Image classification method of Bayesian machine hardware based on diffusion type memristor

Abstract

The invention discloses an image classification method of Bayesian machine hardware based on a diffusion type memristor. The method comprises the following steps: firstly, preparing a diffusion type memristor and constructing a cross array circuit structure, and taking the diffusion type memristor as a probability bit of Bayesian machine hardware in a voltage modulation stage; obtaining a reference image with a preset category label, and carrying out statistics on a prior probability and a conditional probability of a pixel point; and on the basis of the prior probability and the conditional probability, voltage-probability modulation is carried out on to-be-classified images, a modulation voltage is obtained in combination with a nonlinear transformation method, the modulation voltage is input into the cross array circuit structure, and the images are classified according to the number of times of conduction. According to the invention, the diffusion type memristor is introduced as a probability bit, the structure of the cross array circuit is optimized through the true physical random characteristic of the device, and the use of a large number of gate circuits and pseudo-random number generators is avoided, so that the calculation cost is reduced, and the calculation capability of realizing complex tasks is realized.

Description

Technical Field

[0001] The present invention relates to a Bayesian machine hardware, belonging to the field of probability calculation, and particularly to an image classification method for a Bayesian machine hardware based on diffusive memristors. Background Art

[0002] With the rapid development of artificial intelligence, traditional von Neumann computers have encountered problems such as memory walls, high time costs, and high energy costs. Probabilistic computing not only has the potential to solve complex problems more effectively than traditional computing paradigms, but also exhibits stronger robustness and anti-interference capabilities. A Bayesian machine is a probabilistic framework that allows decisions to be made in the presence of incomplete information, maximizing the combination of all available hypotheses and prior knowledge.

[0003] Although Bayesian machines exhibit advantages such as robustness to device defects and single-event perturbations, their hardware implementations suffer from problems such as repeated memory access, low energy efficiency, and high computational costs. Traditional hardware implementations of Bayesian machines have not been improved for the problem of separating data storage and computing, making it difficult to improve the computing power of hardware Bayesian machines. Therefore, when deploying Bayesian machines on FPGAs and CMOS-based application-specific integrated circuits, the computing power is often greatly reduced due to the neglect of these non-ideal factors.

[0004] To address these limitations, some researchers have proposed Bayesian machines based on non-volatile memristor crossbar arrays for probabilistic computing. The working mode of the crossbar array structure is naturally compatible with Bayesian rules, and the significant advantages of the crossbar array structure in scalability and parallel computing can greatly reduce data movement during the computing process. However, when implementing non-volatile memristor crossbar arrays, a large number of pseudo-random generators and gate circuits must be used, which will bring additional time costs and increase energy consumption. Summary of the Invention

[0005] To solve the problems in the background art, the present invention provides an image classification method for a Bayesian machine hardware based on diffusive memristors. The present invention aims to utilize the physical randomness of diffusive memristors and the parallel computing ability of crossbar arrays to achieve higher computing efficiency and ability with a more streamlined circuit structure. Compared with traditional methods, the circuit structure is significantly simplified, and it has the computing ability to implement complex tasks.

[0006] The technical solution adopted by the present invention is as follows:

[0007] The image classification method for a Bayesian machine hardware based on diffusive memristors of the present invention includes:

[0008] Step 1) Prepare a number of diffusive memristors to construct a crossbar array circuit structure, and use each diffusive memristor in the crossbar array circuit structure as a probability bit of the Bayesian machine hardware to achieve voltage-probability modulation.

[0009] Step 2) Obtain a number of reference images with preset class labels and construct an image dataset. Statistically obtain the prior probability and conditional probability of each pixel point's binary brightness of each reference image in the image dataset for subsequent test set image classification.

[0010] Step 3) Based on the prior probability and conditional probability of each reference image under each preset class obtained in Step 2), perform voltage-probability modulation on the image to be classified and process it by combining a non-linear transformation method to obtain the modulation voltage to be classified. Input the modulation voltage to be classified into the crossbar array circuit structure, and classify the image according to the conduction times of the crossbar array circuit structure.

