A method, apparatus, device and signal equalizer based on reservoir computing

Abstract

The present application provides a method, apparatus, device, and signal equalizer based on reservoir computing. The method includes: when a predetermined trigger condition is satisfied, generating neural network parameter information corresponding to a second communication device according to channel condition information reported by the second communication device, where the neural network parameter information includes a jump size corresponding to a reservoir; and sending the neural network parameter information to the second communication device. The present application proposes the concept of DRJ. The second communication device can configure DRJ based on the jump size allocated by the first communication device for it, so as to implement a signal equalizer based on DRJ, which can support adaptive equalization of a high-speed communication system in an easy-to-train and low-power consumption manner. Due to the structural characteristics brought by its jump-carrying deterministic structure and relatively simple linear regression training characteristics, it can reduce the complexity and power consumption of the second communication device and enable it to support a simpler 、 Lower power consumption and faster hardware implementation.

Description

Technical Field

[0001] This application relates to the field of communication technologies, and in particular, to a method, apparatus, device, and signal equalizer based on reservoir computing (RC) with low complexity, low power consumption, and higher speed. Background Art

[0002] To meet the requirements of emerging applications, such as fixed-mobile convergence in fifth-generation or later mobile networks, high-speed access networks are widely used in various communication systems (such as passive optical network PON, wireless local area network WiFi, and cellular communication networks, etc.). The importance of equalization of high-speed signals cannot be overemphasized, because such software-based algorithms can essentially make full use of cost-effective hardware devices. In existing communication systems, traditional neural networks (NNs), such as fully-connected neural networks (FCNNs) and recurrent neural networks (RNNs), are used to achieve signal equalization. However, traditional neural networks require an energy-consuming and time-consuming training phase to correctly equalize the channel, because their training requires backpropagation through the entire network, and thus they do not have self-adaptability. Summary of the Invention

[0003] The objective of this application is to provide a method, apparatus, device, and signal equalizer based on reservoir computing, which have a simpler, lower-power-consumption, and higher-speed hardware implementation.

[0004] According to one aspect of this application, a method based on reservoir computing is provided, which is applied to a first communication device. Wherein, the method includes:

[0005] When a predetermined trigger condition is satisfied, generate neural network parameter information corresponding to the second communication device according to the channel condition information reported by the second communication device, where the neural network parameter information includes the jump size corresponding to the reservoir;

[0006] Send the neural network parameter information to the second communication device.

[0007] According to another aspect of this application, a method based on reservoir computing is provided, which is applied to a second communication device. Wherein, the method includes:

[0008] Receive the neural network parameter information sent by the first communication device, where the neural network parameter information includes the jump size corresponding to the reservoir;

[0009] Configure a neural network according to the neural network parameter information, where the neural network includes an input layer, a Deterministic Reservoir, and a trainable readout layer, and configure the Deterministic Reservoir based on the jump size;

[0010] Train the readout layer by using simple linear regression to obtain a trained signal equalizer.

[0011] According to another aspect of the present application, there is provided a first apparatus based on reservoir computing in a first communication device, where the first apparatus includes:

[0012] An apparatus for generating neural network parameter information corresponding to the second communication device according to channel condition information reported by the second communication device when a predetermined trigger condition is satisfied, where the neural network parameter information includes a jump size corresponding to the reservoir;

[0013] An apparatus for sending the neural network parameter information to the second communication device.

[0014] According to another aspect of the present application, there is provided a second apparatus based on reservoir computing in a second communication device, where the second apparatus includes:

[0015] An apparatus for receiving neural network parameter information sent by the first communication device, where the neural network parameter information includes a jump size corresponding to the reservoir;

[0016] An apparatus for configuring a neural network according to the neural network parameter information, where the neural network includes an input layer, a Deterministic Reservoir, and a trainable readout layer, and configure the Deterministic Reservoir based on the jump size;

[0017] An apparatus for training the readout layer by using simple linear regression to obtain a trained signal equalizer.

[0018] According to another aspect of the present application, there is provided a signal equalizer based on reservoir computing, where the signal equalizer includes a reservoir module and a readout module, and the reservoir module includes a Deterministic Reservoir, and the Deterministic Reservoir is configured by using a jump size allocated by the first communication device for the second communication device.

[0019] According to another aspect of the present application, there is provided a first communication device, where the first communication device includes:

[0020] A memory for storing one or more programs;

[0021] One or more processors connected to the memory,

[0022] When the one or more programs are executed by the one or more processors, the one or more processors are caused to perform the following operations:

[0023] When a predetermined trigger condition is satisfied, neural network parameter information corresponding to the second communication device is generated according to channel condition information reported by the second communication device, wherein the neural network parameter information includes a jump size corresponding to a reservoir;

[0024] The neural network parameter information is sent to the second communication device.

[0025] According to another aspect of the present application, a second communication device is provided, wherein the second communication device includes:

[0026] A memory for storing one or more programs;

[0027] One or more processors connected to the memory,

[0028] When the one or more programs are executed by the one or more processors, the one or more processors are caused to perform the following operations:

[0029] Receive neural network parameter information sent by a first communication device, wherein the neural network parameter information includes a jump size corresponding to a reservoir;

[0030] Configure a neural network according to the neural network parameter information, wherein the neural network includes an input layer, a deterministic reservoir, and a trainable readout layer, and configure the deterministic reservoir based on the jump size;

[0031] Train the readout layer by using simple linear regression to obtain a trained signal equalizer.

