Reservoir computing systems, methods, and media based on spin torque nanooscillators

Abstract

The invention relates to a reserve pool calculation system and method based on a spin torque nanometer oscillator and a medium, belongs to the field of neuromorphic calculation and hardware acceleration, and adopts the inherent non-linear current-voltage characteristic of the spin torque nanometer oscillator to directly generate a non-linear signal from a linear signal. The generation of a nonlinear part of a feature vector in next-generation reserve pool calculation is realized, and the three problems of feature dimension combination explosion in NG-RC, high calculation overhead caused by a large amount of matrix operation and insufficient multi-scale dynamics capture capability are effectively solved by replacing a traditional polynomial algorithm with a physical law. Self-adaptive frequency domain response is achieved on the hardware level, complex reconfiguration is not needed, and the time sequence processing efficiency can be improved on the premise that the hardware cost is not remarkably increased in combination with low calculation overhead of simulation domain processing.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of neuromorphic computing and hardware acceleration, and relates to a reservoir computing system based on a spin-torque nano-oscillator, a method and a medium. BACKGROUND

[0002] The spin-torque nano-oscillator (STNO) based on a magnetic tunnel junction (MTJ) is a magnetic device composed of an ultrathin barrier layer and two ferromagnetic (FM) layers, one of which has a fixed magnetic orientation (fixed layer / reference layer), and the magnetic orientation of the other FM layer (free layer) can be changed. The core working mechanism of the spin-torque nano-oscillator is based on the precise coupling of spin transfer torque (STT) and magnetic dynamics. When the injected current intensity exceeds the critical threshold, the spin transfer torque effect precisely offsets the magnetic damping force of the free layer, forcing the magnetization vector to deviate from the static equilibrium and enter a stable continuous precession state. This continuous magnetization oscillation is converted into a measurable voltage oscillation signal in real time through the magnetic resistance effect, and the amplitude directly maps the strength of the magnetization precession. Moreover, the voltage amplitude as a function of the injected current is highly nonlinear and essentially depends on the past input. The amplitude dynamics of the spin-torque nano-oscillator thus combines the two core features of neuromorphic computing, nonlinearity and memory, in a single nanodevice.

[0003] Reservoir computing (RC) is a computing framework derived from recurrent neural networks (RNNs) and is specifically designed for processing time series data. It is widely used in natural language processing and time series analysis tasks that require context association. In the reservoir computing framework, low-dimensional input signals are first mapped to a high-dimensional dynamic system called a reservoir, which is usually composed of an RNN with a fixed random input matrix and an internal connection matrix. After nonlinear transformation by the reservoir, the input data is converted into high-dimensional spatiotemporal dynamic features, which are then decoded by a lightweight trainable output layer. The important meta-parameters in the core structure of traditional reservoir computing are randomly generated, lacking explicit design rules, which leads to uncertainty in the results.

[0004] The research of next generation reservoir computing (NG-RC) theoretically proves that the reservoir computing with a linear reservoir node and a nonlinear output layer is mathematically equivalent to a nonlinear vector autoregressive (NVAR) method. This discovery means that the powerful function of reservoir computing can be achieved without actually constructing and running a complex stochastic RNN (i.e., a "reservoir"). The NVAR method directly uses the time delay and nonlinear combination of the time series data itself to construct a feature vector, thereby implicitly defining an equivalent reservoir computing, and the parameter amount of the output weight matrix to be trained is much smaller than that of the traditional reservoir computing. However, there is still a technical problem of low time series signal processing performance in the existing design of NG-RC. SUMMARY

[0005] In view of the problems in the above-mentioned traditional method, the application provides a spin-torque nano-oscillator-based reservoir computing method, a spin-torque nano-oscillator-based reservoir computing system and a computer readable storage medium, which can realize efficient and adaptive time series signal processing.

