Implementation method, device and system of physical intrinsic neural network and medium

Abstract

The invention provides a physical intrinsic neural network implementation method, device and system and a medium, and relates to the field of artificial intelligence. The multi-layer neural network includes a first network layer implemented by a first physical intrinsic system configured to receive a physical input signal of the first network layer and output a physical output signal of the first network layer based on a response characteristic of the first physical intrinsic system. The method comprises the following steps: determining input data of a multi-layer neural network, and determining a to-be-trained target layer in the multi-layer neural network; inputting the input data into a multi-layer neural network, and obtaining an output signal of a target layer; based on the output signal of the target layer, the goodness of the input data is determined, adjustable parameters of the target layer are adjusted based on the goodness, and the goodness describes the response quality of the target layer to the input data.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence, and in particular to a method for implementing a physical intrinsic neural network, an implementation device, a system, a computer-readable storage medium, and a computer program product of the physical intrinsic neural network. Background Art

[0002] An artificial neural network is a computational model that mimics the biological nervous system. It consists of multiple interconnected layers, with nodes in each layer acting like neurons, transmitting signals through synapses. During computation, each neuron performs weighted summation and nonlinear transformations based on input activation values combined with model parameters such as weights and biases, enabling feature extraction and representation. This network can continuously learn through parameter updates during the training phase and generate or discriminate outputs during the inference phase. It is widely used in fields such as image processing, speech, and natural language processing.

[0003] The approaches described in this section are not necessarily approaches that have been previously conceived or employed. Unless otherwise indicated, it should not be assumed that any approach described in this section is prior art simply by virtue of its inclusion in this section. Similarly, unless otherwise indicated, the issues raised in this section should not be considered as having been recognized in any prior art. Summary of the Invention

[0004] It would be advantageous to provide a mechanism that alleviates, mitigates, or even eliminates one or more of the above-mentioned problems.

[0005] According to one aspect of the present disclosure, a method for implementing a physical intrinsic neural network is provided. The multi-layer neural network includes a first network layer implemented by a first physical intrinsic system, and the first physical intrinsic system is configured to receive a physical input signal of the first network layer and output a physical output signal of the first network layer based on the response characteristics of the first physical intrinsic system. The method includes: determining the input data of the multi-layer neural network and determining the target layer to be trained in the multi-layer neural network; inputting the input data into the multi-layer neural network and obtaining the output signal of the target layer; and determining the goodness of the input data based on the output signal of the target layer, and adjusting the adjustable parameters of the target layer based on the goodness, wherein the goodness describes the response quality of the target layer to the input data.

[0006] According to another aspect of the present disclosure, a device for implementing a physical intrinsic neural network is provided. The multi-layer neural network includes a first network layer implemented by a first physical intrinsic system, the first physical intrinsic system is configured to receive a physical input signal of the first network layer, and output a physical output signal of the first network layer based on the response characteristics of the first physical intrinsic system. The device includes: a first determination unit configured to determine input data of the multi-layer neural network and determine a target layer to be trained in the multi-layer neural network; an acquisition unit configured to input the input data into the multi-layer neural network and acquire the output signal of the target layer; and a determination unit configured to determine the goodness of the input data based on the output signal of the target layer, and adjust the adjustable parameters of the target layer based on the goodness, wherein the goodness describes the response quality of the target layer to the input data.

[0007] According to yet another aspect of the present disclosure, a system is provided for implementing a multi-layer neural network. The system includes: a first physical eigensystem, configured to: implement a first network layer in a multi-layer neural network; and receive a physical input signal of the first network layer, and output a physical output signal of the first network layer based on the response characteristics of the first physical eigensystem; and at least one of a first control unit and a second control unit. The first control unit includes: at least one processor; and a memory connected to the at least one processor in communication, wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor and the first physical eigensystem can collaboratively perform at least one step of the above method. The second control unit includes: a third eigenmapping channel of the physical eigen, wherein the third eigenmapping channel and the first physical eigensystem can collaboratively perform at least one step of the above method in the physical domain.

[0008] According to yet another aspect of the present disclosure, a computer-readable storage medium is provided, on which a computer program is stored. The computer program is used to implement a multi-layer neural network, wherein the multi-layer neural network includes a first network layer implemented by a first physical eigensystem, the first physical eigensystem being configured to receive a physical input signal of the first network layer and output a physical output signal of the first network layer based on a response characteristic of the first physical eigensystem, wherein when the computer program is executed by a processor, the processor and the first physical eigensystem are enabled to collaboratively perform the above method.

[0009] According to yet another aspect of the present disclosure, a computer program product is provided, comprising a computer program for implementing a multi-layer neural network, the multi-layer neural network comprising a first network layer implemented by a first physical eigensystem, the first physical eigensystem being configured to receive a physical input signal of the first network layer and output a physical output signal of the first network layer based on a response characteristic of the first physical eigensystem, wherein when the computer program is executed by a processor, the processor and the first physical eigensystem are enabled to collaboratively perform the above-mentioned method.

[0010] According to one or more embodiments of the present disclosure, compared to the traditional method of using only silicon-based integrated circuits and updating the parameters of the entire network through the backpropagation algorithm, the first network layer in the neural network is implemented by utilizing the response characteristics of the physical intrinsic system, and combined with the parameter adjustment method based on forward propagation and goodness, not only the complex numerical calculations of the silicon-based integrated circuit are reduced, but also the gradient is not acquired and transmitted using the backpropagation algorithm, thereby effectively solving the problem that the physical intrinsic system is difficult to transmit the gradient in the reverse direction, which makes the neural network training difficult. In addition, the parameter adjustment method based on forward propagation and goodness enables the training of each network layer to be carried out independently, and the parallel signal processing capability of the physical intrinsic system can be matched with the above parameter adjustment method, thereby efficiently realizing asynchronous reasoning and training of different layers. Because the physical intrinsic system can spontaneously complete the forward propagation process of the first network layer, under certain circumstances, it is more efficient, high-speed and low-energy than the numerical calculation of silicon-based integrated circuits.

[0011] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present disclosure, nor are they intended to limit the scope of the present disclosure. Other features of the present disclosure will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Further details, features and advantages of the present disclosure are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which:

[0013] Figure 1 A flowchart illustrating a method for implementing a physical intrinsic neural network according to an exemplary embodiment of the present disclosure is shown;

[0014] Figure 2 A flowchart of determining the goodness of input data based on an output signal of a target layer and adjusting an adjustable parameter of the target layer based on the goodness is shown according to an exemplary embodiment of the present disclosure;

[0015] Figure 3 A schematic diagram illustrating calculation of goodness and an adjustment amount of an adjustable mechanism for a physical intrinsic system using a silicon-based integrated circuit according to an exemplary embodiment of the present disclosure is shown;

[0016] Figure 4 A flowchart of determining the goodness of input data based on an output signal of a target layer and adjusting an adjustable parameter of the target layer based on the goodness is shown according to an exemplary embodiment of the present disclosure;

[0017] Figure 5 A schematic diagram illustrating determining goodness and adjusting response characteristics of a physical eigensystem in a purely physical domain according to an exemplary embodiment of the present disclosure is shown;

[0018] Figure 6 A flowchart illustrating a method for implementing a physical intrinsic neural network according to an exemplary embodiment of the present disclosure is shown;

[0019] Figure 7 A schematic diagram of a forward-backward propagation algorithm according to an exemplary embodiment of the present disclosure is shown;

[0020] Figure 8 A schematic diagram illustrating forward propagation and parameter adjustment of multiple network layers according to an exemplary embodiment of the present disclosure is shown;

[0021] Figure 9 A structural block diagram of a device for implementing a physical intrinsic neural network according to an exemplary embodiment of the present disclosure is shown;

[0022] Figure 10 A structural block diagram of a system according to an exemplary embodiment of the present disclosure is shown; and

[0023] Figure 11 A block diagram of an exemplary computer device that can be used in accordance with exemplary embodiments is shown. DETAILED DESCRIPTION

[0024] In this disclosure, unless otherwise specified, the use of terms such as "first" and "second" to describe various elements is not intended to limit the positional relationship, temporal relationship, or importance relationship of these elements. Such terms are only used to distinguish one element from another. In some examples, the first element and the second element may refer to the same instance of the element, while in some cases, based on the context of the description, they may also refer to different instances.

[0025] The terms used in the description of various examples in this disclosure are for the purpose of describing specific examples only and are not intended to be limiting. Unless the context clearly indicates otherwise, if the number of elements is not specifically limited, the element can be one or more. As used herein, the term "plurality" means two or more, and the term "based on" should be interpreted as "based at least in part on". In addition, the terms "and / or" and "at least one of..." cover any one of the listed items and all possible combinations.

[0026] Some concepts and terms involved in this disclosure are introduced below.

[0027] In related technologies, traditional silicon-based integrated circuits usually rely on complex digital calculations and large-scale data storage to implement neural networks, which not only makes the calculation process cumbersome and time-consuming, but also uses transistors and other components to achieve high energy consumption for digital operations and data storage.