[0032] Compared with the prior art, the present application has the following advantages: The concept of DRJ (Deterministic Reservoir with Jumps) is proposed. The second communication device can configure DRJ based on the jump size allocated by the first communication device for it, so as to implement a signal equalizer based on DRJ, which can support the adaptive equalization of a high-speed communication system in an easy-to-train and low-power consumption manner. And due to the structural characteristics and relatively simple linear regression training characteristics brought by its deterministic structure with jumps, it can reduce the complexity and power consumption of the second communication device and make it support more simplicity 、Lower power consumption and higher speed hardware implementations (such as passive integrated photonic chips, System on Chip (SoC), or Field-Programmable Gate Array (FPGA)); and, by using the allocated jump size and a smaller reservoir size, the solution of this application has better BER performance compared with the prior art. For example, through experiments, for [17 / 20i, (50,1 / 4)] adopted in this application compared with [17 / 20i, (500,1 / 4)] adopted by the existing RC, the reservoir size is reduced by 90%, but the performance is better than that of the existing RC; in addition, this application can also support a wide range of applications, such as speech recognition, robot control, time series prediction, image processing, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Other features, objects, and advantages of this application will become more apparent by reading the detailed description of the non-limiting embodiments with reference to the following drawings:

[0034] Figure 1 The flowchart shows the method of reservoir computing in the first communication device according to an embodiment of this application;

[0035] Figure 2 The flowchart shows the method of reservoir computing in the second communication device according to an embodiment of this application;

[0036] Figure 3 The diagram shows the first device of reservoir computing in the first communication device according to an embodiment of this application;

[0037] Figure 4 The diagram shows the second device of reservoir computing in the second communication device according to an embodiment of this application;

[0038] Figure 5 The diagram shows the schematic diagram of establishing a signal equalizer based on DRJ according to an example of this application;

[0039] Figure 6 The flowchart shows the process of establishing a signal equalizer based on DRJ according to an example of this application;

[0040] Figure 7 The diagram shows a single-output architecture according to an example of this application;

[0041] Figure 8 The diagram shows a parallel-output architecture according to an example of this application;

[0042] Figure 9 The diagram shows the experimental setup of 100Gbps / λPAM4-based IMDD PON according to an example of this application;

[0043] Figure 10 shows the BER corresponding to different skip sizes tested based on Figure 7 the single-output architecture shown;

[0044] Figure 11 is the test result of Experiment 1 (single output) (the solid line is DRJ and the dashed line is RC);

[0045] Figure 12 is the test result of Experiment 2 (parallel output) (the solid line is DRJ and the dashed line is RC);

[0046] Figure 13 shows a schematic diagram of a DRJ-based signal equalizer according to an example of the present application;

[0047] Figure 14 shows a schematic diagram of a DRJ-based signal equalizer according to another example of the present application;

[0048] Figure 15 shows an exemplary system that can be used to implement the various embodiments described in the present application.

[0049] Identical or similar reference numerals in the drawings represent identical or similar components. Detailed Description of the Invention

[0050] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the operations as sequential processes, many of the operations can be implemented in parallel, concurrently, or simultaneously. In addition, the order of the operations can be rearranged. The process can be terminated when its operations are completed, but it can also have additional steps not included in the drawings. The process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0051] As used herein, a "device" refers to an intelligent electronic device that can perform a predetermined processing process such as numerical calculation and / or logical calculation by running a predetermined program or instruction. It can include a processor and a memory, and the processor executes the program instructions pre-stored in the memory to perform the predetermined processing process, or a predetermined processing process is performed by a hardware such as an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), a digital signal processor (DSP), or a combination of the above two.

[0052] The methods discussed later in this document, some of which are illustrated by flowcharts, can be implemented by hardware, software, firmware, middleware, microcode, hardware description language, or any combination thereof. When implemented in software, firmware, middleware, or microcode, the program code or code segments for performing the necessary tasks can be stored in a machine or computer-readable medium, such as a storage medium. One or more processors can perform the necessary tasks.

[0053] The specific structural and functional details disclosed herein are merely representative and are for the purpose of describing exemplary embodiments of the present application. However, the present application can be embodied in many alternative forms and should not be construed as being limited only to the embodiments set forth herein.

[0054] It should be understood that although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are only used to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly, the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the associated listed items.

[0055] The terms used herein are for the purpose of describing particular embodiments only and are not intended to limit the exemplary embodiments. Unless the context clearly dictates otherwise, the singular forms "a", "an" used herein are also intended to include the plural. It should also be understood that the terms "comprises" and / or "comprising" specify the presence of the stated features, integers, steps, operations, units, and / or components, and do not preclude the presence or addition of one or more other features, integers, steps, operations, units, components, and / or combinations thereof.

[0056] It should also be noted that in some alternative implementations, the functions / actions mentioned may occur in a different order than that indicated in the figures. For example, depending on the functions / actions involved, two consecutively shown figures may actually be executed substantially simultaneously or sometimes in the reverse order.

[0057] This application discovers that, compared with traditional neural networks (NNs), a recently proposed reservoir computing (RC) based on silicon photonics has been proposed to avoid training the hidden part (i.e., the reservoir) of the NN. RC has the advantages of being easy to train, low power consumption, and high speed, which are beneficial to photonic applications. The hardware implementation of photonics-based RC allows for making full use of the advantages of light (low power consumption, high bandwidth, etc.) for computing. As an emerging technology, the RC-based non-linear equalizer has been regarded as a powerful adaptive equalization tool for compensating transmission impairments in the field of optical communication, including linear and non-linear distortions. However, although existing RCs allow avoiding training the hidden layer of the NN, they require a random reservoir matrix W res and a large reservoir size. This application also discovers that even with a medium reservoir size, existing RCs based on random reservoir matrices cannot meet the bit error rate (BER) requirements regarding the complexity of the second communication device in high-speed communication systems. To address the above technical problems, this application aims to propose a new low-complexity technical solution for establishing a signal equalizer based on deterministic RC. The signal equalizer based on deterministic RC can utilize the equalization performance and reduce the complexity of the second communication device.