[0011] In the above-mentioned Step 1), when preparing the diffusive memristor, first use photolithography to pattern one side of the silicon substrate. When patterning, the positions of bit lines, etc. also need to be reserved. Then, at the patterned position, first use sputtering to deposit a titanium (Ti) metal layer and a platinum (Pt) metal layer in sequence. The thickness ranges of both the titanium (Ti) metal layer and the platinum (Pt) metal layer are 1 - 40 nm. Then, prepare an aluminum oxide (Al 2 O 3 dielectric layer with a thickness of 7 - 20 nm on the platinum (Pt) metal by atomic layer deposition (ALD). Finally, use thermal evaporation and lift-off processes to prepare a silver (Ag) metal layer and a gold (Au) metal layer in sequence on the aluminum oxide (Al 2 O 3 dielectric layer. The thickness ranges of both the silver (Ag) metal layer and the gold (Au) metal layer are 20 - 40 nm. Use the titanium (Ti) metal layer, the platinum (Pt) metal layer, the aluminum oxide (Al 2 O 3 dielectric layer and the silver (Ag) metal layer together as the bottom electrode, and use the gold (Au) metal layer as the top electrode. Finally, prepare an Au / Ag / Al 2 O 3 / Pt / Ti diffusive memristor.

[0012] In the above-mentioned Step 1), Au / Ag / Al 2 O 3The / Pt / Ti diffusive memristor is fabricated through an anodic oxidation process, and it exhibits fluctuating threshold switching characteristics within the switching voltage range; in the diffusive memristor, when the compliance current is 1 nA, an external voltage of 5 - 10 V is applied to move silver particles to form a conductive channel, and the diffusive memristor switches from the initial high-resistance state to the low-resistance state, with the current increasing by several orders of magnitude. After removing the external voltage, the diffusive memristor quickly returns to the initial high-resistance state within a time less than the preset time threshold to continue the voltage-probability modulation process; due to the residual nanoparticles on both sides of the electrodes of the diffusive memristor, when the reapplied voltage is within the probability switching range, the diffusive memristor can easily switch to the low-resistance state; when performing image classification, the modulated voltage obtained after voltage-probability modulation is within the probability switching range of the diffusive memristor, and the probability switching range is as low as 0.07 - 0.14 V.

[0013] In the described step 1), the crossbar array circuit structure further includes I bit lines, N word lines, and I probability lines. Each word line is arranged horizontally in parallel at intervals, each bit line and probability line are arranged vertically in parallel at intervals, each probability line is located between two adjacent bit lines, and each word line is perpendicularly and crosswise arranged with each bit line and probability line. A diffusive memristor and a transistor are disposed in the rectangular region formed between every two adjacent bit lines and every two adjacent word lines, thereby forming an array unit. For each array unit, the bottom electrode of the diffusive memristor is connected to the bit line shared with the previous array unit in the current row, the top electrode of the diffusive memristor is connected to the word line shared with the previous array unit in the current column, and the transistor is connected between the probability line and the bit line to which the diffusive memristor is connected; one end of each word line and bit line is connected to a pulse generator, and one end of each word line receives a pulse input voltage. The other ends of each word line are sequentially connected to the diffusive memristors in the last array unit of each row according to the arrangement order, the other ends of each bit line are used as output terminals, and both ends of each probability line are connected to a pulse receiving module.

[0014] In the described step 2), the number of preset categories is the same as the number of bit lines and probability lines in the crossbar array circuit structure, and each image has its own one preset category; the number of rows of pixel points of each image is the same as the number of word lines in the crossbar array circuit structure.

[0015] In the described step 2), each reference image has its own one preset category label, that is, each pixel point in each reference image has the same one preset category label. After binarizing the brightness of each pixel point in each reference image, the binarized brightness is obtained. For each reference image under each preset category, the prior probability and conditional probability of the binarized brightness of each pixel point in each reference image are statistically obtained.

