Method, device, apparatus and medium for optical communication

By exchanging feedback information between optical communication devices and updating the initial parameter values ​​of the neural network, the problem of low training speed and accuracy of neural networks in optical communication is solved, and the efficiency and adaptability of signal processing are improved.

CN114793305BActive Publication Date: 2024-11-26ALCATEL LUCENT SHANGHAI BELL CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202110096879.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-01-25
Publication Date
2024-11-26
Estimated Expiration
2041-01-25

AI Technical Summary

Technical Problem

In optical communication, it is difficult for the prior art to quickly and efficiently train neural networks for signal processing, and the trained neural networks perform poorly under different communication channel conditions, resulting in low accuracy and energy efficiency of signal processing.

Method used

By exchanging feedback information between optical communication devices, the initial parameter value set of neural networks is updated, so that each optical communication device can train based on better initial values, thereby improving training speed and accuracy and adapting to different channel conditions.

Benefits of technology

It realizes fast and reliable training of neural networks, improves the signal processing performance and energy efficiency of optical communication equipment, and shortens training time and computing resource consumption.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114793305B_ABST
    Figure CN114793305B_ABST
Patent Text Reader

Abstract

Embodiments of the present disclosure relate to methods, devices, apparatuses, and media for optical communications. The method includes receiving feedback information for training of a neural network from at least one second optical communication device at a first optical communication device, the neural network being configured to process a signal received from the first optical communication device, the feedback information at least including a training performance indication of each training of the neural network at at least one second optical communication device; updating a first initial parameter value set for the neural network maintained at the first optical communication device based on the feedback information to obtain a second initial parameter value set for the neural network; and transmitting the second initial parameter value set to at least one other second optical communication device for training the neural network based on the second initial parameter value set at at least one other second optical communication device. In this way, the training speed and the accuracy and reliability of the neural network can be improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] Embodiments of the present disclosure relate to the field of optical communications, and more particularly to methods, devices, apparatuses, and computer-readable media for optical communications. Background Art

[0002] For the new generation of high-speed passive optical networks (PON), such as PON with 50Gbps or even higher rates, the transmitted signals may experience high distortion and noise interference due to poor channel response and high link loss budget requirements. Optical communication equipment as the signal receiving end needs to perform signal equalization through signal processing to eliminate or reduce signal distortion and noise interference.

[0003] In recent years, signal processing based on neural networks (NN) has gradually been recognized as a promising signal processing method. Compared with traditional linear equalizers, NN-based signal processing can automatically learn channel characteristics and more effectively recover the original signal from received signals with linear and nonlinear distortions. Such high-performance signal processing is particularly suitable for PON. Summary of the invention

[0004] Generally, embodiments of the present disclosure relate to improved training schemes for neural networks (NNs) used for signal processing in optical communications.

[0005] In a first aspect of the present disclosure, a method for optical communication is provided. The method includes receiving, at a first optical communication device, feedback information for training of a neural network from at least one second optical communication device, the neural network being configured to process a signal received from the first optical communication device, the feedback information at least including a training performance indication of training the neural network at the at least one second optical communication device respectively; updating a first initial parameter value set for the neural network maintained at the first optical communication device based on the feedback information to obtain a second initial parameter value set for the neural network; and transmitting the second initial parameter value set to at least one other second optical communication device, for training the neural network based on the second initial parameter value set at the at least one other second optical communication device.

[0006] In a second aspect of the present disclosure, a method for optical communication is provided. The method includes training a neural network at a second optical communication device, the neural network being configured to process a signal received from a first optical communication device; and transmitting feedback information for the training of the neural network to the first optical communication device, the feedback information at least including a training performance indication of training the neural network at the second optical communication device.

[0007] In a third aspect of the present disclosure, a first optical communication device is provided. The first optical communication device comprises at least one processor; and a memory coupled to the at least one processor, the memory containing instructions stored therein, and when the instructions are executed by the at least one processing unit, the first optical communication device executes the method of the first aspect.

[0008] In a fourth aspect of the present disclosure, a second optical communication device is provided. The second optical communication device includes at least one processor; and a memory coupled to the at least one processor, wherein the memory contains instructions stored therein, and when the instructions are executed by at least one processing unit, the second optical communication device executes the method of the second aspect.

[0009] In a fifth aspect of the present disclosure, a first optical communication device is provided. The first optical communication device includes a component for receiving feedback information for training of a neural network from at least one second optical communication device, the neural network being configured to process a signal received from the first optical communication device, the feedback information at least including a training performance indication of training the neural network at the at least one second optical communication device; a component for updating a first initial parameter value set for the neural network maintained at the first optical communication device based on the feedback information to obtain a second initial parameter value set for the neural network; and a component for transmitting the second initial parameter value set to at least one other second optical communication device, so as to train the neural network based on the second initial parameter value set at the at least one other second optical communication device.

[0010] In a sixth aspect of the present disclosure, a second optical communication device is provided. The second optical communication device includes a component for training a neural network, the neural network being configured to process a signal received from a first optical communication device; and a component for transmitting feedback information for the training of the neural network to the first optical communication device, the feedback information at least including a training performance indication of training the neural network at the second optical communication device.

[0011] In a seventh aspect of the present disclosure, a computer-readable medium is provided, wherein instructions are stored on the computer-readable medium, and when the instructions are executed by at least one processing unit, the at least one processing unit is configured to execute the method of the first aspect.

[0012] In an eighth aspect of the present disclosure, a computer-readable medium is provided, wherein instructions are stored on the computer-readable medium, and when the instructions are executed by at least one processing unit, the at least one processing unit is configured to execute the method of the second aspect.

