A reservoir computing system based on memcapacitors

Abstract

The present disclosure provides a reservoir computing system based on memcapacitors. The reservoir computing system based on memcapacitors comprises: an input mask module comprising a mask unit, the mask unit being configured to modulate an amplitude of an input signal and determine a voltage signal proportional to the amplitude of the input signal; a reservoir module comprising reservoir units, each reservoir unit comprising a memcapacitor, each memcapacitor being configured to construct a perisach model based on the memcapacitor, an output end of each perisach model based on the memcapacitor outputting a polarization intensity of the memcapacitor, and the perisach model based on the memcapacitor serving as a node of a reservoir network; and an output module configured to obtain output weights of the reservoir network in a training stage of the reservoir network, multiply the polarization intensity of the memcapacitor with the output weights of the reservoir network in a test stage of the reservoir network, and output a prediction result. By the present disclosure, the reservoir network is constructed by using the memcapacitors, thereby improving the working efficiency of the reservoir computing system and reducing the static power consumption.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of reservoir computing, in particular, to a reservoir computing system based on a memistor. BACKGROUND

[0002] Artificial intelligence is developing rapidly, and the explosive growth of data volume puts new requirements on computing power. Due to its structural characteristics, the traditional feedforward neural network cannot remember the short-term characteristics of signals, and it is widely used in processing static spatial image signals, such as image recognition, etc., but it is not suitable for processing time signals. In the recurrent neural network, the recursive structure can realize the function of short-term memory, and it is suitable for processing space-time signals, but the recursive structure not only consumes a large amount of computing resources, but also the problems of gradient explosion or disappearance make the training of the recurrent neural network very difficult.

[0003] Therefore, the concept of reservoir is proposed. The reservoir network fixes the weights of the recurrent neural network except the output layer, and only trains the output layer weight. The training will become linear, and the output weight of the reservoir network can be calculated by using a simple algorithm such as linear regression method. In this way, the problems of gradient disappearance or gradient explosion caused by recursive connection can be solved, and the consumption of computing resources can be greatly reduced.

[0004] The traditional reservoir neural network uses random connection of reservoir nodes to obtain the ability to map time signals to high-dimensional space. Such structure is not only difficult to implement at the hardware level, but also the richness of the reservoir state cannot be adjusted after the structure of the reservoir network is fixed, and the performance of the reservoir cannot be optimized. SUMMARY

[0005] In view of the defects in the prior art, the purpose of the present disclosure is to provide a reservoir computing system based on a memistor.

[0006] To achieve the above-mentioned purpose, according to one aspect of the present disclosure, a reservoir computing system based on a memistor is provided, comprising:

[0007] An input mask module comprising n parallel mask units, the mask units being used for modulating and processing the amplitude of the input signal to determine that the amplitude is proportional to the voltage signal of the input signal;

[0008] a reserve pool module, comprising n parallel reserve pool units, each of the reserve pool units comprising a memcapacitor, a perisach model based on the memcapacitor is constructed according to parameters of the memcapacitor, an input end of each of the perisach models based on the memcapacitor is connected with an output end of one of the mask units, and an output end of each of the perisach models based on the memcapacitor outputs a polarization intensity of the memcapacitor, and the perisach model based on the memcapacitor serves as a node of the reserve pool network;

[0009] an output module, which obtains output weights of the reserve pool network in a training stage of the reserve pool network, and performs multiplication processing on the polarization intensity of the memcapacitor and the output weights of the reserve pool network and outputs a prediction result in a test stage of the reserve pool network.

[0010] Optionally, the mask unit generates a mask signal in a random manner, and the mask signal is a discrete matrix with a maximum element of 1.

[0011] Optionally, the memcapacitor in each of the perisach models based on the memcapacitor dynamically responds to the input voltage signal, establishes a feedback connection between different time step signals, and changes the voltage signal to a high-dimensional space in a nonlinear change processing manner to determine the polarization intensity of the memcapacitor.