[0058] First of all, it should be noted that this application first proposes the concept of a deterministic reservoir with jumps (DRJ). The deterministic reservoir with jumps is configured using a determined jump size configured by the first communication device for the second communication device. The deterministic reservoir includes all two-way jumps based on this jump size, and all two-way jumps share the same weight. This application can establish a new signal equalizer for high-speed communication based on the low-complexity DRJ. In some embodiments, the first communication device includes, but is not limited to, an optical line terminal (OLT), a base station (BS), etc. In some embodiments, the second communication device includes, but is not limited to, an optical network unit (ONU), a user equipment (UE), etc. It should be noted that the first communication device and the second communication device in the above embodiments are only examples and not limitations to this application.

[0059] The following further describes this application in detail with reference to the accompanying drawings.

[0060] Figure 1The flowchart shows a method based on reservoir computing in a first communication device according to an embodiment of the present application. The method includes step S11 and step S12. In step S11, when a predetermined trigger condition is satisfied, the first communication device generates neural network parameter information corresponding to the second communication device according to the channel condition information reported by the second communication device, where the neural network parameter information includes the jump size corresponding to the reservoir. In step S12, the first communication device sends the neural network parameter information to the second communication device.

[0061] In step S11, when a predetermined trigger condition is satisfied, the first communication device generates neural network parameter information corresponding to the second communication device according to the channel condition information reported by the second communication device, where the neural network parameter information includes the jump size corresponding to the reservoir.

[0062] In some embodiments, the predetermined trigger condition includes any condition for triggering the first communication device to start the process of establishing a signal equalizer. In some embodiments, the predetermined trigger condition includes any one of the following: the second communication device is newly connected; the channel condition of the second communication device has changed. As an example, the first communication device continuously monitors the status of the second communication device. When it is found that a second communication device is newly connected and has never been seen by the first communication device before, it is determined that the predetermined trigger condition is satisfied. It should be noted that it can be set that the predetermined trigger condition is satisfied only when the second communication device is newly connected and has never been seen by the first communication device, or it can be set that the predetermined trigger condition is satisfied as long as the second communication device is newly connected (including the case where the second communication device that has been connected to the first communication device reconnects to the first communication device). As another example, whenever the first communication device receives the channel condition information sent by the second communication device, it determines whether the channel condition of the second communication device has changed (or whether the change exceeds the constraint range). If so, it is determined that the predetermined trigger condition is satisfied. In some embodiments, the first communication device re-determines the new jump size according to the latest channel condition information reported by the second communication device at regular time intervals. If the new jump size is the same as the jump size determined last time, no further operation is performed. If they are different, it is determined that the predetermined trigger condition is satisfied (that is, a new jump size needs to be re-allocated for the second communication device), and then the first communication device generates neural network parameter information corresponding to the second communication device based on the new jump size.

[0063] In some embodiments, the jump size is greater than or equal to 2. In some embodiments, the neural network parameter information not only includes the jump size corresponding to the reservoir, but also includes other parameter information other than the jump size, such as the number of neural network layers, the number of neurons in each layer, etc. In some embodiments, the first communication device determines the jump size corresponding to the reservoir and the values of other neural network parameters based on the channel condition information reported by the second communication device, such as CSI (Channel State Information), and generates corresponding neural network parameter information. As an example, the first communication device generates neural network parameter information including the jump size corresponding to the reservoir, the number of neural network layers, and the number of neurons in each layer based on the CSI reported by the second communication device.

[0064] In step S12, the first communication device sends the neural network parameter information to the second communication device. As an example, the newly connected second communication device reports CSI to the first communication device. The first communication device determines that a predetermined trigger condition is met and generates neural network parameter information NN_param based on the CSI. NN_param includes Jump_size (jump size) and other parameter information. After that, the first communication device sends NN_param to the second communication device so that the second communication device can establish a DRJ-based signal equalizer based on the NN_param.

[0065] In some embodiments, the method further includes step S13 and step S14. In step S13, the first communication device receives the ready signal sent by the second communication device after the training is completed; in step S14, the first communication device sends load data to the second communication device. In some embodiments, after the second communication device completes the configuration and training of the neural network (only the readout layer of the neural network needs to be trained in this application), it sends a ready signal to the first communication device. After receiving the ready signal, the first communication device transmits load data to the second communication device. After receiving the load data, the second communication device will use the trained signal equalizer to achieve equalization.

[0066] Figure 2The flowchart shows a method based on reservoir computing in a second communication device according to an embodiment of the present application. The method includes step S21, step S22, and step S23. In step S21, the second communication device receives neural network parameter information sent by the first communication device, where the neural network parameter information includes the jump size corresponding to the reservoir. In step S22, the second communication device configures a neural network according to the neural network parameter information, where the neural network includes an input layer, a deterministic reservoir, and a trainable readout layer, and configures the deterministic reservoir based on the jump size. In step S23, the second communication device trains the readout layer by using simple linear regression to obtain a trained signal equalizer.

[0067] In step S21, the second communication device receives neural network parameter information sent by the first communication device, where the neural network parameter information includes the jump size corresponding to the reservoir. The relevant content of the neural network parameter information has been described in detail in the foregoing embodiment and will not be elaborated herein.

[0068] In step S22, the second communication device configures a neural network according to the neural network parameter information, where the neural network includes an input layer, a deterministic reservoir, and a trainable readout layer, and configures the deterministic reservoir based on the jump size. It should be noted that only the readout layer in the neural network needs to be trained.

[0069] In some embodiments, the deterministic reservoir may be one or more. In some embodiments, the trainable readout layer may be single-output or parallel-output.

