Online learning system based on memristor cross array

Abstract

The invention relates to an online learning system based on a memristor cross array, and belongs to the technical field of machine learning. The system comprises a forward calculation subsystem, a memristor resistance value online adjustment subsystem and a visualization subsystem, the forward calculation subsystem comprises a memristor cross array simulation device and a computer module, and derivative calculation in neural network training and memristor resistance value dynamic adjustment are combined through a simulation platform. A memory resistor is used for achieving storage and calculation integration, a computer module completes derivative calculation in a differential mode, a differential derivative calculation mechanism of the computer module directly obtains a derivative through actual response of a physical device, the defect that a storage and calculation integration framework cannot store intermediate variables is overcome, and calculation autonomy and efficiency are remarkably improved. Meanwhile, an introduced resistance updating mechanism combining coarse adjustment and fine adjustment can realize efficient and accurate control of the resistance of the memristor. Finally, the neural network learning efficiency and performance based on the memristor cross array are improved.

Description

Technical Field

[0001] The present invention is applicable to the field of machine learning technology, and in particular relates to an online learning system based on a memristor crossbar array. Background Art

[0002] Current deep learning neural network training suffers from low computational efficiency, constrained by the traditional "von Neumann" architecture. This problem is exacerbated during real-time processing in large-scale inference tasks.

[0003] Memristors are a new type of electronic component whose resistance can be controlled by electrical signals. Their resistance (and therefore the data stored) is determined by the charge or magnetic flux passing through them. By directly applying voltage or current to the memristor array, calculations can be performed in the same location where the data is stored, eliminating the need for frequent data transfer between the processor and memory. This achieves "storage-computing integration," creating a new opportunity to address the aforementioned inefficient neural network training. However, existing methods for updating the resistance of memristor crossbar arrays, which are derived from integrated memristors, still lack versatility, autonomy, and performance, and are unable to efficiently meet the online learning requirements of neural network computing.

[0004] That is, the current neural network training calculation based on memristor crossbar arrays has technical problems of low learning efficiency and learning accuracy. Summary of the Invention

[0005] In view of this, the present invention provides an online learning system based on a memristor crossbar array, which can realize dynamic update and optimization of the resistance value of the memristor, improve the learning efficiency and accuracy of the neural network calculation, and solve the technical problem of low learning efficiency and learning accuracy of the current neural network training calculation based on the memristor crossbar array.

[0006] The present invention provides an online learning system based on a memristor crossbar array, comprising a forward computing subsystem, a memristor resistance online adjustment subsystem, and a visualization subsystem; The forward computing subsystem includes a memristor crossbar array simulation device and a computer module. The memristor crossbar array simulation device is used to receive original input data and convert the original input data into an analog signal that matches the memristor crossbar array inside the memristor crossbar array simulation device, and input the analog signal into the memristor crossbar array to complete the forward computing. The computer module is used to complete the calculation of the objective function of the neural network to be trained and calculate the derivative of each memristor in the memristor crossbar array in a differential manner. The memristor resistance online adjustment subsystem is used to convert the derivative into an analog excitation signal, and perform coarse and fine adjustment on the resistance of the memristor crossbar array according to the analog excitation signal; The visualization subsystem is used to track the calculation process of the derivative and the resistance adjustment process of the memristor crossbar array, and to display the learning process of the neural network to be trained.

[0007] Optionally, the calculating the derivative of each memristor in the memristor crossbar array in a differential manner includes: The current forward calculation result of the memristor cross array is recorded as the original output, and then a disturbance is applied to each memristor resistance value in the memristor cross array and the forward calculation is performed again to obtain the disturbed output. Based on the original output and the disturbed output, the derivative of each memristor in the memristor cross array is calculated.

