Probabilistic artificial synaptic unit, 1T1R array structure and neural network information recognition method

By introducing probabilistic artificial synaptic units and 1T1R array structures into hardware neural networks, the synaptic state is controlled by randomness of volatile memory, and the problem of unreliable recognition results caused by model perception uncertainty in the prior art is solved. , achieving higher recognition accuracy and uncertainty evaluation, and improving the recognition reliability of key applications.

CN114169509BActive Publication Date: 2024-11-19HUAZHONG UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111432442.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-11-29
Publication Date
2024-11-19
Estimated Expiration
2041-11-29

AI Technical Summary

Technical Problem

The existing memristor-based hardware neural network lacks consideration of model perception uncertainty, resulting in insufficient reliability of identification results in key application areas such as clinical medical diagnosis and intelligent driving.

Method used

Design a probabilistic artificial synaptic unit, including volatile memristors, voltage-dividing resistors and transistors, realize the probability control of synapses through random high and low resistance switching of volatile memristors, build a 1T1R array structure, and pass The neural network model performs information recognition, calculates and evaluates the uncertainty of the recognition results to improve recognition accuracy and reliability.

Benefits of technology

By probability controlling the randomness of synapses, the recognition accuracy of neural networks and the accurate evaluation of output uncertainty are improved, and the reliability of identification results in key application areas is improved, especially in the fields of clinical medical diagnosis and intelligent driving.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114169509B_ABST
    Figure CN114169509B_ABST
Patent Text Reader

Abstract

The present application discloses a probabilistic artificial synaptic unit, including a volatile memristor, a voltage-dividing resistor, a transistor and a non-volatile memristor, one end of the volatile memristor is used as a synaptic control end to obtain a control signal, and the other end is grounded through a voltage-dividing resistor; one end of the non-volatile memristor is used as a synaptic input end, and the other end is connected to one of the source / drain ends of the transistor, and the other end of the source / drain end of the transistor is used as a synaptic output end, and the gate of the transistor is connected to the connection end of the volatile memristor and the voltage-dividing resistor. The present application also relates to a 1T1R array structure constructed based on the above-mentioned probabilistic artificial synaptic unit, and a neural network information recognition method based on the above-mentioned 1T1R array structure. The neural network model in the present application considers the perception uncertainty of the model itself on the basis of the traditional model, so the recognition rate and uncertainty reliability of the recognition result obtained are higher.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application belongs to the field of semiconductor information technology, and more specifically, relates to a probabilistic artificial synaptic unit, a 1T1R array structure and a neural network information recognition method. Background Art

[0002] The bottleneck problem of separation of storage and computing in the traditional von Neumann computing architecture has become prominent, and the existing architecture and devices cannot meet the demand for computing power in the era of artificial intelligence due to the explosive growth of information. Compared with traditional transistors, memristors have the characteristics of low power consumption, high speed, high integration, and multiple conductance states. Brain-like computing is based on the design of efficient artificial synapses and neurons based on memristors, directly building hardware neural networks.

[0003] The uncertainty of brain-like computing is divided into accidental uncertainty and perceived uncertainty. Accidental uncertainty is caused by the inherent noise in the input information, which cannot be eliminated, while perceived uncertainty is related to the model and is caused by insufficient model knowledge and incomplete training. However, the current research on hardware neural networks based on memristors is mainly concentrated in the traditional deterministic computing field of artificial neural networks, that is, the model is considered to be deterministic, and the uncertainty of the output result is only calculated based on the noise of the sample itself, but the uncertainty of the model perception is not considered, making the final uncertainty unreliable. In particular, in several key application areas such as clinical medical diagnosis and intelligent driving decision-making, the demand for reliability of probabilistic computing is more obvious. Summary of the invention

[0004] In response to the above defects or improvement needs of the prior art, the present application provides a probabilistic artificial synaptic unit, a 1T1R array structure and a neural network information recognition method, the purpose of which is to improve the reliability of brain-like computing of artificial neural networks.

[0005] To achieve the above-mentioned object, according to a first aspect of the present application, a probabilistic artificial synaptic unit is provided, which includes a volatile memristor, a voltage divider resistor, a transistor and a non-volatile memristor, wherein:

[0006] One end of the volatile memristor is used as a synaptic control end to obtain a control signal, and the other end is grounded through the voltage-dividing resistor;

[0007] One end of the non-volatile memristor serves as a synaptic input end, and the other end is connected to one of the source / drain ends of the transistor, the other end of the source / drain end of the transistor serves as a synaptic output end, the gate of the transistor is connected to the connection end of the volatile memristor and the voltage divider resistor, and the on and off states of the transistor are switched as the high and low resistance states of the volatile memristor are switched.