[0006] To achieve the above-mentioned purpose, the embodiments of the application adopt the following technical solutions: In one aspect, a spin-torque nano-oscillator-based reservoir computing system is provided, the spin-torque nano-oscillator serving as a nonlinear node of next generation reservoir computing for generating a nonlinear part of a feature vector in the next generation reservoir computing; the reservoir computing system comprises: a data unfolding module configured to unfold external data according to the dimension and time sequence relationship of the external data to be processed to form a linear part of a one-dimensional feature vector; an injection oscillation module configured to inject the linear part into the spin-torque nano-oscillator in a linear order after converting the linear part into an input current, and generate a corresponding oscillation voltage through the spin transfer torque effect of the spin-torque nano-oscillator; a nonlinear module configured to obtain an amplitude of the oscillation voltage to form a nonlinear part with the same dimension as the linear part; a splicing output module configured to splice the linear part and the nonlinear part to obtain a complete feature vector, and multiply the complete feature vector by an output weight of the spin-torque nano-oscillator-based reservoir computing system to obtain a target value corresponding to the external data; wherein the output weight is obtained by linear regression training of the spin-torque nano-oscillator-based reservoir computing system using prior external data.

[0007] In one embodiment, the spin-torque nano-oscillator-based reservoir computing system further comprises: a current scaling module for scaling the input current converted from the linear part to a current range in which the spin-torque nanoscale oscillator can generate a stable oscillation voltage.

[0008] In one of the embodiments, in the process of linear regression training of the spin-torque nanoscale oscillator-based reservoir computing system based on prior external data, the sampling time step of the amplitude of the output oscillation voltage is consistent with the time interval of the current injection.

[0009] In another aspect, a spin-torque nanoscale oscillator-based reservoir computing method is also provided, comprising the steps of: unfolding the external data into a linear part of a one-dimensional feature vector according to the dimension and time sequence relationship of the external data to be processed; injecting the linear part into the spin-torque nanoscale oscillator in a linear order after converting the linear part into an input current, and generating a corresponding oscillation voltage through the spin transfer torque effect of the spin-torque nanoscale oscillator; obtaining an amplitude of the oscillation voltage to form a nonlinear part of the same dimension as the linear part; splicing the linear part and the nonlinear part to obtain a complete feature vector, multiplying the complete feature vector by the output weight of the spin-torque nanoscale oscillator-based reservoir computing system to obtain a target value corresponding to the external data; wherein the output weight is obtained by linear regression training of the spin-torque nanoscale oscillator-based reservoir computing system based on prior external data.

[0010] In one of the embodiments, before injecting the linear part into the spin-torque nanoscale oscillator in a linear order after converting the linear part into an input current, the above-mentioned reservoir computing method further comprises the step of: scaling the input current converted from the linear part to a current range in which the spin-torque nanoscale oscillator can generate a stable oscillation voltage.

[0011] In another aspect, a computer readable storage medium having a computer program stored thereon is also provided, and the computer program, when executed by a processor, implements the steps of the above-mentioned spin-torque nanoscale oscillator-based reservoir computing method.

[0012] One of the above technical solutions has the following advantages and beneficial effects: The spin-torque nanowire oscillator-based reservoir computing system, method and medium described above directly generates a nonlinear signal from a linear signal by using the inherent nonlinear current-voltage characteristics of a spin-torque nanowire oscillator (STNO), and realizes the nonlinear part generation of a feature vector in a next-generation reservoir computing (NG-RC). By replacing the traditional polynomial algorithm with a physical law, the scheme effectively solves the three major problems of feature dimension combination explosion, high computational overhead caused by a large number of matrix operations, and insufficient multi-scale dynamics capture capability in NG-RC. Compared with the prior art, the scheme realizes adaptive frequency domain response (strong current triggers high-frequency oscillation to capture fast-changing processes, and weak current maintains low-frequency oscillation to focus on slow-changing dynamics) at the hardware level without the need for complex reconfiguration. Combined with the low computational overhead of analog domain processing (multiplexing linear dimensions to avoid matrix operations), the scheme can improve the timing processing efficiency without significantly increasing the hardware cost, providing a new way to build a high-performance, high-energy-efficiency neuromorphic computing accelerator, and is suitable for applications such as speech recognition and time series analysis. BRIEF DESCRIPTION OF DRAWINGS