[0028] Compared to the traditional method of using only silicon-based integrated circuits and updating the parameters of the entire network through the backpropagation algorithm, the present disclosure implements the first network layer in the neural network by utilizing the response characteristics of the physical intrinsic system, and combines it with a parameter adjustment method based on forward propagation and goodness. This not only reduces the complex numerical calculations of silicon-based integrated circuits, but also eliminates the need to use the backpropagation algorithm to obtain and transfer gradients, thereby effectively solving the problem that the physical intrinsic system is difficult to reversely transfer gradients, which makes neural network training difficult. In addition, the parameter adjustment method based on forward propagation and goodness enables the training of each network layer to be carried out independently, and the parallel signal processing capability of the physical intrinsic system can be coordinated with the above parameter adjustment method, thereby efficiently realizing asynchronous reasoning and training of different layers. Because the physical intrinsic system can spontaneously complete the forward propagation process of the first network layer, it is more efficient, high-speed and low-energy than the numerical calculation of silicon-based integrated circuits under certain circumstances. The exemplary embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0029] According to one aspect of the present disclosure, a method for implementing a physical eigenvalue neural network is provided. The multi-layer neural network includes a first network layer implemented by a first physical eigenvalue system. The first physical eigenvalue system is configured to receive a physical input signal of the first network layer and output a physical output signal of the first network layer based on a response characteristic of the first physical eigenvalue system.

[0030] Figure 1 FIG. 1 shows a flow chart of a method for implementing a physical intrinsic neural network according to an exemplary embodiment of the present disclosure. Figure 1 As shown, method 100 includes: step S101, determining the input data of the multi-layer neural network, and determining the target layer to be trained in the multi-layer neural network; step S102, inputting the input data into the multi-layer neural network, and obtaining the output signal of the target layer; and step S103, determining the goodness of the input data based on the output signal of the target layer, and adjusting the adjustable parameters of the target layer based on the goodness, wherein the goodness describes the response quality of the target layer to the input data.

[0031] Therefore, compared with the traditional artificial neural network implemented only using silicon-based integrated circuits and the method of updating the parameters of the entire network through the backpropagation algorithm, the present disclosure implements the first network layer in the neural network by utilizing the response characteristics of the physical intrinsic system, and combines it with the parameter adjustment method based on forward propagation and goodness. It not only reduces the complex numerical calculations of silicon-based integrated circuits, but also eliminates the need to use the backpropagation algorithm to obtain and transfer gradients, thereby effectively solving the problem that the physical intrinsic system is difficult to reversely transfer gradients, which leads to difficulties in neural network training. In addition, the parameter adjustment method based on forward propagation and goodness enables the training of each network layer to be carried out independently, and the parallel signal processing capability of the physical intrinsic system can be coordinated with the above-mentioned parameter adjustment method, thereby efficiently realizing asynchronous reasoning and training of different layers. Because the physical intrinsic system can spontaneously complete the forward propagation process of the first network layer, under certain circumstances, it is more efficient, high-speed and low-energy than the numerical calculation of silicon-based integrated circuits.

[0032] Exemplary embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings.

[0033] The neural network implemented in this disclosure does not rely on traditional backpropagation algorithms. Unlike traditional neural networks that update weights through backpropagation, the neural network implemented in this disclosure can update the parameters of each network layer layer by layer in a forward direction without backpropagation. Parameter updates can be achieved through simulated annealing, particle swarm optimization, greedy training, evolutionary algorithms, and other methods.

[0034] A multi-layer neural network can have one or more modes of training or inference (e.g., generation, discrimination, or solution), and has a structure with two or more connected layers. As defined in the field of neuroscience, each layer can have "synapses," "neurons," and / or other structures; as defined in the field of computational science, each layer can have "activation values," "weights / biases (model parameters)," and / or other structures.

[0035] In inference mode, given fixed or time-determined model parameters, each layer's output can have a deterministic or probabilistic response to each layer's input. The network's output can have a deterministic or probabilistic response to the network's input. This response can be complex and high-dimensional. Data flow in inference mode should be forward propagated, with no backpropagation. The activation values of each layer are forward propagated through the parameters to produce the output layer by layer. Each layer can also have cross-layer connections with different structures.

[0036] In training mode, unlike traditional back-propagation gradient descent chain derivation, the multi-layer neural network implemented in the present disclosure can update parameters layer by layer. Each layer can have a corresponding target (which can include a target for physical signals and / or an objective function for digital signals). Each training data flow will obtain the current result of the layer once, and this result will affect the parameter update, and the parameter update will be updated in the direction of a better target, or probabilistically updated in the direction of a better target. The training mode will be explained in more detail below in conjunction with steps S101-S103.

[0037] Various algorithms can be run on the neural network implemented by the above method.

[0038] According to some embodiments, a multi-layer neural network can run at least one algorithm selected from the group consisting of: deterministic algorithms, stochastic or probabilistic algorithms, reasoning or problem-solving algorithms, expert system-based algorithms, simple logic or addition / multiplication algorithms, algorithms in the fields of combinatorial optimization or operations research, algorithms in the fields of machine learning or statistical learning, and algorithms related to complex systems or deep neural networks. Neural networks can be used to perform various tasks such as discrimination, generation, and evolution.

[0039] A physical intrinsic system can be formed by heterogeneous / homogeneous integration of multiple materials in a specific structure through multi-layer stacking, or by a combination of one or more physical devices based on physical, chemical, or biological principles. In this disclosure, "physics" emphasizes the entity of the system and its constituent elements, as well as a general reference to various natural science principles, and is not limited to physical laws in the traditional sense. The system exhibits static or dynamic response functions to multiple inputs / outputs, thereby exhibiting specific response characteristics, and these response characteristics are adjustable, thereby achieving an effect similar to parameter optimization in a neural network.

[0040] It should be noted that although traditional silicon-based integrated circuits also involve the use of physical principles, their implementation of physical intrinsic neural networks is different from the implementation of neural networks using physical intrinsic systems in the context of this disclosure. For example, a field-effect transistor (MOSFET) uses the electric field effect: after a voltage is applied to the gate, an electric field is generated through the insulating layer, thereby inducing a conductive channel in the semiconductor channel, so that a response characteristic is formed between the output current and the gate voltage. However, silicon-based integrated circuits mainly rely on transistors to construct logic circuits with computing functions, and realize the calculation process and the storage of weights, biases and intermediate results through digital operations, and do not directly use the inherent response characteristics of the components to process the input signals of the network layer.

[0041] The disclosed physical eigensystem utilizes its own response characteristics to autonomously implement neural network operator functions such as feature extraction, forward propagation, objective function calculation, gradient acquisition, random perturbation, and parameter update. This response characteristic can be either time-varying or time-invariant.

[0042] In summary, the physical intrinsic system disclosed in the present invention is fundamentally different from traditional silicon-based integrated circuits when implementing some functions of neural networks.

[0043] According to some embodiments, the physical input signal and the physical output signal are analog signals and may be based on at least one selected from the group consisting of: magnetic field, heat, pulse, microwave, radio frequency signal, electrical signal, optical signal, acoustic signal, mechanical signal, vibration signal, time signal and space signal respectively.

[0044] In some embodiments, a temporal signal can refer to changes in a signal along a time axis and may also include the duration, delay, and dynamic time-varying characteristics of a response. This refers to the duration, rate, and pattern of changes in the response of a physical intrinsic system at different time points or time periods. Similarly, a spatial signal can refer to the spatial distribution of a signal, such as the distribution of the response across different regions, local structure or gradient characteristics, and the spatial distribution pattern within a physical intrinsic system.

[0045] In step S101 , input data of a multi-layer neural network is determined, and a target layer to be trained in the multi-layer neural network is determined.

[0046] The input data of the multi-layer neural network can be training data for training mode or data to be inferred for inference mode. In some embodiments, the input data can be used as training data and data to be inferred at the same time, that is, the multi-layer neural network implemented by the present disclosure can be in inference mode and training mode at the same time. Specifically, in the subsequent step S102, the target layer and the previous part of the multi-layer neural network perform layer-by-layer inference on the input data by forward propagation, thereby obtaining the output signal of the target layer, and then using the output signal to complete the training of the target layer in step S103. It can be seen that the layer-by-layer forward propagation from the first layer to the target layer of the multi-layer neural network performed in step S102 has completed partial inference of the input data. At this time, the output signal of the target layer is further input into the subsequent layers of the multi-layer neural network and layer-by-layer forward propagation is performed, which can complete the complete inference of the input data. Therefore, the input data can be used as training data and data to be inferred at the same time.

[0047] In some embodiments, the multi-layer neural network may include only a first network layer implemented by a first physical eigensystem (the first network layer may include one or more layers), or may include a second network layer implemented by a traditional silicon-based integrated circuit (as described below). The target layer may include the first network layer, the second network layer, or both.

[0048] In step S102, the input data of the multi-layer neural network can first be converted into the input signal of the first layer, and then the output of each layer is used as the input of the next layer in sequence until the output signal of the target layer is generated. In this process, from the first layer to the target layer, each layer performs a forward propagation operation on the input signal received by the layer. In some embodiments, the forward propagation of the first network layer implemented in the physical eigensystem can be driven by the physical input signal and performed based on the inherent response characteristics of the system.

[0049] In an exemplary embodiment, the physical intrinsic system can be realized by nanomagnets. For example, after cutting 3mm×3mm units from a wafer, artificial spin ice (such as Figure 3 (as shown), on which are engraved multiple nano-units that can be used for data storage. By applying a magnetic field from eight directions, namely, directly above, upper right, directly right, lower right, directly below, lower left, directly left and upper left, the magnetization state of these units can be biased to one side, thereby reflecting their inherent response characteristics. Artificial spin ice can be integrated on a waveguide, and its input excitation (i.e., input physical signal) uses microwaves (for example, multi-spectrum signals), and the physical output signal can also be microwaves. Since the system constitutes an ordered and complex physical platform, there is a high-order complex function mapping between its output and input, which can be rigorously analyzed by statistical physics methods. Therefore, adjusting the way of applying the magnetic field can directly control the response characteristics of artificial spin ice. Its complexity and nonlinearity mainly come from the scattering effect of the magnetization state on the propagation of spin waves and the coupling between magnets. This complexity can be further enhanced through precision micromachining and material growth.