[0008] Preferably, the volatile memristor is a TiW / CuS / GeSe / Pt conductive bridge threshold switching memristor.

[0009] According to a second aspect of the present application, a 1T1R array structure is provided, comprising:

[0010] A plurality of bit lines and a plurality of word lines are arranged in a cross pattern;

[0011] The probabilistic artificial synapse unit as described above is located at the intersection node of the bit line and the word line, the synaptic input terminal of the probabilistic artificial synapse is connected to the bit line, the synaptic control terminal of the probabilistic artificial synapse is connected to the word line, and the synaptic output terminal of the probabilistic artificial synapse is connected to the source line.

[0012] According to a third aspect of the present application, a neural network information recognition method is provided, comprising:

[0013] Acquire a neural network model, wherein the neural network model is a neural network model obtained by training a neural network including the 1T1R array structure;

[0014] Input the information to be recognized x into the neural network model multiple times, and obtain the recognition result classification probability p obtained by the t-th input t (y=c|x), where y represents the output information of the neural network model, and c is the category corresponding to the output information;

[0015] Calculate the final recognition result classification probability

[0016]

[0017] Wherein, T represents the number of times the information to be identified x is input.

[0018] Preferably, it also includes calculating the uncertainty of the output result according to the classification probability of the final recognition result

[0019] H pred (y|x)=-∑p(y=c|x)logp(y=c|x).

[0020] Preferably, it also includes:

[0021] Select the uncertainty threshold;

[0022] When the uncertainty of the current output result classification is equal to or greater than the threshold, the current output result classification is determined to be unreliable, and when the uncertainty of the current output result classification is less than the threshold, the current output result classification is determined to be credible.

[0023] Preferably, the threshold is 0.5.

[0024] Preferably, in the neural network model, the conduction probability of each probabilistic artificial synaptic unit in the same layer is the same.

[0025] Preferably, the conduction probability of the probabilistic artificial synapse unit is 0.5 or 0.7.

[0026] Preferably, in the neural network model, the control signal of the corresponding probabilistic artificial synapse unit is adjusted according to the conduction probability of the probabilistic artificial synapse unit.

[0027] In general, the present application relates to a probabilistic artificial synaptic unit, which connects the gate of a transistor to a volatile memristor. Based on the properties of the volatile memristor itself, the turn-on voltage threshold of the volatile memristor changes randomly. When a fixed control signal is connected, the volatile memristor is in a high-resistance state or a low-resistance state with a certain randomness, resulting in a certain randomness in the on and off states of the transistor. That is, when a control signal is applied, whether the corresponding synaptic unit can transmit the synaptic signal has a certain randomness, and a probabilistically controllable synaptic random selection can be achieved. Based on the 1T1R array structure constructed by the probabilistic artificial synaptic unit, a neural network is constructed without changing the cross array structure of the 1T1R memristor, and information recognition is performed through the neural network. Compared with the traditional technology in which the synapse is in a certain state, the present application can be different when the neural network is inferred, and the effective synapse is different each time the forward propagation is carried out, which can improve the recognition accuracy of the neural network and accurately evaluate the uncertainty of the network output. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a schematic diagram of the structure of a probabilistic artificial synapse in one embodiment of the present application;

[0029] FIG. 2( a ) is a current-voltage scanning curve of a volatile memristor in one embodiment of the present application;

[0030] FIG2( b ) is a threshold voltage distribution diagram of a volatile memristor in an embodiment of the present application;

[0031] Figure 3 is a simulation diagram of the electrical performance of a probability transistor in an embodiment of the present application;

[0032] FIG4( a ) is a schematic diagram of a 1T1R array structure in an embodiment of the present application;

[0033] FIG4( b ) is a schematic diagram of an artificial neural network based on probabilistic synapses in one embodiment of the present application;

[0034] Figure 5 is a flowchart of the steps of a neural network information recognition method in one embodiment of the present application;

[0035] FIG6( a ) is a comparison diagram of the accuracy of artificial neural networks based on probabilistic synapses and standard synapses in one embodiment of the present application;