[0013] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0014] Figure 1 A module architecture schematic diagram of a spin-torque nanowire oscillator-based reservoir computing system in an embodiment; Figure 2 A schematic diagram of a spin-torque nanowire oscillator-based reservoir computing scheme composed of linear reservoir nodes and a nonlinear output layer in an embodiment; Figure 3 A schematic diagram of an experimental device for processing data by a spin-torque nanowire oscillator in an embodiment; Figure 4 An offline training process schematic diagram of a spin-torque nanowire oscillator-based reservoir computing experimental device composed of linear reservoir nodes and a nonlinear output layer in an embodiment; Figure 5 A flowchart of a spin-torque nanowire oscillator-based reservoir computing method in an embodiment. DETAILED DESCRIPTION

[0015] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not intended to limit the present application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terms used in the specification of the present application are only for the purpose of describing the specific embodiments and are not intended to limit the present application.

[0016] It should be noted that the reference to "embodiments" herein means that the specific features, structures or characteristics described in connection with the embodiments can be included in at least one embodiment of the present application. The phrase is shown at various places in the specification does not necessarily refer to the same embodiment, nor is it mutually exclusive or alternative to other embodiments. Those skilled in the art can understand that the embodiments described herein can be combined with other embodiments. The term "and / or" used herein refers to any combination of one or more of the associated listed items and all possible combinations, and includes these combinations.

[0017] The embodiments of the present application will be described in detail below with reference to the accompanying drawings of the embodiments of the present application.

[0018] The linear part (Olin) of the feature vector in the NG-RC is composed of the observation data at the current time and the historical time to form a time delay embedded sequence, and the nonlinear part (Ononlin) of the feature vector is a nonlinear function of Olin. In existing research, the linear feature vector Olin is combined with polynomials to generate high-order nonlinear terms. Therefore, the dimension of the nonlinear feature vector (i.e. the dimension Dnonlin of the nonlinear part) is determined by the polynomial order, the time series dimension of the original input data and the time delay step number. When any one parameter increases, the feature dimension will increase geometrically, which will cause the problem of feature dimension explosion. Moreover, the polynomial needs to perform a large number of matrix operations, which has large calculation overhead and high energy consumption. Therefore, the advantage of the polynomial NG-RC is significant in low-dimensional systems, but the calculation complexity increases nonlinearly with the dimension. Wherein, Dlin represents the dimension of the linear part.

[0019] However, complex systems require the simultaneous capture of dynamics at different time scales, but the polynomial order in polynomial methods must be pre-set manually and cannot be dynamically adjusted. For example, the Lorenz63 system (a simplified system describing atmospheric convection, exhibiting chaotic characteristics, and a classic model in nonlinear dynamical system research) uses a fixed-order second-order polynomial, while the double-vortex system, due to the presence of non-polynomial vector fields, raises it to third order. Fast-changing processes require high-frequency components, but the cutoff frequency of low-order polynomials is insufficient, leading to aliasing errors; slow-changing processes need to retain DC / low-frequency components, but higher-order terms are overly sensitive to low-frequency signals, amplifying the basis noise. While higher-order terms can extend the bandwidth, they can cause dimensionality explosion. Complex systems need to analyze both fast and slow-changing processes simultaneously, which static polynomials cannot adequately address.

[0020] In one embodiment, such as Figure 1 As shown, a reservoir computing system 100 based on a spin-torque nano-oscillator is provided. The spin-torque nano-oscillator serves as a nonlinear node in next-generation reservoir computing, used to generate the nonlinear part of the eigenvector in the next-generation reservoir computing. Figure 1 As shown, the reservoir computing system 100 may include a data expansion module 11, an injection oscillation module 13, a nonlinear module 15, and a splicing output module 17. The data expansion module 11 expands the external data to form a linear part of a one-dimensional feature vector based on the dimension and temporal relationship of the external data to be processed. The injection oscillation module 13 converts the linear part into an input current and injects it into a spin torque nano-oscillator in a linear sequence, generating a corresponding oscillation voltage through the spin-transfer torque of the spin torque nano-oscillator. The nonlinear module 15 obtains the amplitude of the oscillation voltage to form a nonlinear part with the same dimension as the linear part. The splicing output module 17 splices the linear and nonlinear parts to obtain a complete feature vector, and multiplies the complete feature vector by the output weights of the reservoir computing system based on the spin torque nano-oscillator to obtain the target value corresponding to the external data. The output weights are obtained by linear regression training of the reservoir computing system based on the spin torque nano-oscillator using prior external data.