[0050] In another exemplary embodiment, a physical intrinsic system can be implemented through sound frequency regulation. Specifically, multiple needles can be arranged on a spherical object, each with a different inherent response frequency, thus forming a complex response system. In this system, sound serves as the physical input signal, and the vibration amplitude of each needle serves as the physical output signal (which can be monitored in a variety of ways). In addition, by installing one or more annular structures on the needle body, or adopting other adaptive adjustment methods, the response frequency of the needle can be dynamically adjusted, thereby achieving fine control of the response characteristics of the entire system.

[0051] In step S103 , the goodness of the input data is determined based on the output signal of the target layer, and the adjustable parameters of the target layer are adjusted based on the goodness.

[0052] In the technical solutions disclosed herein, "goodness" can be understood as a measure of the quality of the target layer's response to input data. When the target layer's processing results for the input data are more consistent with expectations or closer to the network's target requirements, the goodness will be higher; conversely, it will be lower. Therefore, by determining the goodness of input data based on the target layer's output signal, it is possible to determine whether the target layer's performance on the input data is "good" or "bad," and this information can then be used to optimize the target layer's parameters.

[0053] In some embodiments, goodness can be obtained in a variety of ways. For example, an objective function can be predefined based on task requirements, which reflects the degree of match between the output of the target layer and the desired target. After a forward propagation is performed using the target layer, the actual output of the target layer can be brought into the objective function (or a physical device that characterizes the objective function) to obtain goodness. A reference signal can also be preset based on experience, prior knowledge or heuristic methods, and the goodness can be obtained by comparing the degree of difference between the actual output of the target layer and the reference signal. In addition, statistical or probabilistic methods can be combined to evaluate the distribution of the response of the target layer under a large amount of input data or different sampling, and define goodness based on this evaluation to measure the current layer's ability to distinguish or adapt to different types of data. It should be noted that the above-mentioned multiple ways of determining goodness are not mutually exclusive, so the specific goodness definition method can be selected, combined or adjusted according to the application scenario and requirements.

[0054] In an exemplary embodiment, the goodness can be expressed as the square sum of activation values of all neurons in the target layer. The activation value can be an activation value that is not layer-normalized.

[0055] Goodness can take many forms. For example, it can be represented using numerical values in silicon-based integrated circuits, or it can be represented in the physical domain using physical quantities. Both of these representations are discussed in detail below.

[0056] After determining the goodness of fit, it can be used to guide the update of neural network parameters. In some embodiments, if the goodness of fit is low, the adjustable parameters of the target layer can be adjusted to significantly increase the goodness of fit; if the goodness of fit is high, the adjustment can be relatively conservative. This approach allows for training and optimization of the target layer during the forward propagation process without backpropagating gradients. As new data is continuously input and iterative updates are performed, the target layer will eventually tend to exhibit a more optimal response for the desired task.

[0057] This approach can also significantly simplify the hardware design used to implement neural networks. The forward-propagation-based training method enables parameter adjustments at the layer level or even locally at the node level, while the lack of backpropagation eliminates the need for a complete gradient return path or large-scale multiply-add arrays in hardware design. This eliminates the multiply-accumulate units, gradient registers, and reverse links typically used for backpropagation.

[0058] According to some embodiments, the target layer to be trained may include a first network layer implemented by a first physical eigensystem, and the adjustable parameters of the target layer may include at least one adjustable mechanism in the first physical eigensystem for adjusting the response characteristics. The at least one adjustable mechanism may include at least one of the following: the position and / or orientation of devices in the first physical eigensystem, the distance between devices, the magnetic field position, the magnetic field direction and / or strength, the pulse current, the voltage bias, the current strength and / or direction, the resistance, the conductance, the temperature, the pressure, the mechanical strain, the optical parameter, the acoustic parameter, the vibration frequency, the spectral characteristics of the electromagnetic signal, the chemical concentration, and the biomolecule activity.

[0059] Therefore, through the above method, the parameter update method based on forward propagation can be organically combined with the physical eigensystem, and the response characteristics of the physical eigensystem can be changed by adjusting the adjustable parameters of the physical eigensystem based on the goodness, thereby realizing the training of the first network layer implemented by the first physical eigensystem without the need for backpropagation gradient.

[0060] In some embodiments, the tunable mechanism may include localized magnetic field changes, such as increasing the magnetic field at a localized location on the device, or changing the direction and / or intensity of a localized magnetic field. Optical parameters may include light intensity, frequency, wavelength, phase, polarization state, divergence angle, etc. Acoustic parameters may include sound intensity, frequency spectrum, etc.

[0061] In some embodiments, the response characteristic can be static or dynamic. For example, the response characteristic can include the rate of change of the current. Accordingly, the adjustable mechanism can be an adjustment of the dynamic response characteristic, for example, increasing or decreasing the rate of change of the current. It will be understood that the above description is merely exemplary. The dynamic response characteristic can also include, for example, changes in the field in space or time, and the adjustable mechanism can include adjustments for such changes. Both can include a richer range of implementations, and neither is limited herein.

[0062] In some embodiments, before training the first network layer, the relationship between at least one adjustable mechanism of the first physical eigensystem and the response characteristic can be determined in advance through experiments, simulations, or other means. Specifically, the relationship between adjusting the at least one adjustable mechanism to change the response characteristic in a specific direction can be determined. In this disclosure, this relationship is also referred to as a preset adjustment strategy. In step S103, the at least one adjustable mechanism of the first physical eigensystem can be adjusted based on this relationship and the direction of the desired change in goodness.

[0063] In one exemplary embodiment, in an example physical eigensystem implemented using artificial spin ice, the response characteristics of the example physical eigensystem can be altered by adjusting the strengths of multiple magnetic fields. Therefore, the strengths of the multiple magnetic fields can be used as at least one tunable mechanism of the example physical eigensystem. Preliminary experiments can determine the correlation between the strengths of the multiple magnetic fields and the microwaves output by the waveguide, thereby enabling targeted adjustment of the at least one tunable mechanism.

[0064] According to some embodiments, the input data may include positive data and negative data, and the goodness of the input data may include positive goodness and negative goodness. Step S103, based on the output signal of the target layer, determines the goodness of the input data, and adjusting the adjustable parameters of the target layer based on the goodness may include: adjusting at least one adjustable mechanism in a direction of increasing the positive goodness; and / or adjusting at least one adjustable mechanism in a direction of decreasing the negative goodness.

[0065] Therefore, by inputting positive data and negative data into the same target layer respectively and performing two forward propagations, the dual parameter adjustment goals of "enhancing positive goodness" and "suppressing negative goodness" can be achieved simultaneously without backpropagation. On the one hand, the response to correct input is strengthened, and on the other hand, the response to incorrect input is weakened. As a result, the target layer can bidirectionally constrain the response characteristics of the physical eigensystem through forward information without the need for gradient backpropagation, so that the neural network converges faster.

[0066] In some embodiments, "positive data" can be real or expected input samples. For example, in an image classification scenario, positive data are real images and their labels that the network needs to correctly identify. "Negative data" are "erroneous" or "false" samples corresponding to positive data. They can be obtained through network self-generation (for example, the neural network automatically generates sample data that is different from the positive data based on certain rules), artificial destruction (for example, adding random perturbations such as noise), or introduction from external data sources.

[0067] After inputting positive data into the multi-layer neural network and obtaining the positive data output signal of the target layer, the goodness determined using the positive data output signal is the positive goodness. Similarly, after inputting negative data into the multi-layer neural network and obtaining the negative data output signal of the target layer, the goodness determined using the negative data output signal is the negative goodness.

[0068] In an exemplary embodiment, the goodness can be expressed as the sum of the squares of the activation values of all neurons in the target layer. The goal of optimizing the target layer is to make the positive goodness corresponding to the positive data higher than a certain threshold, and to make the negative goodness corresponding to the negative data lower than the threshold. The obtained goodness can be subtracted from a preset threshold, and then a logical function (such as sigmoid) is applied to convert the value into a probability output, which indicates the probability that the input data is positive data. The probability output can be used as a constraint to determine the adjustment amount of at least one adjustable mechanism of the first physical eigensystem.

[0069] In an exemplary embodiment, the cosine similarity between positive data and a random vector x can be determined as the positive goodness, and the cosine similarity between negative data and the same random vector x can be determined as the negative goodness. By adjusting at least one adjustable mechanism in a direction of increasing the positive goodness and / or adjusting at least one adjustable mechanism in a direction of decreasing the negative goodness, the trained target layer has the ability to separate positive data from negative data.

[0070] According to some embodiments, Figure 2 As shown, process 200 can be used to implement step S103, determining the goodness of input data based on the output signal of the target layer, and adjusting the adjustable parameters of the target layer based on the goodness. Process 200 may include: step S201, converting the physical output signal into a digital signal; step S202, performing the following operations using a first silicon-based integrated circuit; step S2021, calculating a first value representing the goodness based on a preset objective function and the digital signal; and step S2022, calculating a second value representing an adjustment amount for at least one adjustable mechanism based on a preset adjustment strategy and the first value, wherein the preset adjustment strategy indicates adjusting the at least one adjustable mechanism in a direction of directional optimization of the goodness; and step S203, converting the second value into a first physical quantity and applying it to the at least one adjustable mechanism.