[0070] Since the jump size configured by the first communication device for the second communication device is determined, the reservoir configured by the second communication device based on this jump size is also determined, that is, the reservoir matrix used for reservoir computing is determined. In some embodiments, the non-zero elements in the matrix corresponding to the deterministic reservoir (that is, the reservoir matrix used for reservoir computing) include all recurrent connections and all bidirectional jumps. All recurrent connections share the first weight, all bidirectional jumps share the second weight, and the remaining elements of the matrix are all set to zero; in some embodiments, the remaining elements of this matrix are all set to zero.

[0071] In step S23, the second communication device trains the readout layer by using simple linear regression to obtain a trained signal equalizer. In some embodiments, the readout layer is trained by simple linear regression at high speed and low power consumption to obtain a trained DRJ-based signal equalizer.

[0072] It should be noted that, due to limited computing power and power consumption, the training of the traditional NN in the second communication device is preferably implemented on the first communication device side. Therefore, it is necessary to upload additional pilots to the first communication device side before training and allocate training parameters (such as weights) to specific second communication devices after training. In this application, since simple linear regression is used to train the readout layer, the newly proposed DRJ-based signal equalizer can be easily trained at high speed and low power at the second communication device end.

[0073] In some embodiments, the method further includes step S24 and step S25. In step S24, the second communication device sends a ready signal to the first communication device after the training is completed; in step S25, the second communication device receives the load data from the first communication device and uses the signal equalizer to achieve load balancing. In some embodiments, after the training of the readout layer is completed, the second communication device sends a ready signal to the first communication device. After receiving the ready signal, the first communication device transmits the load data to the second communication device, and the second communication device receives the load data from the first communication device and uses the signal equalizer to achieve load balancing.

[0074] Figure 3 A first device based on reservoir computing in the first communication device according to an embodiment of the present application is shown. The first device 1 includes a generating device 11 and a first sending device 12. The generating device 11 is configured to generate neural network parameter information corresponding to the second communication device according to the channel condition information reported by the second communication device when a predetermined trigger condition is satisfied, wherein the neural network parameter information includes the jump size corresponding to the reservoir; the first sending device 12 is configured to send the neural network parameter information to the second communication device.

[0075] When a predetermined trigger condition is satisfied, the generating device 11 in the first communication device generates neural network parameter information corresponding to the second communication device according to the channel condition information reported by the second communication device, wherein the neural network parameter information includes the jump size corresponding to the reservoir.

[0076] In some embodiments, the predetermined triggering condition includes any condition for triggering the first communication device to start the process of establishing a signal equalizer. In some embodiments, the predetermined triggering condition includes any one of the following: the second communication device is newly connected; the channel condition of the second communication device changes. As an example, the first communication device continuously monitors the status of the second communication device. When it is found that a second communication device is newly connected and the first communication device has never seen it before, it is determined that the predetermined triggering condition is satisfied. It should be noted that it can be set that the predetermined triggering condition is satisfied only when the second communication device is newly connected and the first communication device has never seen it before, or it can be set that the predetermined triggering condition is satisfied as long as the second communication device is newly connected (including the case where the second communication device that the first communication device has connected to before reconnects to the first communication device). As another example, whenever the channel condition information sent by the second communication device is received, the generating device 11 determines whether the channel condition of the second communication device has changed (or whether the change exceeds the constraint range). If so, it is determined that the predetermined triggering condition is satisfied. In some embodiments, the generating device 11 re-determines the new jump size according to the channel condition information reported by the second communication device at the latest every predetermined time interval. If the new jump size is the same as the jump size determined last time, the operation is not continued. If they are different, it is determined that the predetermined triggering condition is satisfied (that is, a new jump size needs to be re-allocated for the second communication device), and then the generating device 11 generates the neural network parameter information corresponding to the second communication device based on the new jump size.

[0077] In some embodiments, the jump size is greater than or equal to 2. In some embodiments, the neural network parameter information includes not only the jump size corresponding to the reservoir but also other parameter information other than the jump size, such as the number of neural network layers and the number of neurons in each layer. In some embodiments, the generating device 11 determines the jump size corresponding to the reservoir and the values of other neural network parameters based on the channel condition information reported by the second communication device, such as CSI (Channel State Information), and generates the corresponding neural network parameter information. As an example, the generating device 11 generates neural network parameter information including the jump size corresponding to the reservoir, the number of neural network layers, and the number of neurons in each layer based on the CSI reported by the second communication device.

[0078] The first transmitting device 12 transmits the neural network parameter information to the second communication device. As an example, a newly connected second communication device reports CSI to the first communication device. The generating device in the first communication device determines that a predetermined trigger condition is met, and generates neural network parameter information NN_param based on the CSI. NN_param includes Jump_size (jump size) and other parameter information. Then, the first transmitting device 12 in the first communication device transmits NN_param to the second communication device, so that the second communication device establishes a DRJ-based signal equalizer based on NN_param.

[0079] In some embodiments, the first device 1 further includes a first receiving device (not shown in the figure) and a second transmitting device (not shown in the figure). The first receiving device receives the ready signal sent by the second communication device; the second transmitting device transmits load data to the second communication device. In some embodiments, after the second communication device completes the configuration and training of the neural network (only the readout layer of the neural network needs to be trained in this application), it sends a ready signal to the first communication device. After the first receiving device in the first communication device receives the ready signal, it triggers the second transmitting device to transmit load data to the second communication device. After receiving the load data, the second communication device uses the trained signal equalizer to achieve equalization.