[0008] Optionally, the derivative of the memristor in the memristor crossbar array is: in, represents the derivative of the memristor at the i-th row and j-th column in the memristor crossbar array, represents the resistance disturbance increment of the memristor at row i and column j, represents the original resistance value of the memristor at the i-th row and j-th column, represents the original input data input to the memristor crossbar array, Indicates the original resistance value and the original input data The objective function value of the memristor crossbar array is Indicates that it corresponds to and the original input data The objective function value of the memristor crossbar array.

[0009] Optionally, the memristor resistance online adjustment subsystem includes a signal conversion interface, a resistance automatic update simulation circuit module, and a resistance fine-tuning simulation circuit module; The signal conversion interface is used to receive the derivative and convert it into the analog excitation signal, the resistance automatic update simulation circuit module is used to coarsely adjust the resistance of the memristor cross array according to the analog excitation signal, and the resistance fine-tuning simulation circuit module is used to fine-tune the resistance of the memristor cross array according to the analog excitation signal.

[0010] Optionally, the coarsely adjusting the resistance of the memristor crossbar array according to the analog excitation signal includes: The resistance automatic updating simulation circuit module controls the peripheral control circuit module in the memristive crossbar array simulation device to generate a first amplitude pulse corresponding to the integer part of the derivative according to the sign of the derivative and the amplitude of the analog excitation signal, and then applies the first amplitude pulse to the memristive crossbar array to obtain a preliminary updated resistance value; The fine-tuning of the resistance value of the memristor crossbar array according to the analog excitation signal includes: The resistance fine-tuning simulation circuit module controls the peripheral control circuit module in the memristor cross array simulation device to generate a second amplitude pulse corresponding to the fractional part of the derivative based on the amplitude of the analog excitation signal, and then uses an iterative approximation algorithm to correct the error of the initially updated resistance value by applying the second amplitude pulse to the memristor cross array.

[0011] Optionally, the memristor crossbar array simulation device further includes an input interface, an output interface, and an array monitoring module; The input interface is used to convert the original input data into the analog signal, the output interface is used to convert the forward calculation result into a digital signal and transmit it to the computer module, and the array monitoring module is used to periodically read the resistance value of the memristor crossbar array and synchronize it to the visualization subsystem through a message mechanism with a timestamp.

[0012] Optionally, the input interface uses a dynamic range compression technology to convert the original input data into the analog signal.

[0013] Optionally, the computer module includes a communication interface, and the communication interface is used to receive the digital signal and send the derivative to the memristor resistance online adjustment subsystem; The computer module is used to provide a numerical simulation environment and complete the calculation of the objective function in the numerical simulation environment.

[0014] Optionally, the visualization subsystem includes a timing analysis and synchronization module, a source location analysis module, and a data visualization module; The timing analysis and synchronization module is used to time-align the information received by the visualization subsystem, the source location analysis module is used to analyze the generation location of the information received by the visualization subsystem, and the data visualization module is used to visualize the information received by the visualization subsystem.

[0015] Optionally, the visualization subsystem is further configured to trigger an adaptive learning rate adjustment algorithm when the resistance value of the memristor crossbar array drifts or converges stagnantly.

[0016] Compared with the prior art, the present invention has the following beneficial effects: The system of the present invention includes a forward computing subsystem, a memristor resistance online adjustment subsystem, and a visualization subsystem, wherein the forward computing subsystem includes a memristor crossbar array simulation device and a computer module. Through the simulation platform, the derivative calculation in the neural network calculation is combined with the dynamic adjustment of the memristor resistance, and the memristor is used to realize "storage and calculation integration". At the same time, an online update method is proposed in which the computer module completes the derivative calculation in a differential manner and the memristor resistance online adjustment subsystem updates the memristor resistance. The differential derivative calculation mechanism directly obtains the derivative through the actual response of the physical device, avoiding the calculation error of the derivative caused by the possible model mismatch problem when data is migrated between different types of models, and makes up for the defect of the "storage and calculation integration" architecture that cannot store intermediate variables, significantly improving the calculation autonomy and efficiency. At the same time, the system introduces a resistance update mechanism that combines coarse adjustment and fine adjustment. Through the cooperation of the forward computing subsystem and the memristor resistance online adjustment subsystem of the simulation platform, the integer and fractional parts of the derivative can be written step by step, achieving efficient and precise control of the memristor crossbar array. Ultimately, with the help of the system of the present invention, the learning efficiency and performance of the neural network based on the memristor crossbar array can be significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 Schematic diagram of an online learning system based on a memristor crossbar array provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0020] It should be understood that the present invention can be embodied in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to make the disclosure thorough and complete and to fully convey the scope of the invention to those skilled in the art. In the drawings, the dimensions and relative dimensions of layers and regions may be exaggerated for clarity. Like reference numerals denote like elements throughout.