[0036] FIG6( b ) is a relative probability density distribution diagram of the uncertainty of correct classification results and incorrect classification results in one embodiment of the present application. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solutions and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0038] like Figure 1 The structure diagram of the probabilistic artificial synapse unit in one embodiment of the present application is shown, which includes a volatile memristor M1, a voltage-dividing resistor R1, a transistor T1 and a non-volatile memristor M2. The volatile memristor M1 is used to receive a control signal and probabilistically jump to a low-resistance state based on its non-fixed switch threshold. One end of the voltage-dividing resistor R1 is connected to the volatile memristor M1, and the other end is grounded, which is used to control the voltage division and determine the actual voltage signal of the gate of the transistor T1. The gate of the transistor T1 is connected to the voltage-dividing resistor R1, and is used to control whether the signal at the source and drain of the transistor T1 passes based on its own threshold voltage and the random voltage signal from the voltage-dividing resistor R1. When the random voltage signal is greater than the threshold of the transistor T1, the transistor T1 is in the on state, and when the random voltage signal is less than the threshold of the transistor T1, the transistor T1 is in the off state. One end of the non-volatile memristor M2 is connected to the transistor T1, and the other end receives the input signal. Its conductance state is used to store the synaptic weight of the neural network. When the transistor T1 is in the on state, the signal input from the previous level neuron is transmitted to the next level neuron through the non-volatile memristor M2. When the transistor T1 is in the off state, the signal input from the previous level neuron cannot be transmitted to the next level neuron through the non-volatile memristor M2, and the synapse fails. The conductance of the non-volatile memristor M2 is continuously adjustable.

[0039] Among them, the volatile memristor M1 with a random threshold is used to realize the probabilistic on / off of the transistor. As shown in Figure 2(a), after the voltage at both ends of the volatile memristor M1 is higher than the threshold, the resistance state quickly changes from a high resistance state to a low resistance state, and the threshold voltage is different at each transition. When the voltage at both ends of the memristor M1 is lower than the holding voltage, the resistance state will spontaneously change from a low resistance state to a high resistance state. In this embodiment, the volatile memristor M1 uses a TiW / CuS / GeSe / Pt conductive bridge threshold conversion memristor; as shown in Figure 2(b), the threshold voltage of the volatile memristor M1 shows randomness, and the threshold voltages of the memristor that are continuously turned on in the voltage scanning cycle are unrelated.

[0040] It should be noted that the probability of a random event that the threshold voltage of the volatile memristor can be controlled according to the magnitude of the input control voltage, and the probability that the volatile memristor is in a low resistance state is related to the magnitude of the control voltage. The probability that the volatile memristor is in a low resistance state at a certain voltage can be measured through a large number of experiments, and the corresponding relationship between the probability of the volatile memristor being in a low resistance state and the control voltage can be obtained, that is, the corresponding relationship between the probability of the probabilistic transistor being turned on and the control voltage, and then the corresponding control voltage can be selected according to the required probability.

[0041] The resistance value of the voltage divider resistor R1 is obtained according to the high and low resistance states of the volatile memristor M1 and the threshold voltage of the transistor T1. One end of the voltage divider resistor R1 is grounded, and the other end is connected to the gates of the volatile memristor M1 and the transistor T1. The gate voltage of the transistor is the same as the voltage of the voltage divider resistor R1 to ground. The structure composed of the volatile memristor M1, the voltage divider resistor and the transistor T1 is called a probability transistor. The voltage of the control signal connected from the volatile memristor M1 is used as the gate voltage of the probability transistor, as shown in Figure 3 As shown, when the volatile memristor M1 is in a high-resistance state, the voltage of the voltage-dividing resistor R1 to the ground is lower than the threshold of the transistor T1, the transistor T1 is cut off, the input signal cannot be transmitted to the next-level neuron through the synapse, and the synapse fails; when the volatile memristor M1 is in a low-resistance state, the voltage of the voltage-dividing resistor R1 to the ground is higher than the threshold of the transistor T1, the transistor T1 is turned on, the input signal is transmitted to the next-level neuron through the non-volatile memristor M2 and the transistor T1, and the synapse works normally. In this application, the high and low resistance states of the volatile memristor used vary greatly, resulting in a large change in the gate voltage of the transistor before and after the volatile memristor reaches the resistance state conversion threshold voltage. This large change in gate voltage makes the transistor have a very small subthreshold slope (much less than 60mV / dec), and the transistor has a very small leakage current before turning on.