[0016] In step 3), first, the binary brightness of each pixel in the image to be classified is used to obtain the classification probability of each pixel belonging to each preset category based on the prior probability and conditional probability of each reference image under each preset category obtained in step 2). The initial switching voltage-probability curve of the diffusive memristor is non-linearly transformed to obtain an updated switching voltage-probability curve as follows:

[0017]

[0018] where G() is the updated switching voltage-probability curve, that is, the classification probability of the preset category; a and p are the first and second range limit parameters in the non-linear transformation process; V d is the modulation voltage after normalizing the probability switching range of the diffusive memristor, and then it is input into the corresponding word line of the memristor cross array.

[0019] According to the updated switching voltage-probability curve, the modulation voltage of each pixel in the image to be classified is obtained. For each preset category, classification recognition is performed a preset number of times. For each classification recognition, first, the modulation voltages of each row of pixels in the image to be classified are sequentially input into their respective word lines, and at the same time, the column selection pulses generated by the pulse generator are input into the bit lines corresponding to the current preset category, so that a voltage difference is formed between the word line and the bit line. The pulse receiving module saves the number of times the probability line connected to the same diffusive memristor as the current bit line is conducted until all the classification recognitions of the preset number of times are completed, and the total number of times the current probability line is conducted is obtained. Finally, the number of times each probability line is conducted under each preset category is obtained, and the preset category to which the probability line with the most conduction times belongs is used as the category of the current image, and the final image classification is realized based on the full bit line probability comparison.

[0020] During the operation of the diffusive memristor, when one of the bit lines operates in the forward voltage mode, all other bit lines are applied with reverse voltages, so that the diffusive memristor always maintains the probability switching operation mode.

[0021] When the probability line is conducted, all the diffusive memristors in the current column connected to the current probability line are turned on. During the operation of the diffusive memristor, the final on state is as follows:

[0022]

[0023] where p m is the final on state of the m-th diffusive memristor, p m =0 represents not turned on, p m =1 represents turned on; u() is the binary step function; V md is the modulation voltage input to the m-th diffusive memristor; Vo is the boundary voltage of the probability switch range; V S is the preset scaling voltage; rand is the preset random term. To satisfy the probability switch characteristic of the probability bit, the rand term is introduced to define the random state of the probability bit at each turn-on.

[0024] The beneficial effects of the present invention are as follows:

[0025] 1. The present invention introduces a diffusive memristor as the probability bit. Through the true physical random characteristic of the device, the crossbar array circuit structure is optimized. Compared with implementing a Bayesian machine using a non-volatile memristor crossbar array circuit, the use of a large number of gate circuits and pseudo-random number generators is avoided, thereby reducing the computing cost. It has wide applicability, has the computing ability to implement complex tasks, and can complete tasks such as multi-class image recognition.

[0026] 2. The present invention uses a non-linear transformation of the voltage-probability curve in probability calculation to avoid numerical underflow in the calculation process, significantly improving the accuracy and efficiency of the Bayesian machine in the inference process; the inference process based on Bayes' formula is realized through the crossbar array circuit structure, avoiding the use of circuit devices such as multipliers. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 is a schematic diagram of voltage-probability modulation of an input image in an embodiment of the present invention;

[0028] Figure 2 is an optical diagram of a diffusive memristor in an embodiment of the present invention;

[0029] Figure 3 is a voltage-probability curve graph of a diffusive memristor before and after non-linear transformation in an embodiment of the present invention;

[0030] Figure 4 is a schematic diagram of the crossbar array circuit structure of a diffusive memristor in an embodiment of the present invention;

[0031] Figure 5 is a comparison graph of recognition accuracy before and after non-linear transformation of the voltage-probability curve in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0032] In order to enable those skilled in the art to better understand the solution of the present invention and make the above-mentioned objects, features, and advantages of the present invention more clearly understandable, the present invention will be further described in detail below with reference to the embodiments.