[0021] It should be understood that when an element or layer is referred to as being "on," "adjacent to," "connected to," or "coupled to" another element or layer, it may be directly on, adjacent to, connected to, or coupled to the other element or layer, or there may be intervening elements or layers. Conversely, when an element is referred to as being "directly on," "directly adjacent to," "directly connected to," or "directly coupled to" another element or layer, there may be no intervening elements or layers. It should be understood that although the terms first, second, third, etc. may be used to describe various elements, components, regions, layers, and / or parts, these elements, components, regions, layers, and / or parts should not be limited by these terms. These terms are merely used to distinguish one element, component, region, layer, or part from another element, component, region, layer, or part. Thus, a first element, component, region, layer, or part discussed below may be represented as a second element, component, region, layer, or part without departing from the teachings of the present invention.

[0022] Spatially relative terms such as "under," "beneath," "below," "under," "above," "above," etc., may be used herein for convenience of description to describe the relationship of one element or feature shown in the figures to other elements or features. It should be understood that the spatially relative terms are intended to include different orientations of the device in use and operation in addition to the orientations shown in the figures. For example, if the device in the drawings is flipped, then the elements or features described as "under" or "beneath" or "beneath" the other elements will be oriented as "over" the other elements or features. Thus, the exemplary terms "under" and "under" may include both the upper and lower orientations. The device may be oriented otherwise (rotated 90 degrees or in other orientations) and the spatial descriptors used herein are interpreted accordingly.

[0023] The purpose of the terms used herein is only to describe specific embodiments and is not intended to limit the present invention. When used herein, the singular forms "a", "an", and "the" are intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that the terms "comprising" and / or "comprising", when used in this specification, determine the presence of the features, integers, steps, operations, elements and / or parts, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, parts and / or groups. When used herein, the term "and / or" includes any and all combinations of the relevant listed items.

[0024] In order to fully understand the present invention, detailed structures and steps will be provided in the following description to illustrate the technical solutions proposed by the present invention. Preferred embodiments of the present invention are described in detail below. However, in addition to these detailed descriptions, the present invention may also have other implementations.

[0025] like Figure 1, which is a schematic diagram of an online learning system based on a memristor crossbar array provided by an embodiment of the present invention, including a forward computing subsystem, a memristor resistance online adjustment subsystem, and a visualization subsystem.

[0026] The forward computing subsystem includes a memristor crossbar array simulation device and a computer module. The memristor crossbar array simulation device is used to receive original input data and convert the original input data into an analog signal that matches the memristor crossbar array inside the memristor crossbar array simulation device, and input the analog signal into the memristor crossbar array to complete the forward computing. The computer module is used to complete the calculation of the objective function of the neural network to be trained and calculate the derivatives of each memristor in the memristor crossbar array in a differential manner.

[0027] The memristor resistance online adjustment subsystem is used to convert the derivative into an analog excitation signal and perform coarse and fine adjustment on the resistance of the memristor crossbar array according to the analog excitation signal.

[0028] The visualization subsystem is used to track the calculation process of the derivative and the resistance adjustment process of the memristor crossbar array, and to display the learning progress of the neural network to be trained.