[0013] It should be understood that the contents described in the summary of the invention are not intended to limit 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 easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Embodiments of the present disclosure are presented by way of example and their advantages are explained in more detail below with reference to the accompanying drawings, in which

[0015] Figure 1 A schematic diagram showing an optical communication network in which embodiments described in the present disclosure may be implemented;

[0016] Figure 2 is a comparison of the average test errors of the joint training scheme and the self-training scheme for neural networks used for signal processing;

[0017] Figure 3 A schematic diagram of an optical communication network according to some embodiments of the present disclosure is shown;

[0018] Figure 4 A schematic diagram showing the structure of a neural network according to some embodiments of the present disclosure;

[0019] Figure 5 A signaling diagram illustrating an optical communication process between optical communication devices according to some example embodiments of the present disclosure;

[0020] Figure 6 An example of a frame structure for carrying a pilot sequence according to some example embodiments of the present disclosure is shown;

[0021] Fig. 7A A signaling diagram illustrating a process for transmitting an initial parameter value set according to some example embodiments of the present disclosure;

[0022] Figure 7B A signaling diagram illustrating a process for transmitting an initial parameter value set according to some further example embodiments of the present disclosure;

[0023] Figure 8 A flowchart showing a process for updating an initial set of parameter values ​​for a neural network according to some example embodiments of the present disclosure;

[0024] Fig. 9 A comparison chart of average test errors of example training according to the present disclosure and traditional training is shown;

[0025] Fig.10 shows a comparison graph of average test errors of two example trainings according to the present disclosure according to some example embodiments of the present disclosure;

[0026] Fig.11 A comparison chart of average test errors of another example training according to the present disclosure and traditional training is shown;

[0027] Fig.12 A flowchart showing a method implemented at a first optical communication device according to some example embodiments of the present disclosure;

[0028] Fig.13 A flowchart showing a method implemented at a second optical communication device according to some example embodiments of the present disclosure;

[0029] Fig.14 shows a simplified block diagram of an electronic device suitable for implementing embodiments of the present disclosure; and

[0030] Fig.15 A schematic diagram of a computer-readable medium suitable for implementing an embodiment of the present disclosure is shown.

[0031] Throughout the drawings, the same or similar reference numerals refer to the same or similar elements. DETAILED DESCRIPTION

[0032] The principles and spirit of the present disclosure will be described below with reference to several exemplary embodiments shown in the accompanying drawings. It should be understood that these specific embodiments are described only to enable those skilled in the art to better understand and implement the present disclosure, and are not intended to limit the scope of the present disclosure in any way.

[0033] As used herein, the term "including" and similar terms should be understood as open inclusion, i.e., "including but not limited to". The term "based on" should be understood as "based at least in part on". The term "one embodiment" or "the embodiment" should be understood as "at least one embodiment". The terms "first", "second", etc. may refer to different or the same objects. Other explicit and implicit definitions may also be included below.

[0034] As used herein, the term "determine" encompasses a wide variety of actions. For example, "determine" may include computing, calculating, processing, deriving, investigating, searching (e.g., searching in a table, a database, or another data structure), ascertaining, etc. In addition, "determine" may include receiving (e.g., receiving information), accessing (e.g., accessing data in a memory), etc. In addition, "determine" may include resolving, selecting, choosing, establishing, etc.

[0035] The term "circuitry" as used herein refers to one or more of the following: (a) a hardware circuit implementation only (such as an implementation of analog and / or digital circuits only); and (b) a combination of hardware circuits and software, such as (where applicable): (i) a combination of analog and / or digital hardware circuits and software / firmware, and (ii) any portion of a hardware processor and software (including a digital signal processor, software, and memory that work together to enable a device such as an optical communication device or other computing device to perform various functions); and (c) a hardware circuit and / or processor, such as a microprocessor or a portion of a microprocessor, that requires software (e.g., firmware) for operation, but can be without software when software is not required for operation.

[0036] The definition of circuitry applies to all uses of this term in this application, including in any claims. As another example, the term "circuitry" as used herein also covers an implementation of a hardware circuit or processor (or multiple processors), or a portion of a hardware circuit or processor, or its accompanying software or firmware. For example, if applicable to a particular claim notation, the term "circuitry" also covers a baseband integrated circuit or a processor integrated circuit or a similar integrated circuit in an OLT or other computing device.

[0037] The term "optical communication equipment" or "optical communication device" as used herein refers to a device or apparatus used for communication in a communication network based on optical fiber. As an example, examples of optical communication equipment or optical communication devices may include an optical line terminal (OLT), an optical network unit (ONU), or any other optical network element suitable for use in an optical network such as a passive optical network (PON) or any other optical transport network (OTN).

[0038] As used herein, the term "passive optical network" or "PON" refers to an optical distribution network (ODN) comprised of passive components such as optical splitters and optical fibers, without requiring active components.

[0039] The term "neural network (NN)" used in this article can be understood as a machine learning model. The model can learn the association between the corresponding input and output from the training data, so that after the training is completed, the given input is processed based on the parameter value set obtained by the training to generate the corresponding output. "Neural network" may also be sometimes referred to as "neural network model", "learning network", "learning model", "network" or "model", etc. These terms are used interchangeably in this article.

[0040] Machine learning can be divided into three stages, namely the training stage, the testing stage, and the application stage (also called the inference stage). In the training stage, a given learning model can be trained using a large amount of training data, and the values ​​of the model's parameter set are continuously iterated and updated until the model can achieve the desired goal. After training, the values ​​of the parameter set of the machine learning model are determined. In the testing stage, the trained learning model can be tested with test samples to determine the performance of the learning model. In the application stage, actual input data can be applied to the trained learning model. The learning model can process the actual input data based on the parameter set obtained by training to provide corresponding output.

[0041] Figure 1 1 is a schematic diagram of an optical communication network 100 in which the embodiments described in the present disclosure may be implemented. The optical communication network 100 is a communication network based on optical fiber, for example, a PON. The optical communication network 100 includes a plurality of optical communication devices, for example, an OLT 110 and a plurality of (N) ONUs 120-1, 120-2, 120-3, ..., 120-N (hereinafter collectively referred to as ONUs 120 for convenience of discussion), where N is an integer greater than 1.

[0042] OLT 110 and ONU 120 can communicate with each other. For example, OLT 110 can send signals to ONU 120, and the signal transmission in this direction is called a downlink or downstream link. ONU 120 can also send signals to OLT 110, and the signal transmission in this direction is called an uplink or upstream link. Figure 1 In the example shown in FIG. 1 , in the downlink direction, point-to-multipoint communication is formed between the OLT 110 and the plurality of ONUs 120 .

[0043] It should be understood that Figure 1 The number of OLTs and ONUs shown is given for illustration purposes. The optical communication network 100 may include any suitable number of OLTs and ONUs.

