maintain the stability of the entire structure.

[0073] In practical applications, the voltage difference between the second top electrode and the second bottom electrode causes the charge in the third insulating layer to be redistributed, thereby changing the conductance state of the synaptic memristor. This change in conductance state reflects the adjustment of weights and is a key mechanism in the learning process of neural networks. By changing the amount of charge passing through the synaptic memristor, its conductance value can be adjusted to simulate the change of synaptic weights. This structural design enables synaptic memristors to simulate the functions of biological synapses, store and process information by adjusting their conductance values, and provides a basis for building deep neural networks and other artificial intelligence systems.

[0074] Optionally, in a possible implementation, when the above-mentioned synaptic memristor is used to adjust the signal transmission between two neuron memristors according to the weights stored in the neural network, it is specifically used for: the synaptic memristor calculates the product of the voltage value of the signal received by the second top electrode and the conductance value of the synaptic memristor to implement the multiplication algorithm of the synaptic memristor input value and the weight stored in the synaptic memristor.

[0075] Specifically, the signal received by the second top electrode is usually expressed in the form of voltage and applied to the second top electrode of the synaptic memristor. This voltage can represent the output from the neurons in the previous layer, or the input data directly obtained from the external environment. The conductance value of the synaptic memristor represents the weight in the neural network. These weights are adjusted by the learning algorithm during the training process and stored in the memristor, reflecting the importance of the input signal or the degree of influence on subsequent processing.

[0076] In practical applications, when the input voltage is applied to the synaptic memristor, according to Ohm's law, the current through the synaptic memristor I With applied voltage Vand the conductance of the synaptic memristor G is proportional to This current I It can be considered as the product of the input voltage and the conductance value (weight), thus realizing the multiplication operation.

[0077] Furthermore, the neuron memristor receives the output currents (weighted input signals) from multiple synaptic memristors and integrates the output currents of multiple synaptic memristors. The integrated current determines whether the neuron is activated.

[0078] In addition, in one embodiment, the third insulating layer is a metal oxide material with a thickness of 5-30 nanometers; the second top electrode and the second bottom electrode include conductive metals Ti\Pt or Ag.

[0079] In order to illustrate the effectiveness of the scheme in the invention, the following is a description in conjunction with a specific embodiment. Specifically, the present invention first aligns the data of multiple modes to ensure the same input dimension. The weight values ​​trained in advance are set to the size of the conductance value of each synaptic memristor by set / reset. Each mode corresponds to a gate voltage condition of all neurons. For example, when visual data is input, half of the neurons are under high gate voltage and high excitation, and the other half of the neurons are under low gate voltage and low excitation. Then the data of multiple modes are input into the neural network in the form of voltage under the corresponding gate voltage conditions. In each layer of the neural network, the input data voltage value and the synaptic weight conductance value automatically complete the multiplication operation. The neuron collects the current of the synapse in front and transmits electrical signals of different sizes according to its own excitation in its own gate voltage mode. Finally, the electrical signal is transmitted to the output layer, and the serial number of the neuron with the largest output electrical signal is considered to be the result of neural network reasoning.

[0080] Figure 3 This is a light microscope image of the adjustable threshold neuron memristor provided by the present invention. Figure 4 is a box-type statistical diagram of the relationship between gate voltage and discharge current of a large number of neuron memristors provided by the present invention, Figure 4 It is shown that the discharge current increases with the increase of gate voltage, and there is a current difference of nearly 100 times between the highest gate voltage and the lowest gate voltage.

[0081] The deep neural network circuit based on adjustable memristor neurons provided in this embodiment includes: a multi-layer neural network circuit and a control circuit; each layer of the neural network circuit includes a synaptic memristor array circuit and a neuron array circuit; the synaptic memristor array circuit includes multiple rows and columns of synaptic memristors, the synaptic memristors store weights in the neural network, and the synaptic memristors adjust the signal transmission between two adjustable threshold neuron memristors according to the stored weights in the neural network; the control circuit is used to output a control signal to the neuron array circuit according to the current task mode, and the control signal represents the current task mode; the neuron array circuit includes multiple adjustable threshold neuron memristors, and the adjustable threshold neuron memristors are used to respond to the control signal to adjust the size of the excitation threshold of the adjustable threshold neuron memristor and the size of the discharge current during excitation to achieve the current task mode. In the solution of this embodiment, the excitation threshold of the adjustable threshold neuron memristor and the discharge current during excitation can be adapted to different multimodal tasks, that is, a single deep neural network circuit can realize multimodal tasks, which can reduce the hardware cost of the neural network and save hardware costs compared to the traditional late-fusion deep neural network. Furthermore, the synaptic memristor has the characteristics of storage and computing in one, which reduces data handling, so that its characteristics can be used to complete the task of deep neural network reasoning based on very little hardware, saving the power consumption of the deep neural network circuit.