[0029] Among them, Figure 1 As shown, in this embodiment, a memristive cross array simulation device is configured to receive raw input data through an input interface. The type of raw input data is related to the type of the neural network to be trained, so the possible types are various. For example, it may be pixel data in the MNIST classification task, and the memristive cross array inside the memristive cross array simulation device can only recognize analog signals, specifically voltage or current signals. Therefore, in this embodiment, a memristive cross array simulation device is configured to include an input interface, and the received raw input data is converted into an analog voltage / current signal through the input interface to match the physical input port of the memristive cross array, wherein the signal conversion adopts dynamic range compression technology, and the analog signal is ensured to be within the linear response range of the memristive device through adaptive scaling, thereby avoiding calculation errors caused by nonlinear distortion.

[0030] A memristor crossbar array simulation device also includes an output interface. This interface converts the analog current signal output from the forward calculation in the memristor crossbar array into a digital signal after analog-to-digital conversion. This signal is then transmitted to the numerical simulation environment (such as MATLAB / Python) provided by the computer module for derivative calculation. After the analog signal is input into the memristor crossbar array, the forward calculation is completed through matrix multiplication (the product of the memristor resistance weight and the input signal).

[0031] In addition, a memristive crossbar array simulation device is provided, including an array monitoring module, which is used to periodically read the resistance value of the memristive crossbar array and synchronize it to the visualization subsystem through a timestamp message mechanism, so as to enable the system to display and judge the resistance value update status and whether the training of the neural network to be trained has been completed.

[0032] Corresponding to the output interface of the memristor crossbar array simulation device, such as Figure 1 As shown, this embodiment provides a computer module including a communication interface for receiving the forward computation output result, which has been converted into a digital signal, from the output interface, and transmitting the digital signal to the numerical simulation environment provided by the computer module to complete the derivative calculation; and transmitting the calculated derivative to the memristor resistance online adjustment subsystem. Simultaneously, the numerical simulation environment provided by the computer module calculates the objective function (e.g., the cross-entropy loss function) of the neural network to be trained by comparing the digital output with the label.

[0033] The memristor resistance online adjustment subsystem obtains an analog excitation signal for completing the resistance adjustment of the memristor crossbar array based on the received derivative through digital-to-analog conversion. The resistance adjustment includes two stages: coarse adjustment and fine adjustment. Coarse adjustment can set the size of the excitation signal corresponding to the integer part of the derivative, while fine adjustment can set the size of the excitation signal corresponding to the decimal part of the derivative, taking into account both efficiency and accuracy.

[0034] This embodiment of the system combines the derivative calculation in the neural network calculation with the dynamic adjustment of the resistance value of the memristor through the simulation platform, realizing "storage and calculation integration" with the memristor, while the computer module completes the derivative calculation in a differential manner. The differential derivative calculation mechanism of the computer module directly obtains the derivative through the actual response of the physical device, avoiding the calculation error of the derivative caused by the possible model mismatch problem when the data is migrated between different types of models, and making up for the shortcoming of the "storage and calculation integration" architecture that cannot store intermediate variables, significantly improving the calculation autonomy and efficiency. In addition, by introducing a coarse adjustment and fine adjustment mechanism, the present invention can realize the step-by-step writing of the integer and decimal parts of the derivative through the forward calculation subsystem of the simulation platform and the online adjustment subsystem of the memristor resistance, realize the precise control of the memristor crossbar array, and ensure the accuracy of the memristor resistance adjustment. Ultimately, the system of the present invention can significantly improve the learning efficiency and performance of the neural network based on the memristor crossbar array, and can be widely applied to the online learning and autonomous optimization of neural networks.