[0044] like Figure 1 As shown, a plurality of ONUs 120-1, 120-2, 120-3, ..., 120-N respectively include neural networks (NN) 130-1, 130-2, 130-3, ..., 130-N for signal processing (hereinafter collectively referred to as neural network 130 for convenience of discussion). The neural network 130 is configured to process the signal received from the OLT 110 at the ONU 120 to reduce or eliminate the distortion and / or noise in the received signal, thereby recovering the original signal sent by the OLT 110. The signal processing operation of the neural network 130 may include equalization processing, so the neural network 130 may be implemented as or included in the equalizer of the ONU 120.

[0045] As mentioned above, neural network-based signal processing, such as equalization processing, can achieve signal recovery more effectively. For communication channels with poor channel response and high link loss budget requirements, such as high-speed communication channels with a rate of 50Gbps or even higher, neural network-based signal processing can better improve the performance of the receiving end.

[0046] like Figure 1 As shown, OLT 110 may include a signal processing device 112, which may include an optical-to-electrical converter for converting an optical signal into an electrical signal. Signal processing device 112 may also include an equalizer that can restore the original signal sent from ONU 120 from the distorted signal received at OLT 110.

[0047] Although Figure 1 Only the neural network deployed at the ONU for signal processing is shown. In some embodiments, a neural network can also be deployed at the OLT to process the signal received from the ONU.

[0048] Generally, in order to apply a neural network, it is necessary to train the neural network. When training the neural network, the optical communication device at the transmitting end transmits a known pilot signal to the optical communication device at the receiving end. After receiving a distorted pilot signal via a communication channel, the optical communication device uses the received pilot signal and the known pilot signal as training data to train the neural network so that the neural network can correctly recover the known pilot signal from the received pilot signal. In this way, the neural network learns how to reduce or eliminate the distortion and noise interference introduced in the communication channel through which the pilot signal passes. As mentioned above, the training process of the neural network iteratively updates the values ​​of the parameter set of the neural network until the neural network can recover the input received pilot signal into the known pilot signal using the parameter set obtained by training.

[0049] In a network based on optical communication, for an optical communication device as a receiving end, the trained neural network is closely related to the channel conditions between the optical communication device at the transmitting end connected to the device, and therefore cannot be applied to optical communication devices at other receiving ends. This is particularly problematic in point-to-multipoint communication networks such as PON, because the channel conditions between multiple optical communication devices (such as ONU) as receiving ends and the same transmitting end (such as OLT) may be different, for example, they may have different optical fiber lengths, different bandwidths, etc.

[0050] In order to ensure the application performance of the neural network, one solution is to perform self-training at an optical communication device as a receiving end to individually train the neural network to be used for the communication channel between the optical communication device and the corresponding transmitting end.

[0051] In order to make the application performance of the trained neural network reach a satisfactory level (for example, the error reaches the expected level), a large amount of training data needs to be collected, that is, more pilot signals are received from the transmitting end, and the number of training iterations of the neural network is also required to be greater. This will increase the time and computing resources consumed by a single training. The inventors found that in a self-training scheme, starting from a randomly initialized parameter set, more than 200,000 iteration steps and a large amount of training data may be required to train a neural network with a satisfactory error. Therefore, if a neural network is trained separately for each optical communication device as a receiving end in an optical communication network through a self-training scheme, the time and computing resources consumed for training in the entire network will be very large.

[0052] For a certain optical communication device, if the neural network previously trained by other optical communication devices is used to initialize its own neural network and then train the neural network used locally, this solution will not work because the neural network trained for each optical communication device is only suitable for the data set specific to its own communication channel and cannot quickly adapt to new contexts.

[0053] A joint training scheme is also proposed. This scheme mixes pilot signals received at multiple optical communication devices connected to the same transmitting end, and uses them to jointly train a neural network to be deployed to multiple optical communication devices. However, although joint training can reduce the training time of the entire network, the application performance of this universal neural network, such as the performance in terms of signal processing accuracy, is worse than that of the neural network obtained through a self-training scheme at a single optical communication device. Therefore, the neural network obtained by joint training may not meet system requirements in many application scenarios. In addition, the universal neural network may need to be designed to be more complex. Therefore, when used for actual signal processing, the energy efficiency of the universal neural network is lower than that of a neural network trained individually.

[0054] Figure 2 A comparison chart of the average test errors of the joint training scheme and the self-training scheme for a neural network for signal processing is shown, where the horizontal axis represents the number of training steps performed using the stochastic gradient descent (SGD) algorithm during the training phase, and the vertical axis represents the average test error rate, such as the bit error rate (BER). Figure 2 Curve 210 shows the variation of the average test error rate of joint training with the number of training steps, and curve 220 shows the variation of the average test error rate of self-training with the number of training steps. Assume that both joint training and self-training are performed based on a randomly initialized parameter set. By comparing curves 210 and 220, it can be seen that the joint training scheme still has difficulty in obtaining a satisfactory low error rate after many training steps.

[0055] In summary, for the neural networks currently deployed for signal processing in optical communications, it is expected that the training can be completed in a fast and efficient manner while ensuring that the trained neural networks have good performance.

[0056] According to an embodiment of the present disclosure, an improved training scheme for a neural network used for signal processing in optical communications is provided. In this scheme, for a neural network configured to process a signal received from a first optical communication device, the first optical communication device maintains the initial value of a parameter set of the neural network (referred to as an "initial parameter value set"). One or more second optical communication devices as receiving ends each train the neural network and transmit feedback information including at least a training performance indication to the first optical communication device. The first optical communication device updates the maintained initial parameter value set based on the received feedback information. The updated initial parameter value set is provided to one or more additional second optical communication devices so that the additional second optical communication devices can train the neural network based on the updated initial parameter value set.

[0057] It has been proven that the initialization of the parameter set of the neural network is important for achieving fast convergence and obtaining a reliable model. According to the scheme of the present disclosure, by continuously updating the initial values ​​of the parameter set of the neural network, the optical communication device can train the neural network based on better initial values, thereby improving the training speed at the optical communication device and the accuracy and reliability of the neural network. Since the training is performed separately at each optical communication device, the trained neural network can adapt to the channel conditions between the optical communication device and the communication device that sends the signal more quickly and reliably.