[0033] The image classification method of the Bayesian machine hardware based on a diffusive memristor of the present invention is characterized by including:

[0034] Step 1) Prepare a number of diffusive memristors to construct a crossbar array circuit structure, and use each diffusive memristor in the crossbar array circuit structure as a probability bit of the Bayesian machine hardware to achieve voltage-probability modulation. When preparing the diffusive memristor, first use photolithography to pattern one side of the silicon substrate. When patterning, the positions of bit lines, etc. also need to be reserved. Then, at the patterned positions, first use sputtering to deposit a titanium (Ti) metal layer and a platinum (Pt) metal layer in sequence. The thickness ranges of both the titanium (Ti) metal layer and the platinum (Pt) metal layer are 1 - 40 nm. Then, use atomic layer deposition (ALD) to prepare an aluminum oxide (Al) dielectric layer with a thickness of 7 - 20 nm on the platinum (Pt) metal. 2 O 3 Finally, use thermal evaporation and lift-off processes to prepare a silver (Ag) metal layer and a gold (Au) metal layer in sequence on the aluminum oxide (Al) dielectric layer. The thickness ranges of both the silver (Ag) metal layer and the gold (Au) metal layer are 20 - 40 nm. Use the titanium (Ti) metal layer, the platinum (Pt) metal layer, the aluminum oxide (Al) dielectric layer, and the silver (Ag) metal layer together as the bottom electrode, and use the gold (Au) metal layer as the top electrode. Finally, prepare an Au / Ag / Al 2 O 3 / Pt / Ti diffusive memristor. 2 O 3 The Au / Ag / Al 2 O 3 / Pt / Ti diffusive memristor is fabricated by an anodic oxidation process, and it exhibits fluctuating threshold switching characteristics within the switching voltage range. In the diffusive memristor, when the compliance current is 1 nA, apply an external voltage of 5 - 10 V to move silver particles to form a conduction channel. The diffusive memristor switches from the initial high-resistance state to the low-resistance state, and the current increases by several orders of magnitude. After removing the external voltage, the diffusive memristor quickly returns to the initial high-resistance state within a time less than the preset time threshold to continue the voltage-probability modulation process. Since there are residual nanoparticles on both sides of the electrodes of the diffusive memristor, when the voltage applied again is within the probability switching range, the diffusive memristor can easily switch to the low-resistance state. When performing image classification, the modulated voltage obtained after voltage-probability modulation is within the probability switching range of the diffusive memristor, and the probability switching range is as low as 0.07 - 0.14 V.

[0035] Au / Ag / Al 2 O 3 / Pt / Ti diffusive memristor

[0036] In probability calculation, the probability bits that change continuously over time are always "0" or "1". A diffusive memristor is used as the probability bit to implement voltage-probability modulation to generate an input signal, and the modulated signal is input into the memristor crossbar array structure to achieve the probability calculation process. When the applied modulation pulse voltage is within the probability switching range of the diffusive memristor, the memristive device will randomly turn on. The switching voltage-probability curve is obtained based on the statistical data of the device turning on.

[0037] The probability bit structure composed of diffusive memristors exhibits characteristics similar to those of a random Hopfield network, and its behavior is similar to that of a binary random neuron in a neuromorphic computing system, which corresponds to the memristor being on or off.

[0038] The crossbar array circuit structure also includes I bit lines, N word lines, and I probability lines. Each word line is arranged horizontally in parallel at intervals, each bit line and probability line are arranged vertically in parallel at intervals. Each probability line is located between two adjacent bit lines. Each word line is perpendicularly and cross-arranged with each bit line and probability line respectively. A diffusive memristor and a transistor are provided in the rectangular area formed between every two adjacent bit lines and every two adjacent word lines, thereby forming an array unit. For each array unit, the bottom electrode of the diffusive memristor is connected to the bit line shared with the previous array unit in the current row, the top electrode of the diffusive memristor is connected to the word line shared with the previous array unit in the current column, and the transistor is connected between the probability line and the bit line to which the diffusive memristor is connected; one end of each word line and bit line is connected to the pulse generator, and one end of each word line receives the pulse input voltage. The other ends of each word line are sequentially connected to the diffusive memristors in the last array unit of each row in order. The other ends of each bit line are used as output terminals, and both ends of each probability line are connected to the pulse receiving module. It is arranged according to the single-transistor-single-memristor 1T1R (one-transistor-one-memristor) unit architecture of rows and columns.