[0071] Figure 3 A schematic diagram illustrating calculation of goodness and an adjustment amount of an adjustable mechanism for a physical intrinsic system using a silicon-based integrated circuit according to an exemplary embodiment of the present disclosure is shown.

[0072] By converting the physical output signal into a digital signal and then using the first silicon-based integrated circuit to perform the objective function calculation and the numerical solution of the adjustment amount, an organic combination of the physical eigensystem and the digital processing unit is achieved. Specifically, the analog response generated by the physical eigensystem is first digitized, and then the goodness and adjustment amount are calculated in the silicon-based integrated circuit based on the objective function and the preset adjustment strategy. Finally, the adjustment amount is converted into the corresponding physical quantity and applied to at least one adjustable mechanism of the first physical eigensystem. The above method not only utilizes the response characteristics of the physical device for forward propagation, but also utilizes the silicon-based integrated circuit for flexible and precise calculation and parameter adjustment, thereby efficiently completing the optimization of the target layer and improving the adaptability and performance of the overall neural network.

[0073] It is understandable that the specific meanings of the preset objective function and the preset adjustment strategy can be found in the corresponding description above and will not be elaborated here.

[0074] According to some embodiments, Figure 4 As shown, process 400 can be used to implement step S103, determining the goodness of input data based on the output signal of the target layer, and adjusting the adjustable parameters of the target layer based on the goodness. Process 400 may include: step S401, processing the physical output signal using a first eigenmapping channel of a physical intrinsic goodness determination device to obtain a second physical quantity representing the goodness, wherein the first eigenmapping channel represents a preset target function to evaluate the response quality of the first network layer to the physical input signal in the physical domain; step S402, processing the second physical quantity using a second eigenmapping channel of a physical intrinsic response characteristic adjustment device to obtain a third physical quantity that can be directly applied to at least one adjustable mechanism, wherein the second eigenmapping channel represents a preset adjustment strategy to determine the adjustment amount for at least one adjustable mechanism in the physical domain to optimize the goodness in a targeted manner; and step S403, applying the third physical quantity to at least one adjustable mechanism.

[0075] Figure 5 A schematic diagram is shown of determining goodness and adjusting response characteristics of a physical eigensystem in a purely physical domain according to an exemplary embodiment of the present disclosure.

[0076] By setting up a physical intrinsic goodness determination device and a physical intrinsic response characteristic adjustment device, the entire neural network training optimization process is completed in the physical domain, including the calculation of goodness, the solution of the adjustment amount, and the final application of the adjustable mechanism. Compared with the solution of converting the analog signal into a digital signal and then performing the calculation in the silicon-based circuit, this method bypasses the complex digital gradient operation logic, allowing the adjustable mechanism of the physical intrinsic system to be closely integrated with the neural network training process, improving the utilization efficiency of the physical process, and to a certain extent expanding the possibility of deep integration of neural networks and various physical intrinsic devices. In addition, since the entire training optimization process is completed in the physical domain, there is no need to use analog-to-digital conversion elements, thereby avoiding the accuracy loss and additional hardware overhead caused by analog-to-digital conversion, and the coordination between the various physical intrinsic devices is closer.

[0077] An intrinsic mapping channel is a functional channel within a physically intrinsic system or device, formed by the inherent characteristics of the device or the combination of devices. It is used to perform specific signal mapping or similar operations in the physical domain. It uses the system's inherent response mechanisms (such as optical interference, mechanical strain, magnetic coupling, chemical reaction, etc.) to convert input signals into output signals, thereby mapping some physical quantity (such as activation value) into a measure of goodness or regulation.

[0078] The eigenmap pipeline can be viewed as a physical implementation of a "predetermined objective function" or "predetermined regulation strategy." For example, if the predetermined objective function defines a weighted summation or nonlinear activation, the eigenmap pipeline implements a similar operation using the device's inherent physical response without requiring conversion to the digital domain for computation.

[0079] Compared with the pure digital operations of silicon-based integrated circuits, the intrinsic mapping channel "writes" the preset target function or preset adjustment strategy in the physical domain, avoiding additional digital-to-analog conversion or dependence on digital processing units. It can fully utilize the advantages of parallelism, low power consumption or high bandwidth of the physical system, and more directly utilize the inherent characteristics of the device to complete the training of the first network layer, eliminating the need for a complete digital logic and software operation process.

[0080] The goodness determination device can convert the physical output signal into a second physical quantity that characterizes the goodness. In an exemplary embodiment, the goodness determination device can be an optical interference structure. By superimposing and interfering multiple optical signals through optical interference and phase modulation, it is possible to output a light intensity that reflects the input signal strength (such as the sum of squares or other metrics). This light intensity can be regarded as a first physical quantity that characterizes the goodness, which is used to evaluate the response quality of the first network layer to the input data. In this embodiment, the first eigenmapping channel can be a channel that maps multiple optical inputs to an optical interference output, including interference, phase modulation and other links, to achieve a preset objective function at the physical level and obtain the goodness.

[0081] The response characteristic adjustment device can further convert the second physical quantity characterizing the goodness into a third physical quantity that can be directly applied to at least one adjustable mechanism. In an exemplary embodiment, the response characteristic adjustment device can be a magnetic field adjustable module. If the adjustable mechanism of the first physical eigensystem includes magnetic field strength adjustment, the response characteristic adjustment device can use a current-controlled coil to convert the goodness signal into the coil current magnitude, thereby changing the magnetic field strength in real time to affect the magnetic response characteristics of the first physical eigensystem. In this embodiment, the second eigenmapping channel can be a physical path that converts the physical quantity "goodness" into the coil current magnitude in real time, and dynamically adjusts the magnetic field according to the "preset adjustment strategy", and finally completes the update of the adjustable mechanism (magnetic field strength) of the first physical eigensystem.

[0082] According to some embodiments, the input data may include positive data and negative data, and the goodness of the input data may include positive goodness and negative goodness. The second eigenmapping channel may include a first subchannel representing a first preset adjustment strategy and a second subchannel representing a second preset adjustment strategy, the first subchannel may be configured to process a physical quantity representing positive goodness, the first preset adjustment strategy may indicate adjusting at least one adjustable mechanism in a direction of increasing positive goodness, and the second subchannel may be configured to process a physical quantity representing negative goodness, the second preset adjustment strategy may indicate adjusting at least one adjustable mechanism in a direction of reducing negative goodness.

[0083] Therefore, by configuring the first and second subchannels separately within the second eigenmap channel, it is possible to simultaneously process positive and negative goodness signals in the physical domain, employing either enhancement or suppression strategies to adjust the target layer's adjustable mechanism. This approach allows the system to learn from both positive and negative data in parallel and independently, strengthening its response to real data while reducing over-response to erroneous or false inputs, thereby achieving a more flexible and efficient optimization process.

[0084] According to some embodiments, the multi-layer neural network may further include a second network layer implemented by a second silicon-based integrated circuit. The second silicon-based integrated circuit may be configured to receive an electrical input signal of the second network layer and calculate an electrical output signal of the second network layer based on model parameters of the second network layer stored in the second silicon-based integrated circuit. The adjustable parameters of the target layer include the model parameters of the second network layer.

[0085] The implementation method of the physical intrinsic neural network may also include: using a signal conversion device to convert the electrical output signal of the second network layer into a physical input signal of the first network layer; or using a signal conversion device to convert the physical output signal of the first network layer into an electrical input signal of the second network layer.

[0086] The forward transmission of the second network layer implemented by the second silicon-based integrated circuit can be achieved by the operation unit on the second silicon-based integrated circuit, combined with the stored model parameters such as weights and biases, through corresponding mathematical calculations.

[0087] The second silicon-based integrated circuit and the first physical intrinsic system work together, each taking on the computation of a portion of the neural network's layers, thereby collaboratively implementing neural network training and inference. Because the second silicon-based integrated circuit's input and output are based on electrical signals, while the first physical intrinsic system's input and output are based on physical signals, a signal conversion device is used to connect the second silicon-based integrated circuit and the first physical intrinsic system, enabling them to be used collaboratively to implement the neural network.

[0088] In some embodiments, the model parameters of the second network layer may include weights and biases of neurons. The electrical input signal and the electrical output signal may be digital signals for digital circuits or analog signals for analog chips, without limitation.

[0089] According to some embodiments, the second silicon-based integrated circuit may include at least one item selected from the group consisting of: a field programmable gate array (FPGA), an application specific chip (ASIC), a graphics processing unit (GPU), a central processing unit (CPU), a neural processing unit (NPU), a tensor processing unit (TPU), a very large scale integrated circuit (VLSI), a complementary metal oxide semiconductor (CMOS), and an integrated module composed of transistors and / or other electrical devices laid out and wired based on deposition, lithography and / or etching.

[0090] According to some embodiments, the target layer may include a second network layer. Step S103, determining the goodness of the input data based on the output signal of the target layer, and adjusting the adjustable parameters of the target layer based on the goodness may include: utilizing a third silicon-based integrated circuit to perform the following operations: calculating a third value representing the goodness based on a preset objective function and the electrical output signal of the second network layer; and adjusting the model parameters of the second network layer based on the preset adjustment strategy and the third value.