[0021] It is understood that in this embodiment, the inherent nonlinear current-voltage characteristics of the spin torque nano-oscillator (STNO) are used to directly generate a nonlinear signal from a linear signal, thereby realizing the function of generating the nonlinear part of the eigenvector through polynomials in next-generation reservoir calculations. This design in this embodiment belongs to analog domain current-voltage conversion, using physical laws to replace algorithm design, thus resulting in extremely low computational overhead. In this embodiment, the spin torque nano-oscillator (STNO) in... Figure 1As a nonlinear node in the reserve pool calculation, the nonlinear dynamics characteristics (nonlinear conversion of current-voltage) of its physical layer are used to introduce rich nonlinear expression capability for the reserve pool, thereby enhancing the system's performance in processing complex time series (such as chaotic signals and communication signals). This hardware-level nonlinearity has the advantages of low power consumption, high speed, and strong physical robustness compared with traditional software-simulated nonlinearity.

[0022] It should be noted that, Figure 2 a schematic diagram of a reserve pool calculation scheme based on a spin-torque nano-oscillator composed of a linear reserve pool node and a nonlinear output layer, Figure 2 The leftmost part represents a three-dimensional time series signal x , y , z is sampled at time point , as the input sequence of the system. Figure 2 The middle part is the internal process of the module (NG-RC with STNO): first, the linear input generation , that is, first construct a linear input matrix containing the signal values at the current time and the previous time , which is the basic input for linear operation; then perform nonlinear transformation of the STNO, the spin-torque nano-oscillator STNO works as a nonlinear core element, first input current I to the spin-torque nano-oscillator STNO, using its spintronic characteristics (such as spin transfer torque effect), the spin-torque nano-oscillator STNO converts the current into voltage, completes the nonlinear mapping, generates a nonlinear output matrix , corresponding to the transformed ; finally, the total output and the next time prediction, combining the linear output and the nonlinear output (i.e., Ototal,i), and then performing operation through the output weight Wout, finally generating the output at the next time , completing the processing of the time series (such as prediction and feature extraction, etc.). Figure 2 The rightmost part is the display of the system output, after the system processing, the output sequence of the time series signal at the time , which embodies the function of the next step prediction or transformed output of the system to the time series; is the magnetic moment of the free layer of the spin-torque nano-oscillator STNO, is the magnetic moment of the fixed layer of the spin-torque nano-oscillator STNO.

[0023] Firstly, the external data is firstly unfolded according to its own dimension and time sequence relationship to form a linear part of a one-dimensional feature vector; then, the linear part is converted into a current and then injected into a spin-torque nano-oscillator STNO according to the linear order, and the corresponding oscillation voltage is generated through the action of the spin transfer torque STT; then, the amplitude of the oscillation voltage constitutes a nonlinear part, and the nonlinear part inherits the dimension of the linear part, and the dimensions of the two are the same (i.e. Dlin=Dnonlin); finally, the linear part and the nonlinear part are spliced to obtain a complete feature vector, and the dimension of the complete feature vector is the sum of the dimensions of the linear part and the nonlinear part (i.e. Dtotal=Dlin+Dnonlin), and the output weight matrix is obtained through simple linear regression training , the feature vector is multiplied by the output weight to obtain the target value (i.e. the output sequence or signal corresponding to the external data).

[0024] The nonlinear part of the feature vector directly inherits the dimension of the linear part, and the complexity of the overall feature space is significantly reduced through the architecture of multiplexing the linear dimension, and the current time state in the time delay sequence dynamically integrates the residual effect of the historical time data, forming an adaptive time sequence correlation mechanism.

[0025] The voltage oscillation amplitude obtained in the foregoing has robustness to noise, due to its double torque dynamic balance mechanism: the driving effect of the injected current generates a positive torque that promotes precession, and the dissipation effect of the magnetic damping forms an opposite torque that hinders movement, when the intensity of the injected current exceeds the critical threshold, the two torques dynamically offset, forming a stable state similar to an energy trap, and this antagonistic torque system automatically absorbs environmental noise disturbances (such as thermal fluctuations or electromagnetic interference), so that the output oscillation is locked on a stable trajectory.