[0039] Step 2) Obtain a number of reference images with preset category labels and construct an image dataset. Statistically obtain the prior probability and conditional probability of each pixel point of each reference image in the image dataset for subsequent classification of test set images. The number of preset categories is the same as the number of bit lines and probability lines in the crossbar circuit structure, and each image has its own one preset category; the number of rows of pixel points in each image is the same as the number of word lines in the crossbar circuit structure. Each reference image has its own one preset category label, that is, each pixel point in each reference image has the same one preset category label. After binarizing the brightness of each pixel point in each reference image, obtain the binarized brightness. For each reference image under each preset category, statistically obtain the prior probability and conditional probability of the binarized brightness of each pixel point of each reference image.

[0040] Step 3) Based on the prior probability and conditional probability of each reference image under each preset category obtained in Step 2), perform voltage-probability modulation on the image to be classified and process it by combining a non-linear transformation method to obtain the modulated voltage to be classified. Input the modulated voltage to be classified into the crossbar circuit structure, and classify the image according to the number of conduction times of the crossbar circuit structure. First, based on the prior probability and conditional probability of each reference image under each preset category obtained in Step 2), obtain the classification probability of each pixel point belonging to each preset category for each pixel point in the image to be classified; perform a non-linear change on the initial switching voltage-probability curve of the diffusive memristor to change it into an updated switching voltage-probability curve, as follows:

[0041]

[0042] where G() is the updated switching voltage-probability curve, that is, the classification probability of the preset category; a and p are the first and second range limit parameters in the non-linear transformation process; V d is the modulated voltage after normalizing the probability switching range of the diffusive memristor, and then input it into the corresponding word line of the memristor crossbar.

[0043] The modulation voltage of each pixel point in the image to be classified is obtained according to the updated switching voltage - probability curve; for each preset category, classification recognition is performed a preset number of times. For each classification recognition, first, the modulation voltages of the pixel points in each row of the image to be classified are sequentially input into respective word lines, and at the same time, the column selection pulses sent by the pulse generator are input into the bit lines corresponding to the current preset category, so that a voltage difference is formed between the word lines and the bit lines. The pulse receiving module saves the number of times the probability line connected to the same diffusive memristor as the current bit line is turned on until all the classification recognitions of the preset number of times are completed, and the total number of times the current probability line is turned on is obtained; finally, the number of times each probability line is turned on under each preset category is obtained, and the preset category to which the probability line with the most turn - on times belongs is used as the category of the current image, and the final image classification is realized based on the comparison of the probabilities of all bit lines.

[0044] Non - linear transformation is performed on the switching voltage - probability curve of the diffusive memristor to obtain a new switching voltage - probability curve, which can avoid numerical underflow in the calculation process, improve the calculation accuracy while improving the calculation efficiency.

[0045] During the operation of the diffusive memristor, when one of the bit lines operates in the forward voltage mode, all other bit lines are applied with reverse voltages, so that the diffusive memristor always maintains the probability switching operation mode.

[0046] When the probability line is turned on, all the diffusive memristors in the current column connected to the current probability line are turned on. During the operation of the diffusive memristor, the final turn - on states are as follows:

[0047]

[0048] where, p m is the final turn - on state of the m - th diffusive memristor, p m = 0 represents not turned on, p m = 1 represents turned on; u() is the binary step function; V md is the modulation voltage input to the m - th diffusive memristor; V o is the boundary voltage of the probability switching range; V S is the preset scaling voltage; rand is the preset random term. In order to meet the probability switching characteristics of the probability bit, the rand term is introduced to define the random state of the probability bit at each turn - on.

[0049] The probability calculation framework of the Bayesian machine fully conforms to the characteristics of the cross - array based on diffusive memristors. According to the probability of an event occurring at a previous moment and the likelihood coefficient, the posterior probability of the event occurring when the observed value appears can be estimated. In practical applications, each input feature value is usually considered to be conditionally independent, so the Bayesian rule can be simplified as:

[0050] p(X = x|O 1 ,O 2 ,O 3 ,…,O n ) = p(X = x) × p(O 1 |X = x) × p(O 2 |X = x)…× p(O n |X = x)

[0051] Among them, the probability that the event occurs at the previous moment is the prior probability p(X = x), and the likelihood coefficient is p(O 1 ,O 2 ,O 3 ,…,O n ). O 1 ,O 2 ,O 3 ,…,O n is the observed value of the event occurrence, and p(X = x|O 1 ,O 2 ,O 3 ,…,O n ) is the posterior probability that the event X = x occurs when the above observed values occur.