[0082] In addition, in some embodiments, the present invention also provides a control method for a deep neural network circuit based on adjustable memristor neurons, Figure 5 is a flow chart of a control method of a deep neural network circuit based on adjustable memristor neurons provided by the present invention, such as Figure 5 As shown, the control method of the deep neural network circuit based on adjustable memristor neurons includes step 601 and step 602.

[0083] Step 601: The control circuit outputs a control signal to the neuron array circuit according to the current task mode, where the control signal represents the current task mode.

[0084] Step 602: The control circuit applies an electrical signal corresponding to the data to be processed to the first layer neural network circuit.

[0085] The control method of the deep neural network circuit based on adjustable memristor neurons provided in this embodiment is applied to the control circuit 20 in the deep neural network circuit based on adjustable memristor neurons in the above embodiment. The deep neural network circuit based on adjustable memristor neurons includes a multi-layer neural network circuit 10 and a control circuit 20; each layer of the neural network circuit 10 includes a synaptic memristor array circuit 101 and a neuron array circuit 102; the synaptic memristor array circuit 101 includes multiple rows and columns of synaptic memristors, and the synaptic memristors store the weights in the neural network; the neuron array circuit 102 includes multiple adjustable threshold neuron memristors.

[0086] It can be understood that the multi-layer neural network circuit 10 is a deep neural network topology composed of multiple synaptic memristors and multiple adjustable threshold neuron memristors. The deep neural network topology is used to combine these two memristors, and the first half of each layer of the neural network circuit 10 is a synaptic memristor array circuit 101 with multiple rows and columns, and the second half is a plurality of neuron array circuits 102 that can change the gate voltage.

[0087] The synaptic memristor array circuit 101 includes multiple rows and columns of synaptic memristors, and the synaptic memristors store weights in the neural network. Specifically, the synaptic memristor is an electronic device that simulates the function of synapses in biological nervous systems, that is, transmitting and processing signals between neurons. In this embodiment, the synaptic memristors store weights in the neural network, which determine how input signals are propagated and combined in the network.

[0088] Exemplarily, the synaptic memristor may be a non-volatile synaptic memristor, which has the ability to maintain its resistance state even after the power is turned off. This property is similar to the long-term potentiation (LTP) and long-term depression (LTD) of synapses in biological nervous systems, where the strength of a synapse can continuously change based on previous activity history.

[0089] In the present invention, the synaptic memristor has the characteristics of storage and calculation, that is, the synaptic memristor integrates storage and calculation functions, allowing calculation to be performed while storing weights, without the need to frequently read weights. The weights are stored in the synaptic memristor and participate in the calculation, thereby improving the calculation efficiency and reducing power consumption. This method provides new possibilities for realizing efficient and low-power neural networks.

[0090] The following examples illustrate the role and application process of the weights stored in the synaptic memristor. In one example, the synaptic memristor can weight the input signal according to the stored weights to reflect the importance of different input features. In the neural network, each input signal is multiplied by its corresponding weight. This operation simulates the weighted effect of synapses in biological neural systems. In another example, the weights determine how different input features are combined to activate neurons. By adjusting the weights, the network can learn how to extract complex features from the input data. In other examples, during the training process, the weights are adjusted to minimize the difference between the network output and the target output. This process, which is implemented through backpropagation algorithms and other optimization techniques, enables the network to learn and adapt to the training data. In other examples, the weights stored in the synaptic memristor are the conductance values ​​of the synaptic memristors. The weights stored in different synaptic memristors in the synaptic memristor array circuit 101 are dynamically adjusted to adapt to the changing input data and task requirements.