[0026] In dynamic system modeling, noise immunity enhancement refers to improving the robustness of the model to input noise through a specific mechanism to ensure that the prediction result is not affected by weak disturbances. Traditional methods rely on Tikhonov regularization (also known as Ridge Regression) to suppress noise, but the spin-torque nano-oscillator STNO realizes automatic regularization through hardware physical characteristics, and weak noise (such as sensor noise or environmental disturbances) usually appears as a low-amplitude current, which cannot trigger the spin-torque nano-oscillator STNO oscillation, so these noise signals are naturally suppressed, and only the effective signal enters the nonlinear feature generation, i.e. low-amplitude noise is directly filtered out in the feature generation stage, which is equivalent to pre-filtering, which improves the noise robustness, simplifies the noise processing process, and makes the optimization of the regularization strength α more stable (i.e. the sensitivity of α to noise is reduced), and the spin-torque nano-oscillator STNO reduces the difficulty of α tuning through hardware mechanism.

[0027] Spin-torque nano-oscillator (STNO) has a unique current-frequency coupling mechanism, which can realize the intelligent decoupling of multi-scale dynamics at the hardware level. The core of the mechanism is that the strength of the input current is directly converted into a real-time control signal of the oscillation frequency. When a strong current is injected, the STNO is excited to produce high-frequency oscillation, and its response bandwidth is automatically expanded to accurately capture fast-changing processes in the system. Conversely, when the input current is weak, the STNO maintains low-frequency oscillation, and its response naturally focuses on the slow-changing dynamics of the system. This dynamic and continuous frequency-domain adaptive capability enables a single STNO to seamlessly cover multiple orders of magnitude of frequency range without any external reconfiguration or parameter adjustment. The essence of this mechanism is to directly map the mathematical relationship by simulating the physical process, thereby natively solving the modeling problem of multi-scale coupling at the hardware level. It not only decouples fast-changing processes from slow-changing processes in real time, but also automatically balances the transient dynamics and steady-state behavior of the system through its inherent physical response characteristics, providing a flexible and efficient solution for accurate modeling of complex systems.

[0028] The spin-torque nano-oscillator-based reservoir computing system 100 directly generates nonlinear signals from linear signals using the inherent nonlinear current-voltage characteristics of the STNO, realizes the generation of nonlinear parts of feature vectors in next-generation reservoir computing, and effectively solves the three major problems of feature dimension combination explosion, high computational overhead caused by a large number of matrix operations, and insufficient multi-scale dynamics capture capability in next-generation reservoir computing. Compared with the prior art, the present scheme realizes adaptive frequency-domain response (high-frequency oscillation triggered by strong current to capture fast-changing processes, and low-frequency oscillation maintained by weak current to focus on slow-changing dynamics) at the hardware level without complex reconfiguration. Combined with the low computational overhead of analog domain processing (multiplexing linear dimensions to avoid matrix operations), it can improve the timing processing efficiency without significantly increasing the hardware cost, providing a new way to build high-performance and energy-efficient neuromorphic computing accelerators, which is suitable for applications such as speech recognition and time series analysis.

[0029] Figure 3 The settings for the experiment of the spin-torque nano-oscillator (STNO) are introduced. When processing external data using the STNO, a bias current needs to be input first I dc The free layer magnetization of the STNO itself will be affected by magnetic damping, which is like friction resistance, causing any swing or precession to quickly decay and stop. Only when the bias current exceeds a certain threshold current I thAt this time, the energy provided by the spin transfer torque (STT) can completely offset the energy consumed by the damping, and at this time the magnetization can start and maintain stable periodic precession (i.e., oscillation). The oscillation amplitude is strongly dependent on the bias current, and a suitable current value can be selected to set the oscillator at an optimal operating point with strong nonlinear response and high signal-to-noise ratio. This nonlinearity is the key to simulating neuron behavior and implementing complex calculations. Without a nonlinear oscillator, the input signal (i.e., external data) can only be linearly amplified, and the required complex transformation cannot be completed.