[0058] Hereinafter, principles and exemplary embodiments of the present disclosure will be described in detail with reference to the accompanying drawings.

[0059] First reference Figure 3 , which shows a schematic diagram of an optical communication network 100 according to some embodiments of the present disclosure. Figure 3 In the example of the present invention, the training of the neural network is implemented in the OLT 110 and the ONU 120 of the optical communication network 100. In the example embodiments below, the OLT 110 is taken as an example of an optical communication device at the transmitting end (sometimes also referred to as a "first optical communication device"), and the ONU 120 is taken as an example of an optical communication device at the receiving end (sometimes also referred to as a "second optical communication device"). It should be understood that in other embodiments, the OLT 110 can also be operated as an optical communication device at the receiving end, so that a neural network for processing received signals can be deployed therein; and accordingly, the ONU 120 can be operated as an optical communication device at the transmitting end.

[0060] ONU 120 is configured to train neural network 130 for processing signals received via a communication channel with OLT 110, for example, for performing equalization processing on signals received from OLT 110. In some examples, neural network 130 may be implemented as or included in an equalizer of ONU 120. According to an embodiment of the present disclosure, OLT 110 also maintains initial values ​​of a parameter set of neural network 130 (referred to herein as a “parameter value set”).

[0061] The neural network 130 may be configured as any type of neural network, such as a feedforward neural network (FF-NN), a recurrent neural network (RNN), a convolutional neural network (CNN), etc. When training or updating the neural network 130, the OLT 110 and the ONU 120 may maintain the configuration of the structure of the neural network 130, including the number of network layers included in the neural network, the number of processing units (also called neurons) in each network layer, the connection relationship between each network layer, etc. The structure of the neural network 130 is related to the type of the neural network and actual needs, and the present disclosure is not limited in this respect.

[0062] In some embodiments, the parameter set of the neural network 130 may include weights. Optionally, the parameter set of the neural network 130 may also include offsets. Training and updating the neural network 130 will update the specific values ​​of the weights and offsets. The parameter set of the neural network 130 is related to the configuration of its structure.

[0063] For ease of understanding, we will combine Figure 4 Describes the weights and biases included in the parameter set used by the neural network. Figure 4 FIG. 1 is a schematic diagram showing the structure of a neural network 130 according to some embodiments of the present disclosure. Figure 4 As shown, the neural network may include an input layer 410, a plurality of hidden layers 420, 430, etc., and an output layer 440. It is understood that the neural network may also include other numbers of hidden layers and corresponding numbers of neurons. The protection scope of the present disclosure is not limited in this respect.

[0064] exist Figure 4 In the illustrated embodiment, the hidden layer 420 and the hidden layer 430 in the neural network 130 include a plurality of neurons, respectively. Each of these neurons may include an activation function, which may be, for example, a nonlinear or linear function. The function may be, for example, a sine function or other functions. In addition, the same activation function or different activation functions may be specified for each neuron. The protection scope of the present disclosure is not limited in this respect.

[0065] The input of the neural network 130 is provided to the input layer 410, and then is input to the hidden layer 420 connected to the input layer 410. After the neurons of the hidden layer 420 are processed, their outputs are provided to the next hidden layer 430 for further processing. In one example, the inputs in the input layer 410 are fully connected to the neurons in the hidden layer 420, that is, each input is provided to all neurons in the hidden layer 420. Therefore, the number of weights in the hidden layer 420 is shown as formula (1):

[0066] The number of weights in the hidden layer 420 = the number of inputs * the number of neurons in the hidden layer (1)

[0067] For example, when the number of inputs to the hidden layer 420 is 40 and the number of neurons in the hidden layer 420 is 40, the number of weights in the hidden layer 420 is 40*40=1600.

[0068] In addition, one or more neurons in each hidden layer may also have a corresponding offset. That is, the number of offsets may be the same as or less than the total number of neurons in the hidden layer. For example, a hidden layer 420 having 40 neurons may have 40 offsets. Thus, the operation performed by each neuron in the hidden layer 420 is to multiply each input (e.g., including 40 inputs) in the input layer 410 by 40 weights, and then add the corresponding offset (i.e., input 1*weight 1+input 2*weight 2+input 3*weight 3...input 40*weight 40+offset), and then a value may be output through the nonlinear activation function of each neuron. Thus, each of the 40 neurons in the hidden layer 420 obtains an output, and each output is input to the next hidden layer 430, and then a similar operation as that performed in the hidden layer 420 is performed to obtain a corresponding output. Such operations may be performed backward layer by layer. The last hidden layer of the neural network 130 is connected to the output layer 440 to provide the network output.

[0069] It is understood that the fully connected neural network above is only shown as an example, and the neurons in each layer of the neural network may also be connected in a non-fully connected manner. The scope of protection of the present disclosure is not limited in this respect. When the structure of the neural network 130 is determined, the parameter set of the neural network 130 can be determined. The training of the neural network 130 iteratively updates the parameter value set.

[0070] The OLT 110 also includes a parameter update module 320 configured to update a set of parameter values ​​of the neural network 130 based on feedback information received from one or more OUs 120 for the training of the neural network 130. The OLT 110 may maintain training data 332 for the training of the neural network 130, which may be stored in the storage system 330. Specific examples of the training data 332 used at the OLT 110 will be discussed in detail below.

[0071] ONUs 120-1, 120-2, 120-3, ..., 120-N also include model training modules 340-1, 240-2, 240-3, ..., 240-N, respectively, for training the neural network 130 used by them. In this way, the neural network 130 deployed on each ONU 120 can be trained for the communication channel between the ONU 120 and the OLT 110, so that the neural network 130 trained on each device can reliably and accurately perform the processing of the signals received from the respective communication channels.

[0072] One or more ONUs 120 in the optical communication network 110 may receive an initial parameter value set for the neural network 130 from the OLT 110, so as to start training the neural network 130 on a better initialization basis to improve training efficiency and model accuracy.

[0073] The following will refer to Figure 5 And subsequent Figures 6 to 9 The interaction between ONU 120 and OLT 110 to implement the training of neural network 130 will be discussed in detail.