[0052] In the process of classification and recognition, the classification probability that the current image output at the other end of each bit line belongs to one of the preset categories is as follows:

[0053]

[0054] Among them, Q() is the posterior probability, that is, when the pixel values of the input image are P 1 、P 2 、…、P N respectively, the probability that the final classification result is q; x = q represents that the final classification result is q, and in the image classification process, it represents the specific category result to which the image is classified; P 1 、P 2 、…、P n 、…、P N are the pixel values of the current input image respectively, and after modulation, they respectively correspond to the input voltage differences between the 1st, 2nd, …, nth, …, Nth word lines and bit lines; are the switching states of the 1st, 2nd, …, nth, …, Nth diffusive memristors under the pulse in the zth word line respectively.

[0055] According to the probability multiplication characteristic of the diffusive memristor crossbar array, when the pulse generator sends n positive voltage pulses to each word line, the pulse receiving module at the end of the probability line can save the number of times each probability line is turned on. This process is equivalent to obtaining the final posterior probability through Bayesian formula by Bayesian machine hardware calculation. Finally, compare the final conduction times of each probability line, and the category with the highest probability is obtained as the classification result.

[0056] As Figure 2 shown, the Au / Ag / Al2O3 / Pt / Ti diffusive memristor prepared by the present invention. First, a silicon substrate with two-sided electrodes is obtained, and lithography is used to pattern the silicon substrate. Then, sputtering is used to deposit a 3-nm titanium (Ti) metal layer and a 40-nm platinum (Pt) metal layer in sequence. Then, an alumina (Al) 2 O 3 dielectric layer with a thickness of 10 nm is prepared on the platinum (Pt) metal by atomic layer deposition (ALD). Finally, a 40-nm silver (Ag) metal layer and a 20-nm gold (Au) metal layer are sequentially prepared on the alumina (Al) 2 O 3 dielectric layer by thermal evaporation and lift-off processes, and finally the Au / Ag / Al 2 O 3 / Pt / Ti diffusive memristor is obtained.

[0057] The task targeted in this embodiment is handwritten digit recognition, and the dataset used is the large handwritten digit database MNIST (Modified National Institute of Standards and Technology database) collected and sorted by the National Institute of Standards and Technology of the United States. After preprocessing by computer, this dataset has 60,000 training set pictures, and each picture consists of 28×28 binary pixels. In the Bayesian machine based on the diffusive memristor crossbar array circuit of the present invention, the diffusive memristor is used as the probability bit for probability calculation. First, it is necessary to modulate the input picture through the voltage-probability curve, and the input pulse voltage, that is, the modulation voltage, is obtained by modulation according to the gray value normalization of the picture pixel points, as Figure 1 shown. At room temperature, measurements are carried out on the Au / Ag / Al 2 O 3 / Pt / Ti diffusive memristor on a probe station equipped with an FS480 semiconductor parameter analyzer, and the voltage-probability curve of the diffusive memristor and the curve after non-linear transformation are as Figure 3As shown, based on this probability-voltage curve, the prior probabilities and conditional probabilities corresponding to the gray values of different pixels in the image can be modulated into voltages; the probabilities corresponding to the curve after non-linear transformation at low voltages are also increased, effectively avoiding zero-value underflow and improving the calculation efficiency.

[0058] The diffusive memristor crossbar array circuit structure used in the present invention is as Figure 4 shown, and includes a 1T1R layer, word lines, bit lines, and probability lines respectively. According to the sequential circuit logic, square wave pulses are input into the word lines and bit lines respectively to form a voltage difference. Based on the stochastic switching characteristics of the diffusive memristor, each probability bit corresponds to a prior probability or a conditional probability. When all the diffusive memristors on the bit line switch to the low-resistance state, the corresponding transistor is fully turned on, and the corresponding probability line of this bit line will be fully conductive. However, continuous random turn-on behavior will damage the diffusive memristor, causing it to lose the probability turn-on ability and become an ordinary resistor. To prevent this, when one of the bit lines operates in the forward voltage mode, all other bit lines are applied with reverse voltages to ensure that the diffusive memristor always maintains the probability switching operating mode.