[0030] After the external data is input, it is converted into a physical voltage waveform (Vin) by an arbitrary waveform generator (AWG), and then the voltage is converted into an alternating current Iac by a voltage-controlled current source VCCS. The direct current Idc and the alternating current Iac pass through the inductance L and the capacitance C in the bias-tee (also called T-type bias-tee) respectively, to achieve the functions of blocking alternating current and passing direct current, and blocking direct current and passing alternating current. Finally, they are superimposed together and injected into the magnetic tunnel junction MTJ. The amplitude of the oscillation voltage is measured using a microwave diode D0, and the output data is finally sampled by an ADC sampling circuit for subsequent processing.

[0031] As shown in Figure 4 , the output weight is obtained through an independent offline training process. The offline training process first preprocesses the external data as prior data to extract more effective features from the original data for the current task, and then uses the preprocessed external data for subsequent operations to ensure the final effect. Among them, the preprocessing methods of different task data can be different, for example, the silent segment of the voice data is removed, the Fourier transform is performed after framing and windowing to realize time-frequency conversion, MFCC (Mel frequency cepstrum coefficient) feature extraction, etc.; the prediction task needs to create lag features, that is, the data of the previous period (such as the previous day, the previous few hours) is used as new features. Taking the weather prediction task as an example, the temperature data (i.e., external data) of the previous 3 days is usually used as the feature of the predicted temperature Ti of the current day. These temperature data are uniformly unfolded according to their dimensions and time sequence relationship, and arranged into a one-dimensional linear vector , wherein represents the temperature data of the third day before the current day, represents the temperature data of the second day before the current day, represents the temperature data of the first day before the current day.

[0032] In an embodiment, the spin-torque nano-oscillator-based reservoir computing system 100 described above can further include a current scaling module for scaling the input current converted by the linear part to a current range in which the spin-torque nano-oscillator can generate stable oscillation current.

[0033] It can be understood that different spin-torque nano-oscillators STNOs have different ranges of stable oscillation current, and therefore the external data needs to be scaled to the corresponding current range according to the specific STNO requirement, and the scaled current is in a time sequence order is injected into the spin-torque nano-oscillator STNO, wherein, represents the corresponding input current. represents the corresponding input current, represents the corresponding input current. The time interval of data point injection also needs to be adjusted according to the device characteristics of the spin-torque nano-oscillator STNO: for example, the precession of the magnetization of the free layer of the spin-torque nano-oscillator STNO has an intrinsic relaxation time If the time interval of injection is too short, the magnetization state has not yet stabilized, which will lead to the oscillation amplitude not reaching the steady state, and thus cause distortion of the nonlinear response and a decrease in noise suppression capability; if the time interval of injection is too long, the correlation between the sequence points will decay, resulting in insufficient equivalent memory depth of the reservoir.

[0034] Further, the sampling time step of the amplitude of the output oscillation voltage is consistent with the time interval of current injection. In the spin-torque nano-oscillator-based reservoir computing system, one linear input corresponds to one nonlinear output, i.e., the current , the current and the current correspond to the output voltage amplitudes , the output voltage amplitudes and the output voltage amplitudes respectively. After splicing the input current and the output voltage amplitude, the final feature vector (Ii-3, Ii-2, Ii-1, Vi-3, Vi-2, Vi-1) can be formed. Next, the training data set and the test data set used for training and testing respectively can also be constructed (for example, the data set can be divided in a ratio of 7:3 or 6:4), and different training methods are adopted according to the specific task requirement (such as identification classification, time series prediction or regression analysis): for example, for the weather prediction task, the time series prediction method is used, the temperature of the current day is predicted through the temperature data of the previous three days, and the predicted temperature value of the current day is calculated through the ridge regression (an unbiased estimation regression method used to solve the problem of collinear data) to obtain the output weight Wout. Subsequently, the output weight Wout is iteratively updated using all the prior data in the training data set until the end of the sequence, and finally the prediction accuracy of the spin-torque nano-oscillator-based reservoir computing system is tested using the test data set, i.e., whether the prediction accuracy meets the required design requirement can be known, and the system can be used to process real-time time series analysis tasks if the design requirement is met.