[0012] Optionally, a semiconductor parameter analyzer is used to obtain the parameters of each of the memcapacitors.

[0013] Optionally, the memcapacitors in the reserve pool module use a ferroelectric capacitor based on Zr-doped HfO2 (HZO).

[0014] Optionally, the ferroelectric capacitor based on Zr-doped HfO2 (HZO) comprises TiN as a top electrode (TE), HZO as a dielectric layer, and TiN as a bottom electrode (BE).

[0015] Optionally, the HZO as the dielectric layer is obtained by an atomic layer deposition method at a temperature of 250°C and a dose time of 0.1 seconds.

[0016] Optionally, the precursors of the HZO as the dielectric layer are TEMAH and TEMAZ, in the process of obtaining the HZO as the dielectric layer, HfO2 and ZrO2 are circulated, the concentration of Zr in the HZO thin film is the same as that of Hf, after sputtering the TiN as the top electrode (TE) on the HZO as the dielectric layer, the HZO as the dielectric layer is annealed at a temperature of 550°C for 30 seconds in a nitrogen environment.

[0017] Optionally, the output module multiplies the target matrix and the pseudo-inverse of the reservoir state matrix in the training phase of the reservoir network to obtain the output weight of the reservoir network.

[0018] Optionally, the output module trains the output weight of the reservoir network using a linear regression algorithm.

[0019] Compared with the prior art, the embodiments of the present disclosure have at least one of the following beneficial effects:

[0020] Through the above technical solution, the input mask module uses the parallel mask units to perform amplitude modulation processing on the input signal to determine the voltage signal proportional to the amplitude of the input signal, the reservoir module uses the memcapacitor included in the parallel reservoir unit to construct the perisach model based on the memcapacitor, the perisach model based on the memcapacitor as the node of the reservoir network, and the perisach model based on the memcapacitor is used to perform nonlinear transformation on the voltage signal from the mask unit to obtain the polarization intensity of the memcapacitor, and the output module obtains the output weight of the reservoir network in the training phase of the reservoir network, and multiplies the polarization intensity of the memcapacitor and the output weight of the reservoir network in the test phase of the reservoir network and outputs the prediction result. The present disclosure uses the memcapacitor to construct the reservoir network, and the leakage current of the memcapacitor is small, which improves the working efficiency of the reservoir computing system while reducing the static power consumption. BRIEF DESCRIPTION OF DRAWINGS

[0021] Other features, objects, and advantages of the present disclosure will become more apparent from the following detailed description of non-limiting embodiments, made with reference to the accompanying drawings:

[0022] Figure 1 is a block diagram of a reservoir computing system based on a memcapacitor according to an example embodiment.

[0023] Figure 2 is a functional schematic diagram of a reservoir computing system based on a memcapacitor according to an example embodiment.

[0024] Figure 3 is a schematic diagram of the polarization intensity voltage curve simulation effect of a perisach model based on a memcapacitor according to an example embodiment.

[0025] Figure 4 is a structural schematic diagram of a memcapacitor device according to an example embodiment.

[0026] Figure 5 is a result schematic diagram of the identification of the MNIST handwritten digit set by a reservoir computing system based on a memcapacitor according to an example embodiment. DETAILED DESCRIPTION

[0027] The present disclosure will be described in detail below with reference to specific embodiments. The following examples will facilitate further understanding of the present disclosure for those skilled in the art, but do not limit the present disclosure in any form. It should be noted that, for those skilled in the art, without departing from the concept of the present disclosure, a number of modifications and improvements can be made. These all belong to the protection scope of the present disclosure.

[0028] Figure 1 is a block diagram of a reservoir computing system based on a memistor according to an exemplary embodiment. As shown in Figure 1 , the present disclosure provides a reservoir computing system based on a memistor, comprising an input mask module, a reservoir module, and an output module.

[0029] The input mask module comprises n parallel mask units, and the mask units are used to modulate the amplitude of the input signal and determine the voltage signal proportional to the amplitude of the input signal.