[0059] As Figure 5 shown, it is a comparison of the accuracies of the Bayesian machine of the present invention in the MNIST classification task of a large handwritten digit database before and after non-linearly transforming the voltage-probability curve. Based on the probability calculation method of the present invention, when the sizes of the diffusive memristor crossbar arrays are 5×5, 7×7, 9×9, 14×14, and 28×28 respectively, the recognition rates are 66.3%, 76.7%, 80.6%, 83.2%, and 84.4% respectively. Compared with the Bayesian machine without non-linear transformation in the modulation process, the accuracies are improved by 1.1%, 0.4%, 0.6%, 0.5%, and 0.4% respectively. Compared with the traditional probability calculation method, the present invention has the computing ability to handle complex tasks and realizes the classification task of the large handwritten digit database MNIST, which proves the effectiveness of the present invention.

[0060] In addition, for those of ordinary skill in the art, the steps in the embodiments described in the present invention, whether all or part of them, can be implemented by programming instructions to corresponding hardware devices. The corresponding control programs can be stored in various computer-readable media, such as but not limited to read-only memories, hard disks, or optical discs, etc.

[0061] The above embodiments are the implementation manners of the present invention, but the implementation manners of the present invention are not limited by the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement manners and are all included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope defined by the claims.

Claims

1. A Bayesian machine hardware image classification method based on diffuse memristor, characterized in that: include: Step 1) preparing a plurality of diffuse memristors to construct a crossbar array circuit structure, and using each diffuse memristor in the crossbar array circuit structure as a probability bit of the Bayesian machine hardware; Step 2) obtaining a number of reference images with preset category labels and constructing an image dataset, and performing statistics on the binary brightness of each pixel point of each reference image in the image dataset to obtain the prior probability and conditional probability of the image dataset; Step 3) Based on the prior probability and conditional probability of each reference image under each preset category obtained in step 2), the image to be classified is subjected to voltage-probability modulation and processed in combination with a nonlinear transformation method to obtain a modulation voltage to be classified, the modulation voltage to be classified is input into a cross array circuit structure, and the image is classified according to the number of conduction times of the cross array circuit structure.

2. The image classification method based on diffuse memristor Bayesian machine hardware according to claim 1, characterized in that: In the step 1), when preparing a diffused memristor, a photolithography method is first used to pattern one side of a silicon substrate, and then at the patterned position, a titanium Ti metal layer and a platinum Pt metal layer are first deposited in sequence by a sputtering method, and the thickness range of the titanium Ti metal layer and the platinum Pt metal layer are both 1-40 nm, and then an aluminum oxide Al2O3 dielectric layer with a thickness of 7-20 nm is prepared on the platinum Pt metal by an atomic layer deposition method ALD, and finally a silver Ag metal layer and a gold Au metal layer are sequentially prepared on the aluminum oxide Al2O3 dielectric layer using thermal evaporation and stripping processes, and the thickness range of the silver Ag metal layer and the gold Au metal layer are both 20-40 nm, and the titanium Ti metal layer, the platinum Pt metal layer, the aluminum oxide Al2O3 dielectric layer and the silver Ag metal layer are used together as the bottom electrode, and the gold Au metal layer is used as the top electrode, and finally an Au / Ag / Al2O3 / Pt / Ti diffused memristor is prepared.

3. The image classification method based on diffuse memristor Bayesian machine hardware according to claim 1, characterized in that: In the step 1), in the diffuse memristor, when the compliance current is 1 nA, an external voltage of 5-10 V is applied, and the diffuse memristor switches from an initial high-resistance state to a low-resistance state. After the external voltage is removed, the diffuse memristor recovers to the initial high-resistance state in a time less than a preset time threshold to continue the voltage-probability modulation process; when performing image classification, the modulation voltage obtained after the voltage-probability modulation is within the probability switching range of the diffuse memristor, and the probability switching range is 0.07-0.14 V.