[0091] The process for training and optimizing the second network layer using the third silicon-based integrated circuit can refer to the process for training and optimizing the first network layer using the first silicon-based integrated circuit described above. The preset objective function and preset adjustment strategy used for the second network layer can be the same as or different from those used for the first network layer, and are not limited here.

[0092] In some embodiments, the first network layer can be used as a nonlinear feature extraction layer to extract nonlinear features. The computing layer can be trained using a silicon-based integrated circuit (eg, a GPU) without backpropagation.

[0093] According to some embodiments, the first network layer can be used to extract nonlinear features, the multi-layer neural network can further include a linear layer after the nonlinear feature extraction layer, and the multi-layer neural network can further include a third network layer after the first network layer. A second physical eigensystem different from the first physical eigensystem can be used to implement at least one of the following multiple operators for the third network layer: feature extraction, forward propagation, objective function calculation, gradient acquisition, random perturbation, and parameter update.

[0094] In some embodiments, a second physical intrinsic system can be used to implement gradient calculation to complete the training of the linear layer. Specifically, a ReLU-like nonlinear activation function (x+1)*log(x+1)–x can be used, so that the linear layer can be trained using a forward pass using a physical intrinsic-based gradient calculation method (the gradient function is log(x+1)).

[0095] In an exemplary embodiment, when a first physical eigensystem is formed using nanomagnets, and in an embodiment where a multi-layer neural network includes a first network layer and other layers at the same time, the first network layer may not be trained and one of the following may be performed: (1) adding a single or multiple external fields; (2) not adding any magnetic field.

[0096] In some embodiments, the multi-layer neural network may include only the first network layer for extracting nonlinear features.

[0097] Figure 6 FIG. 1 shows a flow chart of a method for implementing a physical intrinsic neural network according to an exemplary embodiment of the present disclosure. Figure 6As shown, method 600 includes: step S604, inputting the output signal of the target layer into the subsequent layer of the multi-layer neural network to perform forward propagation. The forward propagation in the subsequent layer of the multi-layer neural network and the adjustment of the adjustable parameters of the target layer are performed asynchronously. It can be understood that Figure 6 The operations and effects of steps S601 to S603 can refer to the above description of steps S101 to S103 and will not be repeated here.

[0098] Return to step S101. In some embodiments, multiple target layers in a multi-layer neural network can be determined, including a first target layer and a second target layer following the first target layer. In steps S102 and S103, the output signal of the first target layer can be obtained, and a first goodness can be determined, and then the parameters of the first target layer can be adjusted based on the first goodness. Step S205 can further include a training operation for the second target layer: obtaining the output signal of the second target layer; determining the second goodness of the second target layer for the input data based on the output signal of the second target layer; and adjusting the adjustable parameters of the second target layer based on the second goodness.

[0099] In some embodiments, after the parameters of the first target layer are adjusted, the parameters of the first target layer can be decoupled, and other layers can be trained or inferred independently. The forward propagation and the adjustment of the adjustable parameters of the target layer in the subsequent layers of the multi-layer neural network can be performed asynchronously, and the parameter adjustment of the first target layer and the parameter adjustment of the second target layer can also be performed asynchronously. The traditional back-propagation gradient parameter adjustment method requires that the parameters to be adjusted be determined based on the gradient of each neuron after the forward propagation and back-propagation of the entire neural network are completely completed. The implementation method of the physical intrinsic neural network proposed in the present disclosure can adjust the parameters of multiple target layers during the forward propagation process, and the parameter adjustment between different target layers is independent and supports asynchronous processing, thereby significantly improving the training efficiency of the neural network.

[0100] The parallelism and randomness of physical intrinsic systems, combined with forward-propagation-based training, can further improve training efficiency and enable parallel training. Parallel training can include inter-layer parallelism, parallelism between different computing nodes within a layer, and parallelism for batch processing of multiple samples. Furthermore, the combination of the two eliminates the need to wait for full network backpropagation before training, allowing network performance to be improved within local loops of the physical intrinsic system.

[0101] The following will be combined Figure 7 and Figure 8 The traditional forward-backward propagation algorithm and the forward propagation and parameter adjustment methods used in the present disclosure are explained.

[0102] Figure 7Schematic diagram of the forward-backward propagation algorithm according to an exemplary embodiment of the present disclosure is shown. Figure 7 As shown in the figure, process A is first executed, that is, forward propagation of all network layers. Then process B is executed to calculate the global loss based on the output signal of the last network layer and the ground truth. Then process C is executed to calculate the gradient corresponding to each adjustable parameter of each layer through backpropagation. Based on these gradients, the network layer and corresponding parameters that need to be adjusted are determined, and the adjustment amount is determined. Therefore, in the forward-backward propagation algorithm, the forward propagation of all network layers must be completed before the parameters can be adjusted. During this process, the weights and other parameters of all network layers need to be frozen.

[0103] Figure 8 FIG. 1 shows a schematic diagram of forward propagation and parameter adjustment of multiple network layers according to an exemplary embodiment of the present disclosure. Figure 8 As shown, the forward propagation and parameter adjustment of the first and second network layers can be divided into four processes. First, process A is executed. Since both processes B and C depend on the first output signal obtained after process A, and there is no dependency between processes B and C, processes B and C can be executed asynchronously after process A is completed. In other words, after the forward propagation of the first network layer is completed, the forward propagation of the second network layer following the first network layer and the parameter adjustment of the first network layer can be executed asynchronously.

[0104] In addition, process D depends on the second output signal obtained after process C is executed, so process D needs to be executed after process C, but process D and process B can be executed asynchronously. From the above examples, it can be seen that in the method proposed in the present disclosure, each layer is independent of each other and does not need to rely on the back propagation of the global loss of the last layer to guide the update. At the same time, training data and inference data can be fed into the model interspersed to complete the corresponding training and inference tasks. That is, after a certain layer of the network completes inference, it can be trained immediately, unlike the forward-backward propagation algorithm, which needs to freeze the weights of the entire model and suspend the inference task during training until all layers of the network are updated.

[0105] According to another aspect of the present disclosure, a device for implementing a physical intrinsic neural network is provided. The multi-layer neural network includes a first network layer implemented by a first physical intrinsic system, the first physical intrinsic system being configured to receive a physical input signal of the first network layer and output a physical output signal of the first network layer based on a response characteristic of the first physical intrinsic system. Figure 9As shown, the device 900 includes: a first determination unit 910, configured to determine the input data of the multi-layer neural network and determine the target layer to be trained in the multi-layer neural network; an acquisition unit 920, configured to input the input data into the multi-layer neural network and obtain the output signal of the target layer; and a determination unit 930, configured to determine the goodness of the input data based on the output signal of the target layer, and adjust the adjustable parameters of the target layer based on the goodness, wherein the goodness describes the response quality of the target layer to the input data.

[0106] It can be understood that the operations and effects of units 910 to 930 in the apparatus 900 may refer to the above description of steps S101 to S103 in the method 100 and are not described in detail here.

[0107] According to some embodiments, the target layer may include a first network layer, and the adjustable parameters of the target layer may include at least one adjustable mechanism in the first physical intrinsic system for adjusting the response characteristics, and the at least one adjustable mechanism may include at least one of the following: the position and / or orientation of the device in the first physical intrinsic system, the distance between the devices, the magnetic field position, the magnetic field direction and / or intensity, the pulse current, the voltage bias, the current intensity and / or direction, the resistance, the conductivity, the temperature, the pressure, the mechanical strain, the optical parameters, the acoustic parameters, the vibration frequency, the spectral characteristics of the electromagnetic signal, the chemical concentration and the biological molecule activity.

[0108] According to some embodiments, the input data may include positive data and negative data, and the goodness of the input data may include positive goodness and negative goodness. The determining unit may be configured to: adjust at least one adjustable mechanism in a direction to increase the positive goodness; and / or adjust at least one adjustable mechanism in a direction to decrease the negative goodness.

[0109] According to some embodiments, the determination unit may include: a first conversion subunit, configured to convert the physical output signal into a digital signal; a first silicon-based integrated circuit, configured to perform the following operations: calculating a first numerical value representing goodness based on a preset objective function and the digital signal; and calculating a second numerical value representing an adjustment amount for at least one adjustable mechanism based on a preset adjustment strategy and the first numerical value, wherein the preset adjustment strategy indicates adjusting at least one adjustable mechanism in the direction of directional optimization of goodness; and a second conversion subunit, configured to convert the second numerical value into a first physical quantity and apply it to at least one adjustable mechanism.

[0110] According to some embodiments, the determination unit may include: a physical intrinsic goodness determination device, configured to process the physical output signal using a first intrinsic mapping channel therein to obtain a second physical quantity characterizing the goodness, wherein the first intrinsic mapping channel characterizes a preset objective function to achieve evaluation of the response quality of the first network layer to the physical input signal in the physical domain; a physical intrinsic response characteristic adjustment device, configured to process the second physical quantity using a second intrinsic mapping channel therein to obtain a third physical quantity that can be directly applied to at least one adjustable mechanism, wherein the second intrinsic mapping channel characterizes a preset adjustment strategy to achieve determination of the adjustment amount for at least one adjustable mechanism in the physical domain to directionally optimize the goodness; and an adjustment subunit, configured to apply the third physical quantity to at least one adjustable mechanism.