[0035] In one embodiment, the step of calculating the derivative of each memristor in the memristor crossbar array in a differential manner includes: The current forward calculation result of the memristor crossbar array is recorded as the original output. Then, a perturbation is applied to each memristor in the memristor crossbar array and the forward calculation is performed again to obtain the perturbed output. Based on the original output and the perturbated output, the derivative of each memristor in the memristor crossbar array is calculated: in, represents the derivative of the memristor at the i-th row and j-th column in the memristor crossbar array, It represents the resistance disturbance increment of the memristor at the i-th row and j-th column, that is, the small change. represents the original resistance value of the memristor at the i-th row and j-th column, Represents the original input data input to the memristor crossbar array, such as image pixel signals, feature vectors, etc. Indicates the original resistance value and the original input data Under this condition, the objective function value of the memristor crossbar array, such as the loss function value or output response, Indicates that the resistance value changes to After that, and the original input data The objective function value of the memristor crossbar array is shown in Figure 2.

[0036] The calculation method of the sensitivity (derivative) of each memristor in this embodiment does not require the chain derivation of traditional back-propagation, which can significantly improve the calculation efficiency.

[0037] In one embodiment, if Figure 1 As shown, the memristor resistance online adjustment subsystem includes a signal conversion interface, a resistance automatic update simulation circuit module, and a resistance fine-tuning simulation circuit module.

[0038] Among them, the signal conversion interface is used to receive the derivative and convert it into a voltage or current signal as an analog excitation signal, the resistance automatic update simulation circuit module is used to coarsely adjust the resistance of the memristor cross array according to the analog excitation signal, and the resistance fine-tuning simulation circuit module is used to fine-tune the resistance of the memristor cross array according to the analog excitation signal.

[0039] In this embodiment, preferably, during the coarse adjustment phase, the automatic resistance update simulation circuit module controls the peripheral control circuit modules in the memristive crossbar array simulation device to write the resistance value based on the integer portion of the derivative. During the fine adjustment phase, the resistance fine adjustment simulation circuit module controls the peripheral control circuit modules in the memristive crossbar array simulation device to correct the resistance error using a segmented update method based on the fractional portion of the derivative. It will be readily understood that in other embodiments, resistance writing and correction may be performed without depending on the specific magnitude of the derivative value, such as by directly setting a fixed integer value for resistance writing and a fixed correction value for resistance correction.

[0040] Specifically, the resistance value of the memristor crossbar array is roughly adjusted according to the analog excitation signal, including: The resistance automatic update simulation circuit module controls the peripheral control circuit module in the memristor crossbar array simulation device to generate a first amplitude pulse corresponding to the integer part of the derivative according to the sign of the derivative and the amplitude of the analog excitation signal, and then applies the first amplitude pulse to the memristor crossbar array to obtain a preliminary updated resistance value; Fine-tune the resistance of the memristor crossbar array according to the analog excitation signal, including: The resistance fine-tuning simulation circuit module controls the peripheral control circuit module in the memristor crossbar array simulation device to generate a second amplitude pulse corresponding to the fractional part of the derivative based on the amplitude of the analog excitation signal. It then uses an iterative approximation algorithm to correct the error in the initially updated resistance value by applying the second amplitude pulse to the memristor crossbar array.

[0041] It is easy to understand that the correction during the resistance fine-tuning stage typically includes multiple resistance error corrections, thereby gradually adjusting the resistance from the initial updated resistance value to the target resistance value, achieving high-precision adjustment. In this embodiment, the iterative approximation algorithm is preferably a binary method, where the first amplitude pulse (e.g., ±5V) is significantly larger than the second amplitude pulse (e.g., ±0.1V). A hash coding mapping is performed using a topological structure mapping table from the neural network to be trained to the memristor crossbar array, mapping the logical derivative to a physical resistance range (e.g., 1kΩ to 100kΩ), thus resolving the mismatch between the limited resistance range and the theoretical gradient.

[0042] This combined coarse and fine adjustment mechanism balances efficiency and accuracy through a two-stage adjustment process. Coarse adjustment allows the memristor resistance to quickly approach the target value, while fine adjustment eliminates device nonlinearity and noise. This combined coarse and fine adjustment resistance update method reduces the resistance adjustment error from >10% in a traditional single-stage approach to <2%.