[0074] Figure 5 A signaling diagram of an optical communication process 500 between optical communication devices according to some example embodiments of the present disclosure is shown. Figure 3 Let's discuss process 500. The process 500 may involve Figure 3 1 and 2. The OLT 110 and the ONU 120 are shown in FIG. The ONU 120 may be any ONU 120 in the optical communication network 110.

[0075] In process 500 , ONU 120 trains 510 neural network 130 .

[0076] For example, the model training module 340 in the ONU 120 may be configured to train the neural network 130. In some embodiments, when the ONU 120 starts and joins the optical communication network 100, or when a specific communication channel of the OLT 110 is to be activated (e.g., a communication channel with a high-speed wavelength, such as a communication channel of 50 Gbps or higher), the ONU 120 may need to train the neural network 130. The neural network 130 is trained to be able to process subsequent signals received from the OLT 110 to reduce or eliminate distortion, noise interference, etc. introduced during the transmission process from the received signal, thereby restoring the original signal.

[0077] The training of the neural network 130 at the ONU 120 may be based on an initial parameter value set of the neural network 130. The ONU 120 may use the initial parameter value set to initialize a parameter set of the neural network 130. The ONU 120 may obtain the initial parameter value set for training in a variety of different ways.

[0078] In some embodiments, the OLT 110 transmits 506 the initial parameter value set to the ONU 120, and the ONU 120 receives 508 the initial parameter value set from the OLT 110. The OLT 110 may maintain and update the initial parameter value set for the neural network 130. In some embodiments, the initial parameter value set provided to one or more ONUs 120 may be the neural network 130 determined by the OLT 110 based on simulated training data. In some embodiments, the OLT 110 updates the initial parameter value set of the neural network 130 based on feedback information from other ONUs 120 for the training of the neural network 130, and provides the updated initial parameter value set to one or more ONUs 120 that are to perform training. The determination of the initial parameter value set of the neural network 130 at the OLT 110 will be discussed in more detail below.

[0079] In some embodiments, one or more ONUs 120 in the optical communication network 100 may not receive an initial parameter value set from the OLT 110 when performing training of the neural network 130, but may perform training based on a default initial parameter value set. For example, the default initial parameter value set may be a factory default, such as may be stored in a storage device of the ONU 120 or in another storage system accessible to the ONU 120.

[0080] The training of the neural network 130 by the ONU 120 is also based on the training data. In some embodiments, the OLT 110 transmits 502 a plurality of pilot sequences to the ONU 120. The plurality of pilot sequences may be the same or different. The pilot sequences may have a predetermined bit pattern, and such a bit pattern may be known to the OLT 110 and the ONU 120. In some embodiments, the OLT 110 may periodically send or broadcast the plurality of pilot sequences in the downstream traffic. The ONU 120 may receive 504 the plurality of pilot sequences for use as training data for the neural network 130.

[0081] The multiple pilot signals sent by OLT 110 may reach different ONUs 120 via different communication channels, and the conditions of these communication channels may be different (for example, having different optical fiber lengths, having different wavelengths, etc.). Therefore, the pilot sequence received at ONU 120 may be different from the original pilot sequence sent by OLT 110, and the pilot sequences received at different ONUs 120 may also be different. Since ONU 120 is aware of the original pilot sequence sent by OLT 110, ONU 120 can use the received pilot sequence as training data to train neural network 130, so that neural network 130 can recover the original pilot sequence from the received pilot sequence as correctly as possible. During the training process, the parameter set of neural network 130 is iteratively updated starting from the initial parameter value set until the convergence target of the training is reached.

[0082] In some embodiments, the pilot sequence sent by OLT 110 may be carried in a message having a predetermined frame structure. Such a message may include a physical (PHY) layer message. In some examples, the message for carrying the pilot sequence may include a serial number request message that OLT 110 periodically sends to discover new ONUs.

[0083] In some embodiments, the pilot sequence transmitted by OLT 110 may be carried in a predetermined section of a payload portion of a message having a predetermined frame structure. Figure 6 An example of a frame structure 600 for carrying a pilot sequence according to some example embodiments of the present disclosure is shown. The frame structure 600 includes a header 610 and a payload portion 620. A predetermined section of the payload portion 620 is used to carry a pilot sequence 622.

[0084] By carrying the pilot sequence in a predetermined section of the payload portion, it is possible to avoid interfering with the normal frame structure in the downstream link. The OLT 110 can broadcast a message carrying the pilot sequence. An ONU 120 that does not need to collect the pilot sequence (e.g., an ONU 120 that is already in a normal communication state with the OLT 110) can directly ignore the pilot sequence in the payload portion. An ONU 120 that needs to collect the pilot sequence to train the neural network 130 can quickly locate the pilot sequence from the payload, such as by using a correlation window to locate the pilot sequence.

[0085] It should be understood that Figure 6 An example of transmitting a pilot sequence is given. The pilot sequence may also be transmitted to the ONU 120 by other messages or signaling. Generally, in order to complete the training of the neural network 130, the ONU 120 may need to collect a certain amount of pilot sequences from the OLT 110 as training data. In some embodiments,

[0086] After obtaining the initial parameter value set and training data (e.g., the pilot sequence received from the OLT 110) of the neural network 130, the ONU 120 may perform training of the neural network 130. The ONU 120 may implement training using any appropriate training algorithm, and the embodiments of the present disclosure are not limited in this respect. In some embodiments, the training goal of the neural network 130 is that the neural network 130 can reach a predetermined convergence goal, such as the accuracy of the signal processing by the neural network 130 can reach a predetermined goal, such as the error rate (e.g., BER) is lower than a specific threshold. Other convergence goals of the training are also possible. The embodiments of the present disclosure are not limited in this respect.

[0087] After the training of the neural network 130 is completed at the ONU 120, the ONU 120 transmits 515 feedback information for the training of the neural network 130 to the OLT 110. The feedback information includes at least an indication of the training performance of the training of the neural network 130 at the ONU 120. The feedback information may indicate to the OLT 110 the training performance when the ONU 120 performs the training using the current set of initial parameter values. As mentioned above, the set of initial parameter values ​​used by the ONU 120 may be a set of initial parameter values ​​currently maintained by the OLT 110, or may be a default set of initial parameter values.