[0111] According to some embodiments, the input data may include positive data and negative data, and the goodness of the input data may include positive goodness and negative goodness. The second eigenmapping channel may include a first subchannel representing a first preset adjustment strategy and a second subchannel representing a second preset adjustment strategy, the first subchannel may be configured to process a physical quantity representing positive goodness, the first preset adjustment strategy may indicate adjusting at least one adjustable mechanism in a direction of increasing positive goodness, and the second subchannel may be configured to process a physical quantity representing negative goodness, the second preset adjustment strategy may indicate adjusting at least one adjustable mechanism in a direction of decreasing negative goodness.

[0112] According to some embodiments, the physical input signal and the physical output signal may be analog signals, and may be based on at least one item selected from the group consisting of: magnetic field, heat, pulse, microwave, radio frequency signal, electrical signal, optical signal, acoustic signal, mechanical signal, vibration signal, time signal and space signal respectively.

[0113] According to some embodiments, the implementation device of the neural network may further include: a forward propagation unit, configured to input the output signal of the target layer into a subsequent layer of the multi-layer neural network to perform forward propagation, wherein the forward propagation and the adjustment of the adjustable parameters of the target layer in the subsequent layer of the multi-layer neural network are performed asynchronously.

[0114] According to some embodiments, the multi-layer neural network may further include a second network layer implemented by a second silicon-based integrated circuit. The second silicon-based integrated circuit may be configured to receive an electrical input signal of the second network layer and calculate an electrical output signal of the second network layer based on the model parameters of the second network layer stored in the second silicon-based integrated circuit. The adjustable parameters of the target layer may include the model parameters of the second network layer. The device may further include a signal conversion device configured to: convert the electrical output signal of the second network layer into a physical input signal of the first network layer; or convert the physical output signal of the first network layer into an electrical input signal of the second network layer.

[0115] According to some embodiments, the second silicon-based integrated circuit may include at least one item selected from the group consisting of: a field programmable gate array (FPGA), an application specific chip (ASIC), a graphics processing unit (GPU), a central processing unit (CPU), a neural processing unit (NPU), a tensor processing unit (TPU), a very large scale integrated circuit (VLSI), a complementary metal oxide semiconductor (CMOS), and an integrated module composed of transistors and / or other electrical devices laid out and wired based on deposition, lithography and / or etching.

[0116] According to some embodiments, the target layer may include a second network layer. The determining unit may include: a third silicon-based integrated circuit configured to: calculate a third value representing goodness of fit based on a preset target function and an electrical output signal of the second network layer; and adjust a model parameter of the second network layer based on a preset adjustment strategy and the third value.

[0117] According to some embodiments, the first network layer may be used to extract nonlinear features, and the multi-layer neural network may further include a third network layer following the first network layer. A second physical eigensystem, different from the first physical eigensystem, may be used to implement at least one of the following operators for the third network layer: feature extraction, forward propagation, objective function calculation, gradient acquisition, random perturbation, and parameter update.

[0118] According to some embodiments, the multi-layer neural network may be operated with at least one selected from the group consisting of: a deterministic algorithm, a random or probabilistic algorithm, an inference or solution algorithm, an expert system-based algorithm, a simple logic or addition / multiplication algorithm, an algorithm in the field of combinatorial optimization or operations research, an algorithm in the field of machine learning or statistical learning, and an algorithm related to complex systems or deep neural networks.

[0119] According to another aspect of the present disclosure, a system is provided for implementing a multi-layer neural network. Figure 10 As shown, the system 1000 includes: a first physical intrinsic system 1010 , and at least one of a first control unit 1020 and a second control unit 1030 .

[0120] The first physical eigensystem 1010 is configured to implement a first network layer in a multi-layer neural network; and receive a physical input signal of the first network layer, and output a physical output signal of the first network layer based on the response characteristics of the first physical eigensystem.

[0121] The first control unit 1020 includes: at least one processor; and a memory communicatively connected to the at least one processor, wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor and the first physical eigensystem can collaboratively execute at least one step of method 100 or method 600.

[0122] The second control unit 1030 includes: a third physical intrinsic mapping channel, wherein the third physical intrinsic mapping channel and the first physical intrinsic system can cooperate to perform at least one step of the method 100 or the method 600 in the physical domain.

[0123] In some embodiments, the first physical eigensystem 1010 can be composed of different materials or a stack of the same materials or a structure, and can be biological or chemical. It should have multiple inputs and outputs, be sensitive to certain physical quantities, and have linear or nonlinear responses. Such responses can be mechanical, thermal, optical or photonic, electrical, magnetic or spintronic, acoustic, or quantum mechanical, or a combination thereof.

[0124] First, physical intrinsic systems can be combined or integrated on a large scale. This means that one or more devices can be combined through macroscopic mechanics-based methods or microfabrication methods such as film growth, sputtering, chemical deposition, physical deposition, photolithography, and etching, either within the same plane or in different planes of a three-dimensional structure, and then connected using electrical, magnetic, or optical-based devices such as wires of different dielectrics or other physical quantities (connectors also fall within the aforementioned physical intrinsic systems). This combination can range from a few devices to a large-scale combination of tens of thousands or even hundreds of millions of devices, to achieve the functionality of an artificial neural network.

[0125] In some embodiments, the first control unit 1020 or the second control unit 1030 can be used to determine input data and a target layer, and to perform control logic for forward-pass-based training.

[0126] In some embodiments, the system 1000 may further include a silicon-based integrated circuit. The first physical intrinsic system 1010 and the silicon-based integrated circuit may each bear a portion of the neural network computing function.

[0127] In some embodiments, the system 1000 may further include a signal conversion device for converting the electrical signal output by the silicon-based integrated circuit into a physical input signal of the first network layer and / or converting the physical output signal of the first network layer into an electrical signal for input into the silicon-based integrated circuit.

[0128] According to another aspect of the present disclosure, a computer-readable storage medium is provided, having a computer program stored thereon. The computer program is configured to implement a multi-layer neural network, the multi-layer neural network including a first network layer implemented by a first physical eigensystem, the first physical eigensystem being configured to receive a physical input signal of the first network layer and output a physical output signal of the first network layer based on a response characteristic of the first physical eigensystem. When executed by a processor, the computer program enables the processor and the first physical eigensystem to collaboratively perform the aforementioned method 100 or method 600 for implementing a physical eigensystem.

[0129] According to another aspect of the present disclosure, a computer program product is provided, including a computer program for implementing a multi-layer neural network, the multi-layer neural network including a first network layer implemented by a first physical eigensystem, the first physical eigensystem configured to receive a physical input signal of the first network layer and output a physical output signal of the first network layer based on a response characteristic of the first physical eigensystem. When executed by a processor, the computer program enables the processor and the first physical eigensystem to collaboratively perform the aforementioned method 100 or method 600 for implementing a physical eigensystem.

[0130] In the following, combined Figure 11 Illustrative examples of such a computer device, computer-readable storage medium, and computer program product are described.

[0131] Figure 11 An example configuration of a computer device 1100 is shown that may be used to implement the methods described herein.

[0132] Computer device 1100 can be a variety of different types of devices. Examples of computer device 1100 include, but are not limited to, desktop computers, server computers, laptop or netbook computers, mobile devices (e.g., tablet computers, cellular or other wireless phones (e.g., smartphones), notepad computers, mobile stations), wearable devices (e.g., eyeglasses, watches), entertainment devices (e.g., entertainment appliances, set-top boxes communicatively coupled to a display device, game consoles), televisions or other display devices, automotive computers, and the like.

[0133] The computer device 1100 may include at least one processor 1102, memory 1104, communication interface(s) 1106, a physical intrinsic system 1108, input / output (I / O) devices 1110, and one or more mass storage devices 1112, all capable of communicating with one another, such as via a system bus 1114 or other appropriate connection.

[0134] The processor 1102 may be a single processing unit or multiple processing units, all of which may include a single or multiple computing units or multiple cores. The processor 1102 may be implemented as one or more microprocessors, microcomputers, microcontrollers, digital signal processors, central processing units, state machines, logic circuits, and / or any device that manipulates signals based on operational instructions. Among other capabilities, the processor 1102 may be configured to retrieve and execute computer-readable instructions stored in the memory 1104, mass storage device 1112, or other computer-readable media, such as program code for an operating system 1116, program code for application programs 1118, program code for other programs 1120, and the like.

[0135] Memory 1104 and mass storage device 1112 are examples of computer-readable storage media for storing instructions that are executed by processor 1102 to implement the various functions described above. For example, memory 1104 may generally include both volatile memory and non-volatile memory (e.g., RAM, ROM, etc.). In addition, mass storage device 1112 may generally include a hard drive, a solid-state drive, removable media, including external and removable drives, memory cards, flash memory, floppy disks, optical disks (e.g., CDs, DVDs), storage arrays, network attached storage, storage area networks, etc. Memory 1104 and mass storage device 1112 may be collectively referred to herein as memory or computer-readable storage media, and may be non-transitory media capable of storing computer-readable, processor-executable program instructions as computer program code that may be executed by processor 1102 as a specific machine configured to implement the operations and functions described in the examples herein.

[0136] A number of programs may be stored on mass storage device 1112. These programs include an operating system 1116, one or more application programs 1118, other programs 1120, and program data 1122, and may be loaded into memory 1104 for execution.