[0080] Figure 4 A second device based on reservoir computing in the second communication device according to an embodiment of the present application is shown. The second device includes a second receiving device 21, a configuration device 22, and a training device 23. The second receiving device 21 is configured to receive the neural network parameter information sent by the first communication device, where the neural network parameter information includes the jump size corresponding to the reservoir; the configuration device 22 is configured to configure a neural network according to the neural network parameter information, where the neural network includes an input layer, a deterministic reservoir, and a trainable readout layer, and configure the deterministic reservoir based on the jump size; the training device 23 is configured to train the readout layer by using simple linear regression to obtain a trained signal equalizer.

[0081] The second receiving device 21 receives the neural network parameter information sent by the first communication device, where the neural network parameter information includes the jump size corresponding to the reservoir. The relevant content of the neural network parameter information has been described in detail in the foregoing embodiments and will not be repeated here.

[0082] The configuration device 22 configures a neural network according to the neural network parameter information, where the neural network includes an input layer, a deterministic reservoir, and a trainable readout layer, and configures the deterministic reservoir based on the jump size. It should be noted that only the readout layer in the neural network needs to be trained.

[0083] In some embodiments, the deterministic reservoir pool may be one or more. In some embodiments, the trainable readout layer may be single-output or parallel-output.

[0084] Since the hop size configured by the first communication device for the second communication device is determined, the reservoir pool configured by the second communication device based on the hop size is also determined, that is, the reservoir pool matrix for reservoir pool calculation is determined. In some embodiments, the non-zero elements in the matrix corresponding to the deterministic reservoir pool (that is, the reservoir pool matrix for reservoir pool calculation) include all recurrent connections and all bidirectional hops. All recurrent connections share a first weight, all bidirectional hops share a second weight, and the remaining elements of the matrix are all set to zero; in some embodiments, the remaining elements of this matrix are all set to zero.

[0085] The training device 23 trains the readout layer by using simple linear regression to obtain a trained signal equalizer. In some embodiments, the readout layer is trained by simple linear regression at high speed and low power to obtain a trained DRJ-based signal equalizer.

[0086] It should be noted that due to limited computing power and power consumption, the training of the traditional NN in the second communication device is preferably implemented on the first communication device side. Therefore, additional pilots need to be uploaded to the first communication device side before training, and training parameters (such as weights) are allocated to specific second communication devices after training. In this application, since the readout layer is trained by using simple linear regression, the newly proposed DRJ-based signal equalizer can be easily trained at high speed and low power at the second communication device end.

[0087] In some embodiments, the second device 2 further includes a third sending device (not shown in the figure) and a third receiving device (not shown in the figure). The third sending device is used to send a ready signal to the first communication device; the third receiving device is used to receive load data from the first communication device and implement load balancing by using the signal equalizer. In some embodiments, after the training of the readout layer is completed, the third sending device sends a ready signal to the first communication device. After receiving the ready signal, the first communication device transmits load data to the second communication device, and the third receiving device receives the load data from the first communication device and implements load balancing by using the signal equalizer.

[0088] Figure 5 The figure shows a schematic diagram of establishing a signal equalizer based on DRJ according to an example of the present application. In this example, the first communication device is an OLT, and the second communication device is an ONU. As Figure 5 shown, the first communication device is connected to multiple second communication devices. If there is a new second communication device (such as Figure 5The "New ONU" shown in [figure] is connected to the first communication device. The second communication device reports its channel condition information (such as CSI) to the first communication device. If the first communication device detects that the second communication device meets a predetermined trigger condition, it will generate neural network parameter information NN_param based on the channel condition information reported by the second communication device. NN_param includes the jump size (Jump_size) assigned to the second communication device. In this example, the value of Jump_size is 2. After that, the first communication device assigns NN_param to the second communication device. The second communication device configures a deterministic reservoir with jumps based on NN_param, and the second communication device establishes a signal equalizer based on this deterministic reservoir. As Figure 5 shown in the deterministic reservoir, R1 and R8 are neurons (only 8 neurons are shown exemplarily), and all the recurrent connections therein (such as Figure 5 the connections from R1 to R2, R2 to R3, R3 to R4, R4 to R5, R5 to R6, R6 to R7, R7 to R8, and R8 to R1 shown) share the first weight W c , and all the bidirectional jumps (jumps, such as Figure 5 the bidirectional jumps between R1 and R3, R3 and R5, R5 and R7, and R7 and R1 shown. Since the value of Jump_size in this example is 2, R1 jumps 2 neurons to R3, and so on) share the second weight W j , and the deterministic reservoir matrix W res is fixed as follows: the non-zero elements include all the above-mentioned recurrent connections and all the bidirectional jumps, and the remaining elements are all set to 0.

[0089] Figure 6 shows a schematic flow diagram of establishing a signal equalizer based on DRJ in an example of this application. The specific process is as follows: The first communication device continuously monitors the status of the second communication device. The newly connected second communication device reports channel condition information (such as CSI) to the first communication device; when it is found that the second communication device is newly connected and has never been seen by the first communication device before, the first communication device determines that the second communication device meets a predetermined trigger condition (the second communication device is newly connected and has never been seen before), and then generates neural network parameter information (NN_param) including the jump size (Jump_size) for the second communication device according to the channel condition information; after that, the first communication device assigns the generated NN_param to the second communication device. The second communication device configures the NN (only trains the readout layer) based on NN_param. After the training is completed, it sends a "Ready" signal to the first communication device. The first communication device transmits load data to the second communication device, and the second communication device realizes load balancing based on the trained signal equalizer.

[0090] Figure 7 Shows a single - output architecture of an example of this application. The single - output architecture includes an input layer (Input), a deterministic reservoir (DRJ, which may also be called a fixed hidden layer), and a trainable readout layer (Readout). The readout layer has a single output, and the DRJ is configured based on the jump size (Jump_size) assigned by the first communication device to the second communication device. Figure 7 The shown deterministic reservoir is only for exemplary display. Still taking Jump_size = 2 as an example. As Figure 7 shown, the received signals (Received symbols) are input to the input layer. The readout layer has only one output O1, where d represents the received signals with sequential delays (delay), X represents the input, Y represents the output, T represents the time, and W in represents the input matrix, and W out represents the output matrix.