[0035] Each module in the aforementioned reservoir computing system 100 based on spin torque nano-oscillators can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independently of a device with data processing capabilities, or stored in software within the memory of the aforementioned device, so that the processor can call and execute the operations corresponding to each module. The aforementioned device can be, but is not limited to, various types of data processing devices already existing in the art.

[0036] In one embodiment, such as Figure 5 As shown, a method for calculating a reservoir based on a spin torque nano-oscillator is also provided, which may include the following steps S12 to S18: S12, based on the dimension and time sequence relationship of the external data to be processed, expand the external data to form the linear part of a one-dimensional feature vector; S14 converts the linear portion into an input current and injects it into the spin torque nano oscillator in a linear sequence. The corresponding oscillation voltage is generated through the spin-transfer torque effect of the spin torque nano oscillator. S16, obtain the nonlinear part of the oscillation voltage amplitude with the same dimension as the linear part; S18: The linear and nonlinear parts are concatenated to obtain a complete feature vector. The complete feature vector is multiplied by the output weights of the reservoir computing system based on spin torque nano-oscillators to obtain the target value corresponding to the external data. The output weights are obtained by linear regression training using prior external data through the reservoir computing system based on spin torque nano-oscillators.

[0037] The aforementioned reservoir computation method based on spin-torque nano-oscillators (STNOs) directly generates nonlinear signals from linear signals by utilizing the inherent nonlinear current-voltage characteristics of STNOs, thus realizing the generation of the nonlinear part of the eigenvectors in next-generation reservoir computation (NG-RC). By replacing traditional polynomial algorithms with physical laws, this scheme effectively solves three major problems in NG-RC: feature dimension combinatorial explosion, high computational overhead due to numerous matrix operations, and insufficient multi-scale dynamics capture capability. Compared with existing technologies, this scheme achieves adaptive frequency domain response at the hardware level (strong current triggers high-frequency oscillations to capture fast-changing processes, while weak current maintains low-frequency oscillations to focus on slow-changing dynamics), without the need for complex reconfiguration. Combined with the low computational overhead of analog domain processing (reusing linear dimensions and avoiding matrix operations), it can improve time-series processing efficiency without significantly increasing hardware costs, providing a new approach for building high-performance, high-energy-efficiency neuromorphic computing accelerators, suitable for applications such as speech recognition and time series analysis.

[0038] In one embodiment, the above reservoir computing method further comprises the steps of: The input current converted from the linear part is scaled to the current range in which the spin-torque nano-oscillator can generate stable oscillation voltage.

[0039] In one embodiment, the sampling time step of the amplitude of the output oscillation voltage and the time interval of the current injection are consistent during the linear regression training process of the spin-torque nano-oscillator-based reservoir computing system based on the prior external data.

[0040] It can be understood that the explanations of the features of the above-described spin-torque nano-oscillator-based reservoir computing method can be understood in the same way as the corresponding explanations of the embodiments of the above-described spin-torque nano-oscillator-based reservoir computing system 100.

[0041] It should be understood that, although Figure 5 the steps are displayed in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, Figure 5 at least part of the steps of may include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily sequential, but can be alternately executed with other steps or at least part of the sub-steps or stages of other steps.

[0042] In one embodiment, a computer-readable storage medium is provided, which stores a computer program that is executed by a processor to implement the following processing steps: according to the dimension and time sequence relationship of the external data to be processed, the external data is unfolded to form a linear part of a one-dimensional feature vector; the linear part is converted into an input current which is injected into a spin-torque nano-oscillator in a linear order, and a corresponding oscillation voltage is generated by the spin transfer torque effect of the spin-torque nano-oscillator; the amplitude of the oscillation voltage is obtained to form a non-linear part with the same dimension as the linear part; the linear part and the non-linear part are spliced to obtain a complete feature vector, and the complete feature vector is multiplied by the output weight of the spin-torque nano-oscillator-based reservoir computing system to obtain the target value corresponding to the external data; wherein the output weight is obtained by linear regression training of the spin-torque nano-oscillator-based reservoir computing system based on prior external data.