[0030] In some possible embodiments, the mask units generate mask signals in a random manner, and the mask signals are discrete matrices with a maximum element of 1.

[0031] The mask units can generate different mask signals to modulate the amplitude of the input signal, and the determined voltage signal proportional to the amplitude of the input signal is V in .

[0032] The reservoir module comprises n parallel reservoir units, and each reservoir unit comprises a memistor. A perisach model based on the memistor is constructed according to the parameters of the memistor. The input end of each perisach model based on the memistor is connected to the output end of a mask unit, and the output end of each perisach model based on the memistor outputs the polarization intensity of the memistor. The perisach model based on the memistor serves as a node of the reservoir network.

[0033] Each reservoir unit is a single-node delay reservoir unit, and the amplitude proportional to the voltage signal V in of the input signal output by each mask unit is input into the corresponding perisach model based on the memistor.

[0034] In some possible embodiments, the memistor in each perisach model based on the memistor dynamically responds to the input voltage signal V in , establishes a feedback connection between different time step signals, and changes the voltage signal to a high-dimensional space in a nonlinear manner to determine the polarization intensity of the memistor.

[0035] The voltage signal is changed to a high-dimensional space by using a nonlinear change processing manner, so that it is linearly separable, and the reserve pool module outputs the determined polarization intensity of the memistor to the output module.

[0036] The output module trains the output weight of the reserve pool network by using a linear regression algorithm. The output module obtains the output weight of the reserve pool network in the training stage of the reserve pool network, and performs product processing on the polarization intensity of the memistor and the output weight of the reserve pool network in the test stage of the reserve pool network, and outputs a prediction result.

[0037] In the training stage of the reserve pool network, the output module performs product processing on the pseudo-inverse of the target matrix and the reserve pool state matrix, to obtain the output weight of the reserve pool network.

[0038] Through the above technical solutions, the input mask module uses parallel mask units to perform amplitude modulation processing on the input signal to determine the voltage signal proportional to the amplitude of the input signal. The reserve pool module uses the memistor included in the parallel reserve pool unit to construct a perisach model based on the memistor. The perisach model based on the memistor is used as a node of the reserve pool network, and the perisach model based on the memistor is used to perform nonlinear change processing on the voltage signal from the mask unit to obtain the polarization intensity of the memistor. The output module obtains the output weight of the reserve pool network in the training stage of the reserve pool network, and performs product processing on the polarization intensity of the memistor and the output weight of the reserve pool network in the test stage of the reserve pool network, and outputs a prediction result. The present disclosure uses a memistor to construct a reserve pool network. The leakage current of the memistor is small, which improves the working efficiency of the reserve pool computing system while reducing the static power consumption.

[0039] Figure 2 is a functional schematic diagram of a reserve pool computing system based on a memistor according to an example embodiment.

[0040] As shown in Figure 2 In some possible embodiments, the input mask module can be used as the input layer of the reserve pool network. The input mask module uses time division multiplexing to feed the virtual nodes of the reserve pool network and obtain rich reserve pool states.

[0041] As an example, a random mask matrix can be defined, and the input signal is multiplied with the mask matrix.

[0042] One time step of the input signal is τ, and the input signal is defined as I(t0). The random mask matrix is M(NxQ), where N represents the number of virtual nodes (i.e., the number of nodes of the reserve pool network), and Q represents the dimension of the input signal.

[0043] At time t0, the following is performed:

[0044] J(t0) = M x I(t0) when,

[0045] wherein J(t0) represents the input of the reservoir network in the time interval [t0, t0+τ], J(t0) is an N-dimensional vector, and each virtual node of the reservoir network is updated using the corresponding component of J(t0).

[0046] In some possible embodiments, a semiconductor parameter analyzer can be used to obtain the parameters of each memcapacitor. For example, a Keysight E4990A semiconductor parameter analyzer can be used to detect each memcapacitor and obtain the parameters of each memcapacitor.

[0047] In some possible embodiments, a MATLAB tool can be used to construct a perisach model based on the parameters of the memcapacitor.