[0091] Figure 8 Shows a parallel - output architecture of an example of this application. The parallel - output architecture includes an (Input), a deterministic reservoir (DRJ), and a trainable readout layer (Readout). The readout layer has parallel outputs, and the DRJ is configured based on the jump size (Jump_size) assigned by the first communication device to the second communication device. Figure 8 The shown deterministic reservoir is only for exemplary display. Still taking Jump_size = 2 as an example. As Figure 8 shown, the received signals (Received symbols) are input to the input layer. The readout layer has n outputs (denoted as O1,..., O n ), where d represents the received signals with sequential delays (delay), X represents the input, Y represents the output, T represents the time, and W in represents the input matrix, and W out represents the output matrix.

[0092] Figure 9 Shows an experimental setup of an example of a 100Gbps / λPAM4 - based IMDD (Intensity Modulation and Direct Detection) PON of this application. This experimental setup is used to verify the proposed scheme of this application. As Figure 9As shown, on the transmitter side (Tx), the DFB at 1310 nm is externally intensity-modulated by a 50 GBaud PAM4 signal using a Mach-Zehnder Modulator (MZM). After 20 km of SSMF transmission, a Variable Optical Attenuator (VOA) is used for receiver sensitivity measurement. The optical signal is directly detected by a 25 GHz Photodiode (PD), and then, the detected signal is sampled by a Digital Sample Oscilloscope (DSO) with a 59 GHz bandwidth and a 160 GSa / s sampling rate. Finally, the sampled signal is processed offline. The offline processing includes resampling to 50 GSa / s and training of the proposed DRJ-based signal equalizer.

[0093] This application is based on Figure 9 the experimental setup for simulation and experiments. The datasets for training and testing are both random sequences based on the Mersenne Twister. The length of the entire dataset is 100,000, including 40,000 symbols for training and 60,000 symbols for testing. The proposed DRJ method in this application can complete training on the CPU of a laptop in a few seconds, thus enabling adaptive equalization of high-speed PONs with bandwidth-limited optical devices in an easily trainable manner. After careful consideration and detailed evaluation, preferably, a specific jump size can be set to 8 according to the measured channel conditions. Figure 10 shows based on Figure 7 the BER corresponding to different jump sizes tested on the single-output architecture shown.

[0094] Some example experiments of this application are described below.

[0095] In Example Experiment 1, the single-output architecture (DRJ1) shown in Figure 7 is adopted. The number of inputs in the input layer is 17, the number of neurons in the deterministic reservoir is 50 / 500, and the number of outputs in the readout layer is 1, denoted as [17i, (50 / 500, 1)]. In this experiment, the above-mentioned specific jump size (i.e., 8) is used. In Example Experiment 2, the parallel-output architecture (ParaDRJ4) shown in Figure 8 is adopted. The number of inputs in the input layer is 20, the number of neurons in the deterministic reservoir is 50 / 500, and the number of outputs in the readout layer is 4, denoted as [20i, (50 / 500, 4)]. The above-mentioned specific jump size is also used in this experiment. Figure 11 For the test results of Experiment 1 (single output), Figure 12For the test results of Experiment 2 (parallel output), based on Figure 11 and Figure 12 As can be seen (the solid line is DRJ and the dashed line is RC), Experiment 1 and Experiment 2 can achieve better BER performance than traditional RC at the same ROP. It should be noted that when the reservoir size is small, such as the reservoir size is 50, the performance of DRJ1 of [17i,(50,1)] and ParaDRJ4 of [20i,(50,4)] is even better than that of the traditional single-output RC (RC1) of [17i,(500,1)] and the parallel-output RC (ParaRC4) of [20i,(500,4)] with a large reservoir size of 500.

[0096] The present application also proposes a signal equalizer based on reservoir computing, or a signal equalizer based on the DRJ proposed in the present application. Among them, the signal equalizer includes a reservoir module and a readout module. The reservoir module includes a deterministic reservoir, and the deterministic reservoir is configured using the jump size allocated by the first communication device for the second communication device. In some embodiments, the non-zero elements in the matrix corresponding to the deterministic reservoir include all cyclic connections and all two-way jumps. All cyclic connections share the first weight, all two-way jumps share the second weight, and the remaining elements of the matrix are set to zero.

[0097] In some embodiments, the reservoir module is implemented by a passive integrated photonic chip or an SoC or an FPGA, the readout module is implemented based on an optoelectronic readout chip, and the trained weights of the readout module are configured by an electronic microprocessor or an SoC or an FPGA in the optoelectronic readout chip. Figure 13 Shows a schematic diagram of a DRJ-based signal equalizer according to an example of the present application. The signal equalizer includes a trainable readout module and a reservoir module implemented by a fixed passive integrated photonic chip (such as Figure 13 the passive DRJ shown in Figure 13 where the jump size of the passive DRJ still takes 2 as an example. Based on Figure 13 the signal equalizer shown, the input signal (not shown in the figure) is input to the deterministic reservoir. Each output signal of the deterministic reservoir is connected to a PD (Photoelectric Detector) in the readout module, and then handed over to an ADC (Analog to Digital Converter) and modulated by an MP (electronic Microprocessor) to implement the trainable weight, and then output.