[0088] In some embodiments, the training performance indication in the feedback information may indicate the training efficiency of the training at the ONU 120 and / or the accuracy of the trained neural network. Specifically, the training performance indication may include first indication information indicating the time consumption to complete the training of the neural network 130 at the ONU 120. Alternatively or additionally, the training performance indication may also include second indication information indicating the error rate of the neural network 130 after the training of the neural network 130 is completed at the ONU 120. The lower the error rate, the higher the accuracy.

[0089] The time consumption of training the neural network 130 can be indicated by multiple indicators in the training process. In some embodiments, the first indication information may include the number of epochs used by the ONU 120 to complete the training of the neural network 130. In the training of the neural network, one epoch refers to a complete training of the neural network using all the training data, that is, one forward propagation and one backward propagation using all the data. Generally, in order to achieve a set of parameter values ​​that meet expectations (i.e., meet the training objectives), multiple epochs may be required. The larger the number of epochs, the longer the training of the ONU 120 takes. In some embodiments, the first indication information may include the number of iterative steps used by the ONU 120 to complete the training of the neural network 130. In the training of the neural network, all the training data can be divided into several batches, and one iteration refers to the training performed on the neural network using the training data of each batch. One epoch of training may include multiple iterations. It should be understood that the first indication information may also additionally or alternatively include other indicators that can measure the time consumption of the training of the neural network 130 by the ONU 120.

[0090] In some embodiments, in order to indicate the error rate of the neural network 130 after training, the second indication information may include the test error rate (e.g., BER) in the test phase of the neural network 130, or the bit error rate (e.g., BER) in the phase of being put into use. In some embodiments, if the training of the neural network 130 by the ONU 120 is performed by minimizing the constructed loss function, the second indication information may also include the loss value of the loss function after the training is completed. The lower the loss value, the lower the error rate of the trained neural network 130, and thus the higher the accuracy.

[0091] At the OLT 110, the OLT 110 receives 520 feedback information from the ONU 120, and based on the feedback information, updates 525 the maintained initial parameter value set for the neural network 130. For example, the parameter update module 320 in the OLT 110 is configured to perform an update of the initial parameter value set for the neural network 130. The initial parameter value set currently maintained by the OLT 110 is sometimes referred to herein as a “first initial parameter value set”, and the updated initial parameter value set is sometimes referred to herein as a “second initial parameter value set”.

[0092] According to an embodiment of the present disclosure, OLT 110 can continuously optimize the initial parameter value set of neural network 130 based on the training performance of neural network 130 at one or more ONUs 120. In some embodiments, feedback information from ONU 120, such as the training performance indication therein, helps OLT 110 determine whether the initial parameter value set of neural network 130 should continue to be upgraded.

[0093] Specifically, the OLT 110 may determine whether the training performance at the ONU 120 meets the threshold performance based on the training performance indication. For example, if the training performance indication includes first indication information indicating the time consumption of the training, the OLT 110 may determine whether the time consumption of the training (e.g., the number of epochs) exceeds a predetermined time threshold. Alternatively or additionally, if the training performance indication includes second indication information indicating the error rate of the training, the OLT 110 may determine whether the error rate exceeds a predetermined error threshold. If the OLT 110 determines that the training performance at the ONU 120 does not meet the threshold performance, the OLT 110 determines to continue updating the initial parameter value set.

[0094] In the initial stage, the initial parameter value set at the OLT 110 may be the same as the initial parameter value set used at the initially added ONU 120, for example, a default initial parameter value set, or an initial parameter value set determined via simulation training data. If the OLT 110 determines that the training performance at the ONU 120 does not meet the threshold performance, the current initial parameter value set may not help the ONU 120 to train the neural network 130 faster and / or more reliably, and thus further updating is required.

[0095] In some embodiments, the OLT 110 may determine whether to continue updating the initial parameter value set based on whether the training performance at the plurality of ONUs 120 meets the threshold performance after receiving feedback information on the training of the neural network 130 from the plurality of ONUs 120. For example, if the OLT 110 determines that the training performance at a plurality of ONUs 120 (e.g., at a number of ONUs 120 exceeding a threshold) is not satisfactory, for example, the time consumption and / or the error rate exceed the corresponding threshold, then the OLT 110 may determine to update the initial parameter value set.

[0096] The updating of the initial set of parameter values ​​for neural network 130 will be discussed in more detail below.

[0097] A better initial parameter value set helps to train the neural network 130 more quickly and reliably to meet the expected goal. Therefore, after completing the update of the initial parameter value set, the OLT 110 transmits the updated initial parameter value set (i.e., the "second initial parameter value set") to one or more other ONUs 120. The other ONUs 120 can be, for example, new ONUs 120 that want to establish a communication channel with the OLT 110. These ONUs 120 can use the updated initial parameter value set to train the neural network 130 deployed by themselves, so as to use the trained neural network 130 to process the signal received from the OLT 110.

[0098] According to the embodiments of the present disclosure, as the OLT 110 continuously updates and optimizes the initial parameter value set of the neural network 130, the subsequent ONU 120 can continuously improve the training speed of the neural network 130. The ONU 120 can use a small amount of training data to make fine adjustments based on the initial parameter value set, so that the neural network 130 can reach the training target faster. In this way, both the training data collection and the actual training process can be reduced. It can be understood that the process 500 can be continuously executed for multiple ONUs 120.

[0099] In some embodiments, after completing the training of the neural network 130, the ONU 120 may perform a channel activation phase with the OLT 110 to activate a communication channel for a specific wavelength. The ONU 120 may use the trained neural network 130 to process the signal received from the OLT 110 through the activated communication channel, such as performing equalization processing.

[0100] In some embodiments, after activating the communication channel, the ONU 120 may transmit feedback information on the training of the neural network 130 to the OLT 110 through the activated communication channel.

[0101] In process 500, OLT 110 may transmit the locally maintained initial parameter value set for neural network 130 to ONU 120. For example, OLT 110 transmits the initial parameter value set to ONU 120 at 506. In addition, after updating the initial parameter value set, OLT 110 transmits the updated initial parameter value set to another ONU 120 at 530. When transmitting the initial parameter value set to ONU 120, the communication channel between ONU 120 and OLT 110 to perform signal processing by applying neural network 130 may not be activated. In this case, OLT 110 may transmit the initial parameter value set to ONU 120 through other communication methods.