[0042] The present application also relates to a 1T1R array structure, which is constructed based on the above-mentioned probabilistic artificial synapse unit. Specifically, as shown in FIG4(a), the 1T1R array structure includes a plurality of bit lines and a plurality of word lines that cross each other, the probabilistic artificial synapse unit is located at the intersection of the bit line and the word line, the synaptic input end of the probabilistic artificial synapse is connected to the bit line, the synaptic control end of the probabilistic artificial synapse is connected to the word line, and the synaptic output end of the probabilistic artificial synapse is connected to the source line.

[0043] In the present application, without changing the basic structure of the 1T1R memristor cross array and the peripheral circuits of the array, the probabilistic synapses of the present application are used to replace the traditional deterministic synapses. The corresponding neural network structure is shown in Figure 4(b). During the forward propagation of the neural network, the signal from the neuron transmitter is transmitted to the neuron receiver through the synapse. When there are M transmitters and N receivers, there are M*N synaptic paths, and each synapse will be randomly selected according to the set probability.

[0044] The present application also relates to a neural network information recognition method, such as Figure 5 As shown, the method includes:

[0045] Step S100: Acquire a neural network model, where the neural network model is a neural network model trained based on a neural network including a 1T1R array structure.

[0046] First, a trained neural network model is obtained, wherein the synapses in the neural network model adopt the 1T1R memristor cross array structure mentioned above.

[0047] During the training of the model, the corresponding relationship between the probability of conduction of the probabilistic transistor and the control signal is first established. The corresponding control signal is connected according to the target conduction probability required by the probabilistic artificial synaptic unit, and a neural network based on random synapses is built. Then, the network is trained to obtain a trained neural network model.

[0048] When using the trained neural network model for identification, the control signal connected to each probabilistic artificial synapse unit is the same as the control signal connected during training. In one embodiment, the conduction probability of each probabilistic artificial synapse unit in the same layer is the same, and the conduction probability of probabilistic artificial synapse units in different layers may be different. Specifically, the conduction probability of the probabilistic artificial synapse unit is preferably 0.5 or 0.7. Experiments have shown that the recognition effect of the neural network under this value is better. In this neural network model, the effectiveness and failure of the synapse have a certain probability of randomness, so the network results are more diversified and the training effect is better.

[0049] Step S200: Input the information to be recognized x into the neural network model multiple times, and calculate the recognition result classification probability p obtained by the t-th input t (y=c|x).

[0050] After obtaining the neural network model, the information to be recognized x is input into the neural network model multiple times, and the classification probability p of the recognition result output each time is calculated. t (y=c|x), where y is the output result of the neural network, x is the input data of the neural network, and c is the category to which the data belongs.

[0051] Step S300: Calculate the final recognition result classification probability

[0052] Wherein, T is the number of times the information to be identified x is input, and specifically, T may be 20.

[0053] In this application, since the neural network model has a certain randomness, multiple inputs are required and multiple results are weighted averaged to improve the recognition rate. In this embodiment, the neural network corresponding to this hardware structure is simulated by software to realize the recognition of Cifar10 data. The recognition rate of the neural network is shown in Figure 6 (a), and the recognition rate of the neural network using probabilistic synapses is improved.

[0054] In one embodiment, the present application further includes a quantitative calculation of uncertainty, specifically:

[0055] Step S400: Calculate the uncertainty H of the output result based on the final recognition result classification probability pred (y|x)=-∑p(y=c|x)logp(y=c|x).

[0056] The uncertainty of the output result of the neural network model is calculated through step S400. The higher the uncertainty, the lower the trustworthiness of the current result and the higher the error rate of the current recognition result; the lower the uncertainty, the higher the trustworthiness of the current result and the lower the error rate of the current recognition result. Since the neural network model in this application considers the perceptual uncertainty of the model itself on the basis of the traditional model, the uncertainty of the recognition result obtained based on this neural network model is more reliable. In one embodiment, the uncertainty quantification result of the neural network is shown in Figure 6(b), which shows the prediction uncertainty H of all test images drawn using the kernel density estimation of the Gaussian kernel. pred, grouped by correct and incorrect predictions, a total of 20 probability forward propagations (T=20) were performed. The uncertainty of the misclassified data is significantly higher than that of the correctly classified data. That is to say, if the uncertainty of the recognition result obtained by the model is low, it means that the current recognition result is most likely correct and can be trusted to a certain extent. If the uncertainty is high, it means that the recognition result is most likely wrong and the credibility is not high, and further analysis and confirmation is required. In one embodiment, when the uncertainty of the current output result classification is equal to or greater than the threshold, the current output result classification is determined to be unreliable. When the uncertainty of the current output result classification is less than the threshold, the current output result classification is determined to be credible. Specifically, the threshold can be 0.5. If the uncertainty of 0.5 is used as the threshold, 81.3% of classification errors can be avoided. Recognition results with uncertainties lower than 0.5 are classified as trustworthy results, and recognition results with uncertainties higher than or equal to 0.5 are classified as untrustworthy results. We need to further detect and analyze this data. Therefore, when performing information recognition, especially in several key application areas such as clinical medical diagnosis and intelligent driving decision-making, the use of the above neural network model and recognition method can effectively improve recognition reliability.