[0091] It is understandable that the deep neural network circuit based on adjustable memristor neurons provided in the present invention supports multiple task modes. Exemplarily, the task mode can be a visual mode, involving the processing of image or video data, such as object recognition, scene understanding, etc. The task mode can also be an auditory mode, involving the processing of sound or voice data, such as speech recognition, music analysis, etc. The task mode can also be a tactile mode, involving the processing of tactile feedback or force data, such as robot tactile perception. The task mode can also be a text mode, involving the processing of text data, such as natural language processing, sentiment analysis, etc. The task mode can also be a biomedical mode, involving the processing of biomedical signals, such as electroencephalogram (EEG), electrocardiogram (ECG) data analysis, etc. The task mode can also be an environmental mode, involving the processing of environmental data, such as sensor data such as temperature, humidity, and air pressure.

[0092] Specifically, the function of the control circuit 20 is to provide the necessary control signals to the neuron array circuit 102 to adjust the size of the excitation threshold of the adjustable threshold neuron memristor and the size of the discharge current during excitation to adapt it to the current task mode. In practical applications, the control circuit 20 can identify the current task mode. The control circuit 20 generates a corresponding control signal according to the current task mode, and the control signal characterizes the current task mode, and the control signal includes the parameters required to adjust the excitation threshold of the adjustable threshold neuron memristor and the discharge current during excitation. As an example, the gate voltage of the adjustable threshold neuron memristor can be adjusted to adjust the excitation threshold and the discharge current during excitation, and the control signal includes the gate voltage corresponding to each adjustable threshold neuron memristor.

[0093] In this embodiment, there is no specific limitation on the implementation of the control circuit 20. For example, the control circuit 20 may be a microcontroller unit (MCU), a digital signal processor (DSP), a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), an integrated chip (IC), a server, etc.

[0094] Among them, the adjustable threshold neuron memristor is an electronic device that combines the physical properties of a memristor with the dynamic behavior of a biological neuron, especially an adjustable activation threshold. This memristor is designed to simulate the activation characteristics of a biological neuron, where the activation threshold of the neuron can be dynamically adjusted in response to different input signals and network states.

[0095] In the present invention, by adjusting the size of the excitation threshold of the adjustable threshold neuron memristor and the size of the discharge current during excitation, different task modes can be achieved, the attention mechanism of the brain can be simulated, and a neural network that can efficiently process multimodal tasks can be obtained using fewer synapses.

[0096] Among them, the excitation threshold refers to the critical point of activation of the adjustable threshold neuron memristor. In biological neurons, the excitation threshold usually refers to the minimum voltage value reached by the cell membrane potential. When this value is exceeded, the neuron will generate an action potential (or "nerve impulse", "spike"). In the adjustable threshold neuron memristor, the excitation threshold can refer to the minimum voltage or current required to switch the memristor from a high resistance state to a low resistance state. It can be understood that the excitation threshold determines whether the adjustable threshold neuron memristor is activated under a given input signal. Only when the voltage or current corresponding to the input signal exceeds this threshold, the adjustable threshold neuron memristor will change its state and produce an output.

[0097] Specifically, by adjusting the firing threshold, the response sensitivity of the neural network to the input signal can be controlled. A lower threshold makes it easier for the neuron to be activated, while a higher threshold requires a stronger input signal to activate the neuron. The adjustable threshold neuron memristor simulates the dynamic threshold characteristics of biological neurons, where the activation threshold can be adjusted based on the previous activity of the neuron and the needs of the network. By adjusting the firing threshold, the network can flexibly switch between different input signals and task requirements without redesign or retraining. Therefore, the adjustment of the firing threshold is the key to realizing efficient, flexible and adaptable neuromorphic computing systems. By precisely controlling the firing threshold, the neuron memristor can better simulate the behavior of biological neurons and provide powerful computing power for various artificial intelligence applications.

[0098] In practical applications, the magnitude of the discharge current during excitation can reflect the activation strength of the adjustable threshold neuron memristor. The magnitude of the discharge current will affect the activation state of subsequent neuron memristors or circuits. A larger discharge current may lead to stronger signal transmission, thereby propagating farther in the network or affecting more neurons.