[0098] In some embodiments, the reservoir module is implemented by a passive integrated photonics chip, or a SoC, or an FPGA, the readout module is implemented based on an all-optical readout chip, and the trained weights of the readout module are configured by optical modulators in the all-optical readout chip. Figure 14 The schematic diagram of a DRJ-based signal equalizer according to another example of the present application is shown. The signal equalizer includes a trainable readout module and a reservoir module implemented by a fixed passive integrated photonics chip (such as Figure 14 the passive DRJ shown), wherein the readout module is implemented based on an all-optical readout chip (such as an all-optical readout device based on silicon photonics), which supports single-output / parallel output, Figure 14 and the passive DRJ in it still takes the hop size value of 2 as an example. Based on Figure 14 the shown signal equalizer, an input signal (not shown in the figure) is input to a deterministic reservoir, and each output signal of the deterministic reservoir is connected to an OM (Optical Modulator) in the all-optical readout chip, and is modulated by the OM to implement trainable weights, and then output through a detector.

[0099] According to the solution of the present application, the concept of DRJ is proposed. The second communication device can configure the DRJ based on the hop size allocated by the first communication device, so as to implement a DRJ-based signal equalizer, which can support the adaptive equalization of a high-speed communication system in an easy-to-train and low-power consumption manner. And due to the structural characteristics and relatively simple linear regression training characteristics brought by its deterministic structure with hops, the solution of the present application can reduce the complexity and power consumption of the second communication device and enable it to support a simpler 、 lower-power consumption and higher-speed hardware implementation (such as passive integrated photonics chips, SoCs or FPGAs); moreover, by using the allocated hop size and a smaller reservoir size, the solution of the present application has better BER performance than the prior art. For example, through experiments, [17 / 20i, (50,1 / 4)] adopted in the present application compared with [17 / 20i, (500,1 / 4)] adopted by the existing RC, the reservoir size is reduced by 90%, but the performance is due to the existing RC; in addition, the present application can also support a wide range of applications, such as speech recognition, robot control, time series prediction, image processing, etc.

[0100] It should be noted that the names of the various modules or devices involved in the present application are only examples and do not limit the present application. Multiple modules or devices can also be implemented by one module or device through software or hardware, and each module or device can also be divided into multiple modules or devices through software or hardware.

[0101] Figure 15An exemplary system that can be used to implement the various embodiments described in this application is shown.

[0102] In some embodiments, system 1000 can serve as any one of the processing devices in the embodiments of this application. In some embodiments, system 1000 may include one or more computer-readable media having instructions (e.g., system memory or NVM / storage device 1020) and one or more processors (e.g., (one or more) processors 1005) coupled to the one or more computer-readable media and configured to execute the instructions to implement modules to perform the actions described in this application.

[0103] For one embodiment, system control module 1010 may include any suitable interface controller to provide any suitable interface to at least one of (one or more) processors 1005 and / or any suitable device or component communicating with system control module 1010.

[0104] System control module 1010 may include a memory controller module 1030 to provide an interface to system memory 1015. Memory controller module 1030 can be a hardware module, a software module, and / or a firmware module.

[0105] System memory 1015 can be used, for example, to load and store data and / or instructions for system 1000. For one embodiment, system memory 1015 may include any suitable volatile memory, e.g., suitable DRAM. In some embodiments, system memory 1015 may include double data rate type four synchronous dynamic random access memory (DDR4 SDRAM).

[0106] For one embodiment, system control module 1010 may include one or more input / output (I / O) controllers to provide an interface to NVM / storage device 1020 and (one or more) communication interfaces 1025.

[0107] For example, NVM / storage device 1020 can be used to store data and / or instructions. NVM / storage device 1020 may include any suitable non-volatile memory (e.g., flash memory) and / or may include any suitable (one or more) non-volatile storage devices (e.g., one or more hard disk drives (HDDs), one or more optical disc (CD) drives, and / or one or more digital versatile disc (DVD) drives).

[0108] The NVM / storage device 1020 may include storage resources that are physically part of the device on which the system 1000 is installed, or it may be accessible to the device without being part of the device. For example, the NVM / storage device 1020 may be accessed via a network through one or more communication interfaces 1025.

[0109] One or more communication interfaces 1025 may provide an interface for the system 1000 to communicate through one or more networks and / or with any other suitable device. The system 1000 may wirelessly communicate with one or more components of a wireless network according to any one or more of the wireless network standards and / or protocols.

[0110] For one embodiment, at least one of the one or more processors 1005 may be logically encapsulated with one or more controllers (e.g., the memory controller module 1030) of the system control module 1010. For one embodiment, at least one of the one or more processors 1005 may be logically encapsulated with one or more controllers of the system control module 1010 to form a system-in-package (SiP). For one embodiment, at least one of the one or more processors 1005 may be logically integrated with one or more controllers of the system control module 1010 on the same die. For one embodiment, at least one of the one or more processors 1005 may be logically integrated with one or more controllers of the system control module 1010 on the same die to form a system-on-chip (SoC).

[0111] In various embodiments, the system 1000 may be, but is not limited to: a server, a workstation, a desktop computing device, or a mobile computing device (e.g., a laptop computing device, a handheld computing device, a tablet computer, a netbook, etc.). In various embodiments, the system 1000 may have more or fewer components and / or a different architecture. For example, in some embodiments, the system 1000 includes one or more cameras, a keyboard, a liquid crystal display (LCD) screen (including a touchscreen display), a non-volatile memory port, multiple antennas, a graphics chip, an application specific integrated circuit (ASIC), and a speaker.

[0112] This application also provides a first communication device, wherein the first communication device includes:

[0113] A memory for storing one or more programs;

[0114] One or more processors connected to the memory,

[0115] When the one or more programs are executed by the one or more processors, the one or more processors are caused to execute the reservoir computing-based method in the first communication device described in the present application.

[0116] The present application also provides a second communication device, wherein the second communication device includes:

[0117] a memory for storing one or more programs;

[0118] one or more processors connected to the memory,

[0119] When the one or more programs are executed by the one or more processors, the one or more processors are caused to execute the reservoir computing-based method in the second communication device described in the present application.