[0137] Although Figure 61100, but modules 1116, 1118, 1120, and 1122, or portions thereof, may be implemented using any form of computer-readable media accessible by the computer device 1100. As used herein, "computer-readable media" includes at least two types of computer-readable media, namely, computer-readable storage media and communication media.

[0138] Computer-readable storage media include volatile and non-volatile, removable and non-removable media implemented by any method or technology for storing information, such as computer-readable instructions, data structures, program modules or other data. Computer-readable storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disk (DVD), or other optical storage devices, magnetic cassettes, magnetic tape, magnetic disk storage devices or other magnetic storage devices, or any other non-transmission media that can be used to store information for access by a computer device. In contrast, communication media can embody computer-readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transmission mechanism. Computer-readable storage media as defined herein does not include communication media.

[0139] One or more communication interfaces 1106 are used to exchange data with other devices, such as via a network, direct connection, and the like. Such communication interfaces can be one or more of the following: any type of network interface (e.g., a network interface card (NIC)), a wired or wireless interface (such as an IEEE 802.6 wireless LAN (WLAN)), a Worldwide Interoperability for Microwave Access (Wi-MAX), an Ethernet interface, a Universal Serial Bus (USB), a cellular network interface, a Bluetooth™ interface, a Near Field Communication (NFC) interface, and the like. The communication interface 1106 can facilitate communication within a variety of network and protocol types, including wired networks (e.g., LAN, cable, and the like) and wireless networks (e.g., WLAN, cellular, satellite, and the like), the Internet, and the like. The communication interface 1106 can also provide communication with external storage devices (not shown), such as storage arrays, network attached storage, storage area networks, and the like.

[0140] In some examples, the physical intrinsic system 1108 can be a physical (or chemical, biological) device that is a heterogeneous / homogeneous integration of multiple materials in a specific structure, stacked in multiple layers, and exhibits a static or time-varying response function to various inputs / outputs, thereby exhibiting specific response characteristics. This response characteristic can be adjustable, thereby achieving a parameter update effect similar to that of a neural network. A digital signal can be converted into a physical signal by a signal conversion device for input into the physical intrinsic system 1108, or a physical signal output by the physical intrinsic system 1108 can be converted into a digital signal and then connected to the bus.

[0141] In some examples, I / O device 1110 can be a device that receives various inputs from a user and provides various outputs to the user, and can include a touch input device, a gesture input device, a camera, a keyboard, a remote control, a mouse, a printer, a display device, an audio input / output device, and the like.

[0142] The technology described herein can be supported by these various configurations of the computer device 1100 and is not limited to the specific examples of the technology described herein. For example, the functionality can also be implemented in whole or in part on a "cloud" by using a distributed system. The cloud includes and / or represents a platform for resources. The platform abstracts the underlying functionality of the hardware (e.g., servers) and software resources of the cloud. Resources can include applications and / or data that can be used when performing computing processing on a server away from the computer device 1100. Resources can also include services provided over the Internet and / or through a subscriber network such as a cellular or Wi-Fi network. The platform can abstract resources and functionality to connect the computer device 1100 to other computer devices. Therefore, the implementation of the functionality described herein can be distributed throughout the cloud. For example, functionality can be implemented partially on the computer device 1100 and partially through a platform that abstracts the functionality of the cloud.

[0143] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved. This is not limited herein.

[0144] Although the embodiments or examples of the present disclosure have been described with reference to the accompanying drawings, it should be understood that the above-mentioned methods, systems and devices are merely exemplary embodiments or examples, and the scope of the present invention is not limited by these embodiments or examples, but is only limited by the claims after authorization and their equivalents. Various elements in the embodiments or examples may be omitted or replaced by their equivalents. In addition, the steps may be performed in an order different from that described in this disclosure. Further, the various elements in the embodiments or examples may be combined in various ways. It is important that as technology evolves, many of the elements described herein may be replaced by equivalent elements that appear after this disclosure.

Claims

1. A method for realizing a physical intrinsic neural network, wherein: A multi-layer neural network includes a first network layer implemented by a first physical eigensystem, wherein the first physical eigensystem is configured to receive a physical input signal of the first network layer and output a physical output signal of the first network layer based on a response characteristic of the first physical eigensystem. The method includes: Determining input data of the multi-layer neural network and determining a target layer to be trained in the multi-layer neural network; Inputting the input data into the multi-layer neural network and obtaining the output signal of the target layer; and Based on the output signal of the target layer, the goodness of the input data is determined, and the adjustable parameters of the target layer are adjusted based on the goodness, wherein the goodness describes the response quality of the target layer to the input data.

2. The method according to claim 1, wherein The target layer includes the first network layer, and the adjustable parameters of the target layer include at least one adjustable mechanism in the first physical eigensystem for adjusting the response characteristic, wherein the at least one adjustable mechanism includes at least one of the following: The position and / or orientation of devices in the first physical intrinsic system, the distance between devices, the magnetic field position, the magnetic field direction and / or strength, the pulse current, the voltage bias, the current strength and / or direction, the resistance, the conductance, the temperature, the pressure, the mechanical strain, the optical parameters, the acoustic parameters, the vibration frequency, the spectral characteristics of the electromagnetic signal, the chemical concentration and the biological molecule activity.

3. The method according to claim 2, wherein: The input data includes positive data and negative data, and the goodness of the input data includes positive goodness and negative goodness. The determining of the quality of the input data based on the output signal of the target layer, and adjusting the adjustable parameters of the target layer based on the quality includes: adjusting the at least one adjustable mechanism in a direction that increases the positive goodness; and / or The at least one adjustable mechanism is adjusted in a direction to reduce the negative goodness.

4. The method according to claim 2, wherein: Determining the goodness of the input data based on the output signal of the target layer, and adjusting the adjustable parameters of the target layer based on the goodness includes: Converting the physical output signal into a digital signal; The first silicon-based integrated circuit is used to perform the following operations: Calculating a first value representing the goodness based on a preset objective function and the digital signal; and Calculating a second value representing an adjustment amount for the at least one adjustable mechanism based on a preset adjustment strategy and the first value, wherein the preset adjustment strategy indicates adjusting the at least one adjustable mechanism in a direction of optimizing the goodness; and The second numerical value is converted into a first physical quantity and applied to the at least one adjustable mechanism.

5. The method according to claim 2, wherein: Determining the goodness of the input data based on the output signal of the target layer, and adjusting the adjustable parameters of the target layer based on the goodness includes: Processing the physical output signal using a first eigenmapping channel of a physical intrinsic goodness determination device to obtain a second physical quantity representing the goodness, wherein the first eigenmapping channel represents a preset objective function to implement evaluation of a response quality of the first network layer to the physical input signal in the physical domain; Processing the second physical quantity using a second eigenmapping channel of a physical intrinsic response characteristic adjustment device to obtain a third physical quantity that can be directly applied to the at least one adjustable mechanism, wherein the second eigenmapping channel represents a preset adjustment strategy to achieve determination of an adjustment amount for the at least one adjustable mechanism in the physical domain to directionally optimize the goodness of fit; and The third physical quantity is applied to the at least one adjustable mechanism.

6. The method according to claim 4, wherein: The input data includes positive data and negative data, and the goodness of the input data includes positive goodness and negative goodness. The second eigenmapping channel includes a first sub-channel representing a first preset adjustment strategy and a second sub-channel representing a second preset adjustment strategy, the first sub-channel being configured to process a physical quantity representing the positive goodness, the first preset adjustment strategy indicating adjustment of the at least one adjustable mechanism in a direction of increasing the positive goodness, the second sub-channel being configured to process a physical quantity representing the negative goodness, the second preset adjustment strategy indicating adjustment of the at least one adjustable mechanism in a direction of reducing the negative goodness.

7. The method according to any one of claims 1 to 6, wherein The physical input signal and the physical output signal are analog signals and are respectively based on at least one item selected from the group consisting of: Magnetic fields, heat, pulses, microwaves, radio frequency signals, electrical signals, optical signals, acoustic signals, mechanical signals, vibration signals, time signals, and space signals.

8. The method according to any one of claims 1 to 6, further comprising: The output signal of the target layer is input into a subsequent layer of the multi-layer neural network to perform forward propagation, wherein the forward propagation in the subsequent layer of the multi-layer neural network and the adjustment of the adjustable parameters of the target layer are performed asynchronously.

9. The method according to any one of claims 1 to 6, wherein The multi-layer neural network further includes a second network layer implemented by a second silicon-based integrated circuit, wherein the second silicon-based integrated circuit is configured to receive an electrical input signal of the second network layer and calculate an electrical output signal of the second network layer based on model parameters of the second network layer stored in the second silicon-based integrated circuit, and the adjustable parameters of the target layer include the model parameters of the second network layer. The method further comprises: using a signal conversion device to convert the electrical output signal of the second network layer into a physical input signal of the first network layer; or The signal conversion device is used to convert the physical output signal of the first network layer into the electrical input signal of the second network layer.

10. The method according to claim 9, wherein: The second silicon-based integrated circuit includes at least one item selected from the group consisting of: Field-programmable gate array (FPGA), application-specific integrated circuit (ASIC), graphics processing unit (GPU), central processing unit (CPU), neural processing unit (NPU), tensor processing unit (TPU), very large-scale integrated circuit (VLSI), complementary metal oxide semiconductor (CMOS), and integrated modules composed of transistors and / or other electrical devices arranged and wired based on deposition, lithography and / or etching.