[0057] In summary, the present application relates to a probabilistic artificial synaptic unit, by connecting the gate of the transistor to a volatile memristor, based on the properties of the volatile memristor itself, the turn-on voltage threshold of the volatile memristor changes randomly, and when a fixed control signal is connected, the volatile memristor is in a high-resistance state or a low-resistance state with a certain randomness, resulting in a certain randomness in the on and off states of the transistor, that is, when a control signal is applied, whether the corresponding synaptic unit can transmit the synaptic signal has a certain randomness, and a probabilistic controllable synaptic random selection can be achieved. Based on the 1T1R array structure constructed by the probabilistic artificial synaptic unit, a neural network is constructed on the basis of not changing the cross array structure of the 1T1R memristor, and information recognition is performed through the neural network. Compared with the traditional technology in which the synapse is in a certain state, the present application can be different when the neural network is inferred, and the effective synapse is different each time the forward propagation is carried out, which can improve the recognition accuracy of the neural network and accurately evaluate the uncertainty of the network output.

[0058] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A probabilistic artificial synaptic unit, characterized in that: It includes volatile memristors, voltage divider resistors, transistors and non-volatile memristors, among which: One end of the volatile memristor is used as a synaptic control end to obtain a control signal, and the other end is grounded through the voltage-dividing resistor; One end of the non-volatile memristor serves as a synaptic input end, and the other end is connected to one of the source / drain ends of the transistor, the other end of the source / drain end of the transistor serves as a synaptic output end, the gate of the transistor is connected to the connection end of the volatile memristor and the voltage divider resistor, and the on and off states of the transistor are switched as the high and low resistance states of the volatile memristor are switched.

2. The probabilistic artificial synapse unit according to claim 1, characterized in that The volatile memristor is a TiW / CuS / GeSe / Pt conductive bridge threshold conversion memristor.

3. A 1T1R array structure, characterized in that: include: A plurality of bit lines and a plurality of word lines are arranged in a cross pattern; The probabilistic artificial synapse unit according to claim 1 or 2 is located at the intersection node of the bit line and the word line, the synapse input terminal of the probabilistic artificial synapse unit is connected to the bit line, the synapse control terminal of the probabilistic artificial synapse unit is connected to the word line, and the synapse output terminal of the probabilistic artificial synapse unit is connected to the source line.

4. A neural network information recognition method, characterized in that: include: Acquire a neural network model, wherein the neural network model is a neural network model obtained by training a neural network including the 1T1R array structure in claim 3; The information to be identified x Input the neural network model multiple times to obtain the t The classification probability of the recognition result obtained by input ,in, y represents the output information of the neural network model, c The category corresponding to the output information; Calculate the final recognition result classification probability in, T Indicates that the information will be identified x The number of times entered.

5. The neural network information recognition method according to claim 4, characterized in that: It also includes calculating the uncertainty of the output result based on the classification probability of the final recognition result 。 6. The neural network information recognition method according to claim 5, characterized in that: Also includes, Select the uncertainty threshold; When the uncertainty of the current output result classification is equal to or greater than the threshold, the current output result classification is determined to be unreliable, and when the uncertainty of the current output result classification is less than the threshold, the current output result classification is determined to be credible.

7. The neural network information recognition method according to claim 6, characterized in that: The threshold is 0.

5.

8. The neural network information recognition method according to claim 4, characterized in that: In the neural network model, the conduction probability of each probabilistic artificial synaptic unit in the same layer is the same.

9. The neural network information recognition method according to claim 8, characterized in that: The conduction probability of the probabilistic artificial synaptic unit is 0.5 or 0.

7.

10. The neural network information recognition method according to claim 4, characterized in that: In the neural network model, the control signal of the corresponding probabilistic artificial synaptic unit is adjusted according to the conduction probability of the probabilistic artificial synaptic unit.