[0043] In one embodiment, the computer program, when executed by the processor, further implements the steps or sub-steps added in each of the above-mentioned spin-torque nano-oscillator-based reservoir computing method embodiments.

[0044] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, the processes of the above-mentioned embodiments can be included. Any reference to memory, storage, database or other medium used in each embodiment of the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus dynamic random access memory (Rambus DRAM, RDRAM for short) and interface dynamic random access memory (DRDRAM).

[0045] Any combination of the technical features of the above embodiments can be made. In order to make the description simple, all possible combinations of the technical features in the above embodiments are not described, however, as long as the combination of the technical features does not exist, it should be considered as the scope of the present application.

[0046] The above embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the protection scope of the present application. It should be pointed out that, for those skilled in the art, without departing from the concept of the present application, some modifications and improvements can be made, which all belong to the protection scope of the present application.

Claims

1. A spin-torque nanowire oscillator based reservoir computing system, comprising: Spin-torque nano-oscillators as non-linear nodes of next-generation reservoir computing for generating non-linear part of feature vectors in next-generation reservoir computing; The reservoir computing system comprises: a data unfolding module configured to unfold external data to be processed according to a dimension and a time sequence relationship of the external data to form a linear part of a one-dimensional feature vector; an injection oscillation module configured to inject the linear part into the spin-torque nano-oscillator in a linear order after the linear part is converted into an input current, and generate a corresponding oscillation voltage through a spin-transfer torque effect of the spin-torque nano-oscillator; a non-linear module configured to obtain a non-linear part with the same dimension as the linear part by taking an amplitude of the oscillation voltage; a splicing output module configured to splice the linear part and the non-linear part to obtain a complete feature vector, and multiply the complete feature vector by an output weight of the reservoir computing system based on the spin-torque nano-oscillator to obtain a target value corresponding to the external data; wherein the output weight is obtained by linear regression training of the reservoir computing system based on the spin-torque nano-oscillator using prior external data.

2. The spin-torque nanowire oscillator-based reservoir computing system of claim 1, wherein, Further comprising: a current scaling module configured to scale the input current converted from the linear part to a current range in which the spin-torque nano-oscillator can generate a stable oscillation voltage.

3. The spin-torque nanowire oscillator-based reservoir computing system of claim 1 or 2, wherein, In the process of linear regression training of the reservoir computing system based on the spin-torque nano-oscillator using prior external data, a sampling time step of an amplitude of the output oscillation voltage is consistent with a time interval of current injection.

4. A spin-torque nano-oscillator based reservoir computing method, characterized in that, The method comprises the steps of: unfolding external data to be processed according to a dimension and a time sequence relationship of the external data to form a linear part of a one-dimensional feature vector; injecting the linear part into the spin-torque nano-oscillator in a linear order after the linear part is converted into an input current, and generating a corresponding oscillation voltage through a spin-transfer torque effect of the spin-torque nano-oscillator; obtaining a non-linear part with the same dimension as the linear part by taking an amplitude of the oscillation voltage; splicing the linear part and the non-linear part to obtain a complete feature vector, and multiplying the complete feature vector by an output weight of the reservoir computing system based on the spin-torque nano-oscillator to obtain a target value corresponding to the external data; wherein the output weight is obtained by linear regression training of the reservoir computing system based on the spin-torque nano-oscillator using prior external data.

5. The spin-torque nanowire oscillator-based reservoir computing method of claim 4, wherein, Before injecting the linear part into the spin-torque nano-oscillator in a linear order after the linear part is converted into an input current, the reservoir computing method further comprises the step of: scaling the input current converted from the linear part to a current range in which the spin-torque nano-oscillator can generate a stable oscillation voltage.

6. The spin-torque nanowire oscillator-based reservoir computing method according to claim 4 or 5, c h a r a c t e r i z e d b y, In the process of linear regression training of the reservoir computing system based on the spin-torque nano-oscillator using prior external data, a sampling time step of an amplitude of the output oscillation voltage is consistent with a time interval of current injection.

7. A computer readable storage medium having stored thereon a computer program, characterized in that The computer program, when executed by a processor, implements the steps of the reservoir computing method based on the spin-torque nano-oscillator according to any one of claims 4 to 6.

Images