[0120] The present application also provides a computer-readable storage medium, on which a computer program is stored, and the computer program can be executed by a processor to perform the reservoir computing-based method described in the present application.

[0121] The present application also provides a computer program product, which when executed by a device, causes the device to execute the reservoir computing-based method described in the present application.

[0122] For those skilled in the art, it is obvious that the present application is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or basic characteristics of the present application. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present application is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present application. Any reference signs in the claims should not be regarded as limiting the claimed rights. In addition, it is obvious that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. The multiple units or devices stated in the system claims can also be implemented by one unit or device through software or hardware. First, second, etc. are used to denote names and do not denote any particular order.

Claims

1. A method based on reservoir computing, applied to a first communication device, wherein, The method includes: When a predetermined trigger condition is satisfied, generating neural network parameter information corresponding to the second communication device according to the channel condition information reported by the second communication device, where the neural network includes an input layer, a deterministic reservoir, and a trainable readout layer, the neural network parameter information includes the jump size corresponding to the reservoir, and configuring the deterministic reservoir based on the jump size; Sending the neural network parameter information to the second communication device.

2. The method according to claim 1, wherein The predetermined trigger condition includes any one of the following: The second communication device is newly connected; The channel condition of the second communication device changes.

3. The method according to claim 1, wherein, The method further includes: Receiving a ready signal sent by the second communication device after training is completed; Sending load data to the second communication device.

4. A method for a signal equalizer based on reservoir computing, applied to a second communication device, wherein, The method includes: Receiving neural network parameter information sent by the first communication device, where the neural network parameter information includes the jump size corresponding to the reservoir; Configuring a neural network according to the neural network parameter information, where the neural network includes an input layer, a deterministic reservoir, and a trainable readout layer, and configuring the deterministic reservoir based on the jump size; Training the readout layer by using simple linear regression to obtain a trained signal equalizer.

5. The method according to claim 4, wherein, The method further includes: Sending a ready signal to the first communication device after training is completed; Receiving load data from the first communication device and implementing load balancing by using the signal equalizer.

6. The method according to claim 4, wherein The non-zero elements in the matrix corresponding to the deterministic reservoir include all recurrent connections and all bidirectional jumps, all recurrent connections share a first weight, all bidirectional jumps share a second weight, and the remaining elements of the matrix are all set to zero.

7. A first device based on reservoir computing in a first communication device, wherein, The first device includes: A device for generating neural network parameter information corresponding to the second communication device according to the channel condition information reported by the second communication device when a predetermined trigger condition is satisfied, where the neural network includes an input layer, a deterministic reservoir, and a trainable readout layer, the neural network parameter information includes the jump size corresponding to the reservoir, and configuring the deterministic reservoir based on the jump size; A device for sending the neural network parameter information to the second communication device.

8. A second device based on reservoir computing in a second communication device, wherein, The second device includes: A device for receiving neural network parameter information sent by the first communication device, where the neural network parameter information includes the jump size corresponding to the reservoir; A device for configuring a neural network according to the neural network parameter information, where the neural network includes an input layer, a deterministic reservoir, and a trainable readout layer, and configuring the deterministic reservoir based on the jump size; A device for training the readout layer by using simple linear regression to obtain a trained signal equalizer.

9. The second device according to claim 8, wherein, The second device further includes: A device for sending a ready signal to the first communication device; A device for receiving load data from the first communication device and implementing load balancing by using the signal equalizer.

10. The second device according to claim 8, wherein, The non-zero elements in the matrix corresponding to the deterministic reservoir include all cyclic connections and all two-way jumps. All cyclic connections share a first weight, and all two-way jumps share a second weight. The remaining elements of the matrix are set to zero.

11. A signal equalizer based on reservoir computing, wherein, The signal equalizer includes a reservoir module and a trainable readout module. The reservoir module includes a deterministic reservoir, and the deterministic reservoir is configured using the jump size allocated by the first communication device for the second communication device.

12. The signal equalizer according to claim 11, wherein, The reservoir module is implemented by a passive integrated photonics chip or an SoC or an FPGA. The readout module is implemented based on an optoelectronic readout chip, and the trained weights of the readout module are configured by an electronic microprocessor or an SoC or an FPGA in the optoelectronic readout chip.

13. The signal equalizer according to claim 11, wherein, The reservoir module is implemented by a passive integrated photonics chip or an SoC or an FPGA. The readout module is implemented based on an all-optical readout chip, and the trained weights of the readout module are configured by an optical modulator in the all-optical readout chip.

14. A first communication device, wherein, The first communication device includes: a memory for storing one or more programs; one or more processors connected to the memory, when the one or more programs are executed by the one or more processors, causing the one or more processors to perform the following operations: when a predetermined trigger condition is satisfied, generating neural network parameter information corresponding to the second communication device according to the channel condition information reported by the second communication device. The neural network includes an input layer, a deterministic reservoir, and a trainable readout layer. The neural network parameter information includes the jump size corresponding to the reservoir, and configuring the deterministic reservoir based on the jump size; sending the neural network parameter information to the second communication device.

15. A second communication device, wherein, The second communication device includes: a memory for storing one or more programs; one or more processors connected to the memory, when the one or more programs are executed by the one or more processors, causing the one or more processors to perform the following operations: receiving the neural network parameter information sent by the first communication device, where the neural network parameter information includes the jump size corresponding to the reservoir; configuring a neural network according to the neural network parameter information. The neural network includes an input layer, a deterministic reservoir, and a trainable readout layer, and configuring the deterministic reservoir based on the jump size; obtaining a trained signal equalizer by training the readout layer using simple linear regression.

Images