4. The image classification method based on diffuse memristor Bayesian machine hardware according to claim 1, characterized in that: In the step 1), the cross array circuit structure further comprises I bit lines, N word lines and I probability lines, each word line is arranged horizontally in parallel and spaced apart, each bit line and probability line are arranged vertically in parallel and spaced apart, each probability line is located between two adjacent bit lines, each word line is arranged vertically and crosswise with each bit line and probability line, a diffused memristor and a transistor are arranged in a rectangular area formed between every two adjacent bit lines and every two adjacent word lines, thereby forming an array unit, and for each array unit, the bottom electrode of the diffused memristor is connected to the current On the bit line shared with the previous array unit in the row, the top electrode of the diffused memristor is connected to the word line shared with the previous array unit in the current column, and the transistor is connected between the probability line and the bit line connected to the diffused memristor; one end of each word line and bit line is connected to a pulse generator, and one end of each word line receives a pulse input voltage, and the other end of each word line is connected to the diffused memristor in the last array unit of each row in sequence according to the arrangement order, the other end of each bit line serves as an output end, and both ends of each probability line are connected to a pulse receiving module.

5. The image classification method based on diffused memristor Bayesian machine hardware according to claim 4, characterized in that: In the step 2), the number of preset categories is the same as the number of bit lines and probability lines in the cross array circuit structure, and each image has its own preset category; the number of rows of pixel points in each image is the same as the number of word lines in the cross array circuit structure.

6. The image classification method based on diffuse memristor Bayesian machine hardware according to claim 4, characterized in that: In the step 2), each reference image has its own preset category label, that is, each pixel in each reference image has the same preset category label, and the brightness of each pixel in each reference image is binarized to obtain the binary brightness. For each reference image under each preset category, the prior probability and conditional probability of the binary brightness of each pixel in each reference image are obtained by statistics.

7. The image classification method based on diffused memristor Bayesian machine hardware according to claim 4, characterized in that: In the step 3), firstly, the binary brightness of each pixel in the image to be classified is obtained based on the prior probability and conditional probability of each reference image under each preset category obtained in step 2), and the classification probability of each pixel belonging to each preset category is obtained; the initial switching voltage-probability curve of the diffused memristor is nonlinearly changed to an updated switching voltage-probability curve, as follows: Wherein, G( ) is the updated switching voltage-probability curve, i.e., the classification probability of the preset category; a and p are the first and second range limit parameters in the nonlinear transformation process, respectively; V d is the modulation voltage after normalizing the probabilistic switching range of the diffuse memristor; The modulation voltage of each pixel in the image to be classified is obtained according to the updated switching voltage-probability curve; for each preset category, a preset number of classification and recognition are performed, and for each classification and recognition, the modulation voltage of each row of pixels in the image to be classified is first input into one of their respective word lines in turn, and at the same time, the column selection pulse emitted by the pulse generator is input into the bit line corresponding to the current preset category, so that a voltage difference is formed between the word line and the bit line, and the pulse receiving module saves the number of times the probability line connected to the same diffused memristor as the current bit line is turned on, until the preset number of classification and recognition are completed, and the total number of times the current probability line is turned on is obtained; finally, the number of times the probability lines are turned on in each preset category is obtained, and the preset category to which the probability line with the largest number of turns on belongs is taken as the category of the current image.

8. The image classification method based on diffused memristor Bayesian machine hardware according to claim 7, characterized in that: During the operation of the diffused memristor, when one of the bit lines operates in a forward voltage mode, reverse voltages are applied to all other bit lines, so that the diffused memristor always maintains a probabilistic switching operation mode.

9. The image classification method based on diffused memristor Bayesian machine hardware according to claim 7, characterized in that: When the probability line is turned on, all the diffusion memristors in the current column connected to the current probability line are turned on. During the operation of the diffusion memristor, the final turn-on state is as follows: Among them, p m is the final on-state of the mth diffusion memristor, p m =0 means not turned on, p m =1 means on; u() is a binary step function; V md is the modulation voltage input to the mth diffused memristor; V o is the boundary voltage of the probability switching range; V S is the preset scaling voltage; rand is the preset random item.

Images