[0099] In the present invention, there is no specific limitation on the adjustment method of the excitation threshold size and discharge current size corresponding to the adjustable threshold neuron memristor. In one example, the gate voltage of the adjustable threshold neuron memristor is adjusted to achieve the adjustment of the excitation threshold size and discharge current size corresponding to the adjustable threshold neuron memristor. In another example, by changing the material properties used in the adjustable threshold neuron memristor, such as the doping level or the conductivity of the material, its conductivity state and threshold voltage can be changed. In another example, by temperature regulation, the change in temperature can affect the electron mobility of the adjustable threshold neuron memristor material, thereby changing its conductivity and threshold voltage. In another example, the operating point of the neuron array circuit 102 is dynamically adjusted by a software algorithm, thereby indirectly adjusting the excitation threshold size and discharge current size corresponding to the adjustable threshold neuron memristor. It can be understood that for different adjustment methods, the control signal output by the above-mentioned control circuit 20 is in a corresponding form, which will not be described in detail here.

[0100] It can be understood that adjusting the excitation threshold and discharge current of the adjustable threshold neuron memristor can realize the current task mode, and this adjustment mechanism can adapt the deep neural network circuit to different input characteristics and processing requirements. On the one hand, different task modes (such as vision, hearing, touch, etc.) have different data characteristics and statistical characteristics. By adjusting the excitation threshold and discharge current of the adjustable threshold neuron memristor, the deep neural network circuit can better respond to these specific characteristics. On the other hand, the deep neural network circuit needs to dynamically adapt to different input data and task requirements. The adjustable threshold neuron memristor provides a mechanism that allows the network to adjust its parameters at runtime to best handle the current task mode. On the other hand, by adjusting the excitation threshold and discharge current of the adjustable threshold neuron memristor, the activation mode of the neuron can be changed. For example, for task modes that require high sensitivity, the threshold can be lowered to make the neuron more easily activated; while for task modes that need to suppress noise or unimportant information, the threshold can be increased to reduce false activation. On the other hand, by adjusting the excitation threshold and discharge current of the adjustable threshold neuron memristor, biological neurons can be simulated, which have variable thresholds and discharge characteristics, which are essential for the brain to process different types of information. The adjustable threshold neuron memristor mimics these biological characteristics, allowing artificial neural networks to process multimodal data more naturally. Therefore, the scheme in this application can adapt to different task modalities and improve the flexibility, adaptability and performance of the network.

[0101] In the solution of this embodiment, the excitation threshold of the adjustable threshold neuron memristor and the discharge current during excitation can be adapted to different multimodal tasks, that is, a single deep neural network circuit can realize multimodal tasks, which can reduce the hardware cost of the neural network and save hardware costs compared to the traditional late-fusion deep neural network. Furthermore, the synaptic memristor has the characteristics of storage and computing in one, which reduces data handling, so that its characteristics can be used to complete the task of deep neural network reasoning based on very little hardware, saving the power consumption of the deep neural network circuit.

[0102] In practical applications, the control circuit 20 applies the electrical signal of the data to be processed to the first layer neural network circuit 10, and outputs the control signal for the modality of the current task, adjusts the excitation threshold of the adjustable threshold neuron memristor and the discharge current during excitation, and realizes the current task modality. The electrical signal of the data to be processed is processed by the weight of the multi-layer neural network circuit 10 to obtain the inference result corresponding to the data to be processed.

[0103] The control method of the deep neural network circuit based on adjustable memristor neurons provided by the present invention is applied to the control circuit in the deep neural network circuit based on adjustable memristor neurons, and the deep neural network circuit based on adjustable memristor neurons includes: a multi-layer neural network circuit and a control circuit; each layer of the neural network circuit includes a synaptic memristor array circuit and a neuron array circuit; the synaptic memristor array circuit includes a plurality of rows and columns of synaptic memristors, and the synaptic memristors store weights in the neural network; the neuron array circuit includes a plurality of adjustable threshold neuron memristors; the method includes: the control circuit outputs a control signal to the neuron array circuit according to the current task mode, and the control signal represents the current task mode. In the scheme of the present invention, different multimodal tasks can be adapted by adjusting the excitation threshold of the adjustable threshold neuron memristor and the discharge current during excitation, that is, multimodal tasks can be realized by a single deep neural network circuit, which can reduce the hardware cost of the neural network and save hardware cost compared with the traditional late fusion deep neural network. Furthermore, synaptic memristors have the characteristics of integrated storage and computing, which reduces data movement. Therefore, their characteristics can be used to complete deep neural network reasoning tasks based on very little hardware, saving the power consumption of deep neural network circuits.