[0043] In one embodiment, if Figure 1 As shown, the visualization subsystem includes a timing analysis and synchronization module, a source location analysis module, and a data visualization module.

[0044] The timing analysis and synchronization module is used to align the timing of the information received by the visualization subsystem, the source location analysis module is used to analyze the generation location of the information received by the visualization subsystem, and the data visualization module is used to visualize the information received by the visualization subsystem.

[0045] Specifically, such as Figure 1As shown in the figure, the visualization subsystem signal connection input interface, output interface, array monitoring module and signal conversion interface, through asynchronous signal fusion technology, integrates distributed signal sources (such as the resistance of different memristors, interface circuit voltage), and determines the source of signals with different times generated by signal sources at different positions and synchronizes the timing. In addition to receiving the resistance of the memristor cross array sent by the array monitoring module, it also displays key indicators such as the target function curve, derivative distribution, and resistance change trend in real time. Global monitoring is achieved through timing alignment and spatial mapping, providing data support for algorithm optimization.

[0046] Furthermore, the visualization subsystem is used to trigger an adaptive learning rate adjustment algorithm when the resistance value of the memristor crossbar array drifts or converges stagnantly, so as to adjust the learning rate of the neural network to be trained and improve the training efficiency and performance of the model.

[0047] Based on the online learning system based on the memristor crossbar array provided in the above embodiment, taking the MNIST classification task as an example, the complete training process of the neural network to be trained is as follows: 1) Initialization: The resistance values ​​of the memristor crossbar array are randomly distributed, and the topological structure mapping table from the neural network to the memristor crossbar array is loaded into the peripheral control circuit module in the memristor crossbar array simulation device.

[0048] 2) Forward calculation: The pixel data of the image “5” is input to the input interface of the memristor crossbar array simulation device. The input interface circuit converts it into a 28×28 analog voltage matrix and applies it to the memristor crossbar array to complete the forward calculation.

[0049] 3) Derivative generation: After completing the objective function and loss value calculations in the numerical simulation environment provided by the computer module in the memristor crossbar array simulation device, the sensitivity (derivative) of each memristor is calculated one by one through differential means, and a sensitivity matrix is ​​generated.

[0050] 4) Resistance adjustment: Coarse adjustment stage: The peripheral control circuit module applies a 10ms positive pulse to the high-sensitivity memristor (such as the edge detection related unit).

[0051] Fine-tuning stage: Based on the residual feedback, three iterative fine-tuning pulses of 1ms each are applied to the key memristor.

[0052] 5) Convergence judgment: The visualization subsystem displays the loss value in real time. When the loss value drops to a set threshold (such as 0.1), the training of the neural network to be trained is considered complete and the training is terminated.

[0053] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. 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. 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 various embodiments of the present invention, and should all be included in the scope of protection of the present invention.

Claims

1. An online learning system based on a memristor crossbar array, characterized in that: Including forward calculation subsystem, memristor resistance online adjustment subsystem and visualization subsystem; The forward computing subsystem includes a memristor crossbar array simulation device and a computer module. The memristor crossbar array simulation device is used to receive original input data and convert the original input data into an analog signal that matches the memristor crossbar array inside the memristor crossbar array simulation device, and input the analog signal into the memristor crossbar array to complete the forward computing. The computer module is used to complete the calculation of the objective function of the neural network to be trained and calculate the derivative of each memristor in the memristor crossbar array in a differential manner. The memristor resistance online adjustment subsystem is used to convert the derivative into an analog excitation signal, and perform coarse and fine adjustment on the resistance of the memristor crossbar array according to the analog excitation signal; The visualization subsystem is used to track the calculation process of the derivative and the resistance adjustment process of the memristor crossbar array, and to display the learning process of the neural network to be trained.