[0102] Fig. 7A and Figure 7B Example processes 700 and 702 for transmitting an initial parameter value set according to some example embodiments of the present disclosure are shown. Note that in processes 700 and 702, ONU 120 may be any ONU that receives an initial parameter value set from OLT 110 before or after an update.

[0103] exist Fig. 7A In the example process 700 of FIG. 7 , it is assumed that the ONU 120 has a communication channel of multiple wavelengths for communicating with the OLT 110. In the process 700, the ONU 120 performs 705 a channel activation process of a communication channel of a specific wavelength with the OLT 110. The neural network 130 may be trained for processing signals received from the OLT 110 on a communication channel of a first wavelength. And at 705, the ONU 120 may activate a communication channel of a second wavelength with the OLT 110 for receiving an initial parameter value set. The second wavelength is different from the first wavelength. In some embodiments, the communication channel of the second wavelength may have a lower communication rate so that the initial parameter value set can be reliably received.

[0104] In some embodiments, ONU 120 may already have an activated communication channel capable of communicating with OLT 110, and thus the channel activation process in process 700 may be omitted.

[0105] With the communication channel activated, the ONU 120 transmits 710 to the OLT 110, via the activated communication channel, a request for an initial parameter value set for the neural network 130. After receiving 715 the request, the OLT 110 transmits 720 the initial parameter value set to the ONU 120. After receiving 725 the initial parameter value set, the ONU 120 may begin training the neural network 130 based on the initial parameter value set, as described above.

[0106] In some embodiments, OLT 110 can directly carry the initial parameter value set in the periodic downlink control message (such as the sequence number request of ONU, serial number request etc.), so the parameter request process 710, 715 in process 700 can be omitted. In some embodiments, the initial parameter value set can be included in the operation, management and maintenance (PLOAM) message. In some embodiments, the initial parameter value set can be included in the ONU management and control channel (OMCC) message. Of course, other messages can also be used to transmit the initial parameter value set to ONU 120.

[0107] exist Figure 7B In the example process 702 of FIG. 1 , it is assumed that the ONU 120 is communicatively coupled to the auxiliary communication device 730. The auxiliary communication device 730 may also communicate with the OLT 110. The auxiliary communication device 730 is configured to forward the initial parameter value set received from the OLT 110 to the ONU 120.

[0108] Specifically, the auxiliary communication device 730 transmits 735 to the OLT 110 a request for an initial parameter value set for the neural network 130. After receiving 740 the request, the OLT 110 transmits 745 the initial parameter value set to the auxiliary communication device 730. After receiving 750 the initial parameter value set, the auxiliary communication device 730 continues to forward 755 the initial parameter value set to the ONU 120. After receiving 760 the initial parameter value set, the ONU 120 may start training the neural network 130 based on the initial parameter value set as described above.

[0109] In some examples, the auxiliary communication device 730 may be a communication module, such as an Internet of Things (IoT) module, embedded in the ONU 120. The communication module may be implemented in hardware and / or software and communicate with the OLT 110 via other communication technologies, such as communicating with the OLT 110 via mobile Internet.

[0110] In some examples, the auxiliary communication device 730 may include an independent communication device that can be detachably connected to the ONU 120 for communicating therewith. For example, the auxiliary communication device 730 can be coupled to the ONU 120 via a universal serial bus (USB) or other communication interface (e.g., a serial communication interface such as RS232). In actual applications, such an auxiliary communication device 730 can be manually installed to the ONU 120, for example, to obtain an initial parameter value set for the ONU 120.

[0111] It should be understood that the auxiliary communication device 730 may also include any other communication device. The embodiments of the present disclosure are not limited in this regard.

[0112] As mentioned above, OLT 110 updates the initial parameter value set of neural network 130 based on the feedback information from ONU 120. The updating of the parameter set of neural network 130 at OLT 110 is also similar to the training process of neural network, or also called pre-training process, because ONU 120 will further adjust (fine tune) the initial parameter value set obtained by training. The starting point of the training of neural network 130 is the initial parameter value set currently maintained at OLT 110.

[0113] In some embodiments, in order to train the neural network 130, the training data required by the OLT 110, such as the training data 332 in the storage system 330, includes a pilot sequence sent to the ONU 120. In addition, the OLT 110 can also determine the pilot sequence that may be received at each ONU 120 through simulation as the training data 332. For example, the OLT 110 can simulate the communication channel between different ONUs 120, thereby determining the pilot sequence that may be received at each ONU 120 via the simulated communication channel. The OLT 110 can perform an update on the initial parameter value set based on the simulated pilot sequence received at the ONU 120.

[0114] In some embodiments, the feedback information fed back by one or more ONUs 120 to OLT 110 after performing the training of neural network 130 may also include part or all of the training data used to train neural network 130. The training data here includes a pilot sequence received by ONU 120 from OLT 110 via respective communication channels. In this way, OLT 110 can obtain real training data from ONU 120 to more effectively perform the update of the initial parameter value set of neural network 130.

[0115] In some embodiments, the training data fed back from the ONU 120 may be provided to the OLT 110 together with the training performance indication, for example, transmitted to the OLT 110 at 515 in the process 500. In some embodiments, the ONU 120 may transmit the training performance indication and the training data to the OLT 110, respectively. In one embodiment, the OLT 110 may send a request to the ONU 120 that has completed the training of the neural network 130 after determining that the initial parameter value set of the neural network 130 is to be updated, requesting the ONU 120 to transmit the previous update of the initial parameter value of the neural network 130.

[0116] The OLT 110 may collect real training data used by each of the multiple ONUs 120 for updating the initial parameter value set of the neural network 130. In some embodiments, the OLT 110 may train the neural network 130 by utilizing a Model Agnostic Meta Learning (MAML) process, compared to using only training data from one ONU 120, or by mixing training data from different ONUs 120. MAML may combine the learning losses of multiple models learned in different environments (also referred to as tasks), and update the model by calculating the loss with new data, thereby improving the robustness of the model.