[0104] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0105] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep neural network circuit based on adjustable memristor neurons, characterized in that: include: Multi-layer neural network circuits and control circuits; Each layer of neural network circuit includes a synaptic memristor array circuit and a neuron array circuit; The synaptic memristor array circuit includes synaptic memristors in multiple rows and columns, wherein the synaptic memristors store weights in a neural network, and the synaptic memristors are connected to two adjacent adjustable threshold neuron memristors, and the synaptic memristors are used to adjust the signal transmission between the two adjustable threshold neuron memristors according to the weights stored in the neural network; The control circuit is connected to the neuron array circuit, and the control circuit is used to output a control signal to the neuron array circuit according to the current task mode, wherein the control signal represents the current task mode; The neuron array circuit comprises a plurality of adjustable threshold neuron memristors, wherein the adjustable threshold neuron memristors are used to adjust the magnitude of the excitation threshold of the adjustable threshold neuron memristors and the magnitude of the discharge current during excitation in response to the control signal, so as to realize the current task mode; The adjustable threshold neuron memristor is also used to generate a discharge current and output an activation signal when the input signal of the adjustable threshold neuron memristor is greater than the excitation threshold.

2. The deep neural network circuit based on adjustable memristor neurons according to claim 1, characterized in that: The control signal is a gate voltage; the adjustable threshold neuron memristor comprises: a first top electrode, a first insulating layer, a first bottom electrode comprising a two-dimensional material, a second insulating layer, a gate and a first substrate; The gate is located above the first substrate, the gate is connected to the control circuit, and the gate is used to adjust the magnitude of the excitation threshold of the adjustable threshold neuron memristor and the magnitude of the discharge current during excitation in response to the control signal; The second insulating layer is located above the gate and between the first bottom electrode; The first bottom electrode is located above the second insulating layer; The first insulating layer is located between the first bottom electrode and the first top electrode; The first top electrode is located above the first insulating layer.

3. The deep neural network circuit based on adjustable memristor neurons according to claim 1, characterized in that: The synaptic memristor comprises: a second top electrode, a third insulating layer, a second bottom electrode and a second substrate; The second bottom electrode is located above the second substrate; The third insulating layer is located above the second bottom electrode; The second top electrode is located above the third insulating layer.

4. The deep neural network circuit based on adjustable memristor neurons according to claim 3, characterized in that: When the synaptic memristor is used to adjust the signal transmission between two adjustable threshold neuron memristors according to the weight stored in the neural network, it is specifically used to: The synaptic memristor implements a multiplication algorithm of the synaptic memristor input value and the weight stored in the synaptic memristor by calculating the product of the voltage value of the signal received by the second top electrode and the conductance value of the synaptic memristor.

5. The deep neural network circuit based on adjustable memristor neurons according to claim 2, characterized in that: The first insulating layer and the second insulating layer are made of metal oxide material with a thickness of 5-30 nanometers; The first top electrode and the first bottom electrode include conductive metals such as Ti, Pt or Ag.

6. The deep neural network circuit based on adjustable memristor neurons according to claim 2, characterized in that: The two-dimensional material is MoS2.

7. The deep neural network circuit based on adjustable memristor neurons according to claim 3, characterized in that: The third insulating layer is a metal oxide material of 5-30 nanometers; The second top electrode and the second bottom electrode include conductive metal Ti\Pt or Ag.

8. The deep neural network circuit based on adjustable memristor neurons according to any one of claims 1 to 7, characterized in that: The control circuit is also connected to the first layer of the neural network circuit, and the control circuit is also used to apply an electrical signal corresponding to the data to be processed to the first layer of the neural network circuit.

9. A control method for a deep neural network circuit based on adjustable memristor neurons, characterized in that: Control circuits applied to deep neural network circuits based on tunable memristor neurons; The deep neural network circuit based on adjustable memristor neurons includes a multi-layer neural network circuit and a control circuit; each layer of the neural network circuit includes a synaptic memristor array circuit and a neuron array circuit; the synaptic memristor array circuit includes multiple rows and columns of synaptic memristors, and the synaptic memristors store weights in the neural network; The neuron array circuit includes a plurality of adjustable threshold neuron memristors; The method comprises: The control circuit outputs a control signal to the neuron array circuit according to the current task mode, wherein the control signal represents the current task mode.

10. The control method of the deep neural network circuit based on adjustable memristor neurons according to claim 9, characterized in that: The method further comprises: The control circuit applies an electrical signal corresponding to the data to be processed to the first layer of the neural network circuit.

Images