[0048] First, an instance of the Preisach_node class is created in the MATLAB tool, and its attributes are initialized. Second, in the constructor, a perisach model of N domains is generated, and finally, it is stored in a model dictionary model_dict.

[0049] Figure 3 is a schematic diagram of the simulation effect of the polarization intensity voltage curve of the memcapacitor based on the perisach model of the memcapacitor according to an example embodiment.

[0050] As shown in Figure 3 For each input voltage signal, the perisach model domains are traversed, and the polarization state of each domain is updated according to the size of the voltage of the voltage signal.

[0051] As an example, if the voltage of the voltage signal is less than the reverse flip threshold of the perisach model, the polarization state of the domain is updated to -1.

[0052] As another example, if the voltage of the voltage signal is greater than the positive flip threshold of the perisach model, the polarization state of the domain is updated to 1.

[0053] The polarization states of each domain are counted to obtain the polarization intensity of the memcapacitor under the input signal.

[0054] In some possible embodiments, the memcapacitor has short-term memory characteristics and nonlinear characteristics, and the memcapacitor can also map the input signal to a high-dimensional linearly separable space and generate rich reservoir states. In the present disclosure, the memcapacitor takes the voltage signal output by the input mask unit as the input signal of the reservoir module, that is, the voltage signal V inThe dynamic response is performed and the non-linear change processing is mapped to a high-dimensional space.

[0055] By the technical solution, the input signal is generated in a time division multiplexing manner, and a rich reserve pool state is generated, so that the performance of the reserve pool is optimized.

[0056] Figure 4 It is a structure diagram of a memcapacitor device according to an exemplary embodiment. Wherein, 1 represents a substrate, 2 represents TiN of a bottom electrode (BE), 3 represents HZO of a dielectric layer, and 4 represents TiN of a top electrode (TE).

[0057] As shown in Figure 4 In some possible embodiments, the memcapacitor in the reserve pool module adopts a ferroelectric capacitor based on Zr-doped HfO2 (HZO).

[0058] The ferroelectric capacitor based on Zr-doped HfO2 (HZO) includes TiN as a top electrode (TE), HZO as a dielectric layer, and TiN as a bottom electrode (BE).

[0059] The TiN of the top electrode (TE) is 30 nm, the HZO as the dielectric layer is 6 nm, and the TiN of the bottom electrode (BE) is 30 nm.

[0060] The memcapacitor can be obtained on an n-doped silicon substrate, wherein the top electrode (TE) and the bottom electrode (BE) are obtained by a magnetron sputtering machine, and the HZO of the dielectric layer is obtained by an atomic layer deposition method at a temperature of 250 DEG C and a dose time of 0.1 seconds.

[0061] As an example, obtaining the dielectric layer HZO includes:

[0062] The precursors of the HZO as the dielectric layer are TEMAH and TEMAZ, in the process of obtaining the HZO as the dielectric layer, the HfO2 and ZrO2 cycles are controlled, the concentration of Zr in the HZO thin film is the same as that of Hf, after sputtering the TiN of the top electrode (TE) on the HZO as the dielectric layer, annealing at a temperature of 550 DEG C for 30 seconds in a nitrogen environment, finally, the HZO as the dielectric layer is obtained.

[0063] In some possible embodiments, the output module can be used as an output layer of the reserve pool network, and the output module trains the output weight of the reserve pool network by using a linear regression algorithm.

[0064] The polarization strength of the reservoir network, i.e. the polarization strength of the memcapacitor, is configured as a reservoir state matrix. In the training phase of the reservoir network, the pseudo-inverse of the target matrix and the reservoir state matrix are multiplied to obtain the output weight of the reservoir network, which remains fixed after the training of the reservoir network is completed. In the test phase of the reservoir network, after the new input signal is fed into the reservoir network, the time division multiplexing process and the nonlinear mapping process are performed to obtain the reservoir state matrix, and the reservoir state matrix is multiplied with the output weight matrix to output the prediction result at each time.