11. The method according to claim 9, wherein The target layer includes the second network layer, wherein determining the goodness of the input data based on an output signal of the target layer, and adjusting the adjustable parameters of the target layer based on the goodness includes: The third silicon-based integrated circuit is used to perform the following operations: Calculating a third value representing the goodness based on a preset objective function and the electrical output signal of the second network layer; and Based on a preset adjustment strategy and the third value, the model parameters of the second network layer are adjusted.

12. The method according to any one of claims 1 to 6, wherein: The first network layer is used to extract nonlinear features, and the multi-layer neural network further includes a third network layer after the first network layer, wherein a second physical eigensystem different from the first physical eigensystem is used to implement at least one of the following multiple operators for the third network layer: Feature extraction, forward propagation, objective function calculation, gradient acquisition, random perturbation and parameter update.

13. The method according to any one of claims 1 to 6, wherein: The multi-layer neural network is operated with at least one item selected from the group consisting of: Deterministic algorithms, random or probabilistic algorithms, reasoning or solving algorithms, expert system-based algorithms, simple logic or addition / multiplication algorithms, algorithms in the field of combinatorial optimization or operations research, algorithms in the field of machine learning or statistical learning, and algorithms related to complex systems or deep neural networks.

14. A device for implementing a physical intrinsic neural network, wherein: The multi-layer neural network includes a first network layer implemented by a first physical eigensystem, wherein the first physical eigensystem is configured to receive a physical input signal of the first network layer and output a physical output signal of the first network layer based on a response characteristic of the first physical eigensystem. The apparatus includes: a first determining unit configured to determine input data of the multi-layer neural network and determine a target layer to be trained in the multi-layer neural network; an acquisition unit configured to input the input data into the multi-layer neural network and acquire an output signal of the target layer; and A determination unit is configured to determine the goodness of the input data based on the output signal of the target layer, and adjust the adjustable parameters of the target layer based on the goodness, wherein the goodness describes the response quality of the target layer to the input data.

15. The device according to claim 14, wherein The target layer includes the first network layer, and the adjustable parameters of the target layer include at least one adjustable mechanism in the first physical eigensystem for adjusting the response characteristic, wherein the at least one adjustable mechanism includes at least one of the following: The position and / or orientation of devices in the first physical intrinsic system, the distance between devices, the magnetic field position, the magnetic field direction and / or strength, the pulse current, the voltage bias, the current strength and / or direction, the resistance, the conductance, the temperature, the pressure, the mechanical strain, the optical parameters, the acoustic parameters, the vibration frequency, the spectral characteristics of the electromagnetic signal, the chemical concentration and the biological molecule activity.

16. The device according to claim 15, wherein The input data includes positive data and negative data, and the goodness of the input data includes positive goodness and negative goodness. Wherein, the determining unit is configured to: adjusting the at least one adjustable mechanism in a direction that increases the positive goodness; and / or The at least one adjustable mechanism is adjusted in a direction to reduce the negative goodness.

17. The device according to claim 15, wherein The determining unit includes: a first conversion subunit, configured to convert the physical output signal into a digital signal; A first silicon-based integrated circuit is configured to perform the following operations: Calculating a first value representing the goodness based on a preset objective function and the digital signal; and Calculating a second value representing an adjustment amount for the at least one adjustable mechanism based on a preset adjustment strategy and the first value, wherein the preset adjustment strategy indicates adjusting the at least one adjustable mechanism in a direction of optimizing the goodness; and The second conversion subunit is configured to convert the second numerical value into a first physical quantity and apply the first physical quantity to the at least one adjustable mechanism.

18. The device according to claim 15, wherein The determining unit includes: a physical intrinsic goodness determination device, configured to process the physical output signal using a first intrinsic mapping channel therein to obtain a second physical quantity representing the goodness, wherein the first intrinsic mapping channel represents a preset objective function to achieve an evaluation of a response quality of the first network layer to the physical input signal in the physical domain; a physical intrinsic response characteristic adjustment device, configured to process the second physical quantity using a second intrinsic mapping channel therein to obtain a third physical quantity that can be directly applied to the at least one adjustable mechanism, wherein the second intrinsic mapping channel represents a preset adjustment strategy to achieve determination of an adjustment amount for the at least one adjustable mechanism in the physical domain to directionally optimize the goodness; and The regulating subunit is configured to apply the third physical quantity to the at least one adjustable mechanism.

19. The device according to claim 18, wherein The input data includes positive data and negative data, and the goodness of the input data includes positive goodness and negative goodness. The second eigenmapping channel includes a first sub-channel representing a first preset adjustment strategy and a second sub-channel representing a second preset adjustment strategy, the first sub-channel being configured to process a physical quantity representing the positive goodness, the first preset adjustment strategy indicating adjustment of the at least one adjustable mechanism in a direction of increasing the positive goodness, the second sub-channel being configured to process a physical quantity representing the negative goodness, the second preset adjustment strategy indicating adjustment of the at least one adjustable mechanism in a direction of reducing the negative goodness.

20. The device according to any one of claims 14 to 19, wherein The physical input signal and the physical output signal are analog signals and are respectively based on at least one item selected from the group consisting of: Magnetic fields, heat, pulses, microwaves, radio frequency signals, electrical signals, optical signals, acoustic signals, mechanical signals, vibration signals, time signals, and space signals.

21. The apparatus according to any one of claims 14 to 19, further comprising: A forward propagation unit is configured to input the output signal of the target layer into a subsequent layer of the multi-layer neural network to perform forward propagation, wherein the forward propagation in the subsequent layer of the multi-layer neural network and the adjustment of the adjustable parameters of the target layer are performed asynchronously.

22. The device according to any one of claims 14 to 19, wherein: The multi-layer neural network further includes a second network layer implemented by a second silicon-based integrated circuit, wherein the second silicon-based integrated circuit is configured to receive an electrical input signal of the second network layer and calculate an electrical output signal of the second network layer based on model parameters of the second network layer stored in the second silicon-based integrated circuit, and the adjustable parameters of the target layer include the model parameters of the second network layer. The device further includes a signal conversion device configured to: converting the electrical output signal of the second network layer into a physical input signal of the first network layer; or The physical output signal of the first network layer is converted into an electrical input signal of the second network layer.

23. The device according to claim 22, wherein The second silicon-based integrated circuit includes at least one item selected from the group consisting of: Field-programmable gate array (FPGA), application-specific integrated circuit (ASIC), graphics processing unit (GPU), central processing unit (CPU), neural processing unit (NPU), tensor processing unit (TPU), very large-scale integrated circuit (VLSI), complementary metal oxide semiconductor (CMOS), and integrated modules composed of transistors and / or other electrical devices arranged and wired based on deposition, lithography and / or etching.

24. The apparatus according to claim 22, wherein The target layer includes the second network layer, wherein the determining unit includes: The third silicon-based integrated circuit is configured to perform the following operations: Calculating a third value representing the goodness based on a preset objective function and the electrical output signal of the second network layer; and Based on a preset adjustment strategy and the third value, the model parameters of the second network layer are adjusted.

25. The device according to any one of claims 14 to 19, wherein The first network layer is used to extract nonlinear features, and the multi-layer neural network further includes a third network layer after the first network layer, wherein a second physical eigensystem different from the first physical eigensystem is used to implement at least one of the following multiple operators for the third network layer: Feature extraction, forward propagation, objective function calculation, gradient acquisition, random perturbation and parameter update.

26. The device according to any one of claims 14 to 19, wherein The multi-layer neural network is operated with at least one item selected from the group consisting of: Deterministic algorithms, random or probabilistic algorithms, reasoning or solving algorithms, expert system-based algorithms, simple logic or addition / multiplication algorithms, algorithms in the field of combinatorial optimization or operations research, algorithms in the field of machine learning or statistical learning, and algorithms related to complex systems or deep neural networks.

27. A system for implementing a multi-layer neural network, the system comprising: A first physical eigensystem is configured to: Implementing a first network layer in the multi-layer neural network; as well as receiving a physical input signal of the first network layer, and outputting a physical output signal of the first network layer based on a response characteristic of the first physical eigensystem; as well as at least one of the first control unit and the second control unit, Wherein, the first control unit includes: at least one processor; and a memory communicatively coupled to the at least one processor, The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor and the first physical eigensystem to collaboratively perform at least one step of the method according to any one of claims 1 to 13. Wherein, the second control unit includes: The third eigenmapping channel of physical eigenvalues, The third eigenmapping channel and the first physical eigensystem can collaboratively perform at least one step of the method according to any one of claims 1 to 13 in the physical domain.

28. A computer-readable storage medium having stored thereon a computer program for implementing a multi-layer neural network, the multi-layer neural network comprising a first network layer implemented by a first physical eigensystem, the first physical eigensystem being configured to receive a physical input signal of the first network layer and output a physical output signal of the first network layer based on a response characteristic of the first physical eigensystem. in, When the computer program is executed by a processor, the processor and the first physical eigensystem are enabled to collaboratively perform the method according to any one of claims 1 to 13.

29. A computer program product, comprising a computer program for implementing a multi-layer neural network, the multi-layer neural network comprising a first network layer implemented by a first physical eigensystem, the first physical eigensystem being configured to receive a physical input signal of the first network layer and output a physical output signal of the first network layer based on a response characteristic of the first physical eigensystem. in, When the computer program is executed by a processor, the processor and the first physical eigensystem are enabled to collaboratively perform the method according to any one of claims 1 to 13.

Images