2. The online learning system based on a memristor crossbar array according to claim 1, characterized in that: The step of calculating the derivative of each memristor in the memristor crossbar array in a differential manner includes: The current forward calculation result of the memristor cross array is recorded as the original output, and then a disturbance is applied to each memristor resistance value in the memristor cross array and the forward calculation is performed again to obtain the disturbed output. Based on the original output and the disturbed output, the derivative of each memristor in the memristor cross array is calculated.

3. The online learning system based on a memristor crossbar array according to claim 2, characterized in that: The derivative of the memristor in the memristor crossbar array is: in, represents the derivative of the memristor at the i-th row and j-th column in the memristor crossbar array, represents the resistance disturbance increment of the memristor at row i and column j, represents the original resistance value of the memristor at the i-th row and j-th column, represents the original input data input to the memristor crossbar array, Indicates the original resistance value and the original input data The objective function value of the memristor crossbar array is Indicates that it corresponds to and the original input data The objective function value of the memristor crossbar array.

4. The online learning system based on a memristor crossbar array according to claim 1, characterized in that: The memristor resistance online adjustment subsystem includes a signal conversion interface, a resistance automatic update simulation circuit module, and a resistance fine-tuning simulation circuit module; The signal conversion interface is used to receive the derivative and convert it into the analog excitation signal, the resistance automatic update simulation circuit module is used to coarsely adjust the resistance of the memristor cross array according to the analog excitation signal, and the resistance fine-tuning simulation circuit module is used to fine-tune the resistance of the memristor cross array according to the analog excitation signal.

5. The online learning system based on a memristor crossbar array according to claim 4, characterized in that: The coarsely adjusting the resistance value of the memristor crossbar array according to the analog excitation signal includes: The resistance automatic updating simulation circuit module controls the peripheral control circuit module in the memristive crossbar array simulation device to generate a first amplitude pulse corresponding to the integer part of the derivative according to the sign of the derivative and the amplitude of the analog excitation signal, and then applies the first amplitude pulse to the memristive crossbar array to obtain a preliminary updated resistance value; The fine-tuning of the resistance value of the memristor crossbar array according to the analog excitation signal includes: The resistance fine-tuning simulation circuit module controls the peripheral control circuit module in the memristor cross array simulation device to generate a second amplitude pulse corresponding to the fractional part of the derivative based on the amplitude of the analog excitation signal, and then uses an iterative approximation algorithm to correct the error of the initially updated resistance value by applying the second amplitude pulse to the memristor cross array.

6. The online learning system based on a memristor crossbar array according to claim 1, characterized in that: The memristor crossbar array simulation device further includes an input interface, an output interface, and an array monitoring module; The input interface is used to convert the original input data into the analog signal, the output interface is used to convert the forward calculation result into a digital signal and transmit it to the computer module, and the array monitoring module is used to periodically read the resistance value of the memristor crossbar array and synchronize it to the visualization subsystem through a message mechanism with a timestamp.

7. The online learning system based on a memristor crossbar array according to claim 6, characterized in that: The input interface converts the original input data into the analog signal using a dynamic range compression technique.

8. The online learning system based on a memristor crossbar array according to claim 6 or 7, characterized in that: The computer module includes a communication interface, the communication interface is used to receive the digital signal and send the derivative to the memristor resistance online adjustment subsystem; The computer module is used to provide a numerical simulation environment and complete the calculation of the objective function in the numerical simulation environment.

9. The online learning system based on a memristor crossbar array according to claim 1, characterized in that: The visualization subsystem includes a timing analysis and synchronization module, a source location analysis module, and a data visualization module; The timing analysis and synchronization module is used to time-align the information received by the visualization subsystem, the source location analysis module is used to analyze the generation location of the information received by the visualization subsystem, and the data visualization module is used to visualize the information received by the visualization subsystem.

10. The online learning system based on a memristor crossbar array according to claim 1 or 9, characterized in that: The visualization subsystem is further configured to trigger an adaptive learning rate adjustment algorithm when the resistance value of the memristor crossbar array drifts or converges stagnates.

Images