[0065] Figure 5 is a result diagram of the MNIST handwritten digit set recognition of a memcapacitor-based reservoir computing system according to an exemplary embodiment.

[0066] As shown in Figure 5 , in some possible embodiments, the feasibility of the memcapacitor-based reservoir computing system of the present disclosure can be verified. Through experimental verification, the accuracy rate of the prediction result of the present disclosure in the MNIST handwritten digit data set can reach 90.3%, and the prediction error (normalized root mean square error) in the Mackey-Glass time series prediction task can reach 0.13. The single pulse energy consumption of the memcapacitor-based reservoir computing system of the present disclosure is 1.8 pJ, while the single pulse energy consumption of the memristor-based reservoir network is 3.0 nJ, indicating that the memcapacitor-based reservoir computing system of the present disclosure can greatly reduce the electrostatic loss in principle.

[0067] The specific embodiments of the present disclosure are described above. It needs to be understood that the present disclosure is not limited to the above specific embodiments, and various modifications or changes can be made by those skilled in the art within the scope of the claims, which does not affect the essential content of the present disclosure. The above preferred features can be used in combination in the case of not conflicting with each other.

Claims

1. A reservoir computing system based on memcapacitors, characterized in that, The application relates to a neural network model, comprising: an input mask module comprising n parallel mask units for modulating an input signal to determine a voltage signal with an amplitude proportional to the input signal; a reservoir module comprising n parallel reservoir units, each of which comprises a memistor, and a perisach model based on the memistor is constructed according to parameters of the memistor, an input end of each perisach model based on the memistor is connected to an output end of a mask unit, and a polarization intensity of the memistor is output at an output end of each perisach model based on the memistor, and the perisach model based on the memistor serves as a node of a reservoir network; wherein the memistor in each perisach model based on the memistor dynamically responds to the input voltage signal, establishes a feedback connection between signals at different time steps, and changes the voltage signal to a high-dimensional space by using a nonlinear change processing mode to determine the polarization intensity of the memistor; an output module for multiplying the polarization intensity of the memistor and output weights of the reservoir network in a training stage of the reservoir network and outputting a prediction result in a test stage of the reservoir network.

2. The reservoir computing system based on memcapacitors of claim 1, wherein, The mask units generate mask signals in a random manner, and the mask signals are discrete matrices with a maximum element of 1.

3. The reservoir computing system based on memcapacitors of claim 1, wherein, A semiconductor parameter analyzer is used to obtain parameters of each memistor.

4. The reservoir computing system based on memcapacitors of claim 1, wherein, The memistor in the reservoir module is a ferroelectric capacitor based on Zr-doped HfO2 (HZO).

5. The reservoir computing system based on memcapacitors of claim 4, wherein, The ferroelectric capacitor based on Zr-doped HfO2 (HZO) comprises TiN as a top electrode, HZO as a dielectric layer and TiN as a bottom electrode.

6. The reservoir computing system based on memcapacitors of claim 5, wherein, The HZO as the dielectric layer is obtained by an atomic layer deposition method at a temperature of 250 DEG C and a dose time of 0.1 seconds.

7. The reservoir computing system based on memcapacitors of claim 6, wherein, TEMAH and TEMAZ are precursors of the HZO as the dielectric layer, HfO2 and ZrO2 are circulated during the process of obtaining the HZO as the dielectric layer, the concentration of Zr in the HZO thin film is the same as that of Hf, after sputtering the TiN as the top electrode on the HZO as the dielectric layer, the HZO as the dielectric layer is annealed in a nitrogen environment at a temperature of 550 DEG C for 30 seconds.

8. The reservoir computing system based on memcapacitors of claim 1, wherein, The output module multiplies a pseudo-inverse of a target matrix and a reservoir state matrix to obtain the output weights of the reservoir network in the training stage of the reservoir network.

9. The reservoir computing system based on memcapacitors of claim 8, wherein, The output module trains the output weights of the reservoir network by using a linear regression algorithm.

Images