[0117] Since the communication channels connected to different ONUs 120 by OLT 110 are different, these different communication channels are equivalent to different environments. The neural network 130 trained at different ONUs 120 actually needs to learn the characteristics of different communication channels respectively. At OLT 110, through MAML, the common characteristics learned by the neural network 130 at different ONUs 120 can be obtained, so that the updated initial parameter value set can better help the neural network 130 to quickly adapt to the new environment, that is, it can be trained faster at the new ONU 120 to meet the expected goals. In addition, the application of MAML enables OLT 110 to collect a small amount of real training data from a single ONU 120 to complete the update of parameter values.

[0118] For a better understanding, let's first briefly introduce the MAML algorithm. When applying MAML to train a neural network, in each iteration of training, training data D and D' of multiple tasks can be randomly selected from the training data corresponding to different tasks. For D of each task Ti, the learning loss is calculated to update the gradient, such as stochastic gradient (GD), as follows:

[0119]

[0120] in represents the gradient of task Ti, θ represents the total gradient of the neural network in this training step, represents the learning loss of task Ti, the loss value of the loss function of the neural network in task Ti, which is determined, for example, by the forward propagation and backward propagation processes after applying the training data of task Ti to the neural network; α represents the learning rate for task Ti, which can be a predetermined value and can be the same or different for each task.

[0121] Determine the gradient of each task in multiple randomly selected tasks (assuming N tasks) After that, the gradients of multiple tasks are recalculated for the D' model to update the total gradient θ of the neural network. For example, such an update can be expressed as:

[0122]

[0123] in Represents the gradient based on task Ti The learning loss is determined, and β represents the learning rate, which can be a predetermined value.

[0124] The total gradient θ of the entire neural network is iteratively updated in multiple training steps. After such iterative updates reach the expected goal, for example, when the number of iterations reaches a threshold or the loss value of the neural network is less than a predetermined threshold, the update of the initial parameter value set of the neural network is completed.

[0125] In some embodiments, the update of the initial parameter value set of the neural network 130 by the OLT 110 may include an update using simulated training data and an update using real training data obtained from the ONU 120. For example, the OLT 110 may also use the MAML process to train the neural network 130 using simulated training data and real training data.

[0126] Figure 8 FIG. 8 is a flowchart showing a process 800 for updating an initial parameter value set of a neural network according to some example embodiments of the present disclosure. The process 800 may be implemented by the OLT 110, for example, by Figure 2 The parameter updating module 320 in the OLT 110 is implemented.

[0127] At block 810, in the initial stage, OLT 110 randomly initializes a parameter value set of neural network 130 to obtain a random initial parameter value set. Based on the random initial parameter value set, OLT 110 performs a first training phase 802 based on simulation training data obtained by simulating different communication channels.

[0128] Specifically, in block 815, OLT 110 randomly selects simulation training data of N tasks from M simulation tasks for training neural network 130. M and N are positive integers, and M is greater than N. Here, the simulation task refers to the training of neural network 130 performed by simulated ONU 120 for different communication channels, wherein the simulation training data is the pilot sequence sent by OLT 110 and may be received at ONU 120 after passing through the simulated communication channel.

[0129] At 820, OLT 110 calculates the gradient of each of the N tasks As shown in the above formula (2). In box 825, OLT 110 updates the total gradient θ of neural network 130. In box 830, OLT 110 determines whether the number of iterations is greater than a predetermined number threshold (denoted as L), or whether the loss value of the loss function is less than a predetermined loss threshold (denoted as ε). If the number of iterations is not greater than the predetermined number threshold and the loss value is not less than the predetermined loss threshold, OLT 110 performs the next iteration step, including performing the operations of boxes 815, 820 and 825.

[0130] If the number of iterations is greater than a predetermined number threshold L, or if the loss value of the loss function is less than a predetermined loss threshold ε, the OLT 110 performs a second training phase 804 based on real training data obtained from the ONU 120. Specifically, in block 835, the OLT 110 randomly selects real training data of N' tasks from the M' real tasks for training the neural network 130. M and N are positive integers, and M' is greater than N'. Here, the real task refers to the training of the neural network 130 at a certain ONU 120. The real training data includes training data obtained from the ONU 120, such as a pilot sequence received by the ONU 120. Since the OLT 110 may continue to receive training data from more ONUs 120, the value of M' will continue to increase.

[0131] At block 840, OLT 110 calculates the gradient of each of the N' tasks As shown in the above formula (2). In box 845, OLT 110 updates the total gradient θ of neural network 130. In box 850, OLT 110 determines whether the number of iterations is greater than a predetermined number threshold (denoted as L'), or whether the loss value of the loss function is less than a predetermined loss threshold (denoted as ε'). If the number of iterations is not greater than the predetermined number threshold and the loss value is not less than the predetermined loss threshold, OLT 110 performs the next iteration step, including performing the operations of boxes 835, 840 and 845.

[0132] If the number of iterations is greater than a predetermined number threshold L′, or if the loss value of the loss function is less than a predetermined loss threshold ε′, in block 855 , the OLT 110 obtains an updated set of initial parameter values ​​of the neural network 130 .

[0133] In some embodiments, process 800 for updating the initial set of parameter values ​​for neural network 130 may be performed periodically or in response to a particular event. For example, as mentioned above, upon determining that the training performance indication received from one or more ONUs 120 does not meet a performance threshold, OLT 110 may determine to perform process 800 to update the currently maintained initial set of parameter values.

[0134] It should be understood that Figure 8 This is only an example process of updating the initial parameter value set. In some embodiments, OLT 110 may directly perform training on neural network 130 based on real training data from ONU 120 to update the initial parameter value set. In some embodiments, OLT 110 may utilize other meta-learning (metalearning) or other training algorithms other than MAML to perform training on neural network 130. The embodiments of the present disclosure are not limited in this respect.

[0135] Fig. 9 A comparison chart of average test errors of example training according to the present disclosure and traditional training is shown, wherein the horizontal axis represents the number of training steps performed using the SGD algorithm during the training phase, and the vertical axis represents the average test error rate, such as BER. Fig. 9 Curve 910 shows the average test error rate of the neural network 130 trained according to the example of the present disclosure as a function of the number of training steps, curve 920 shows the average test error rate of the neural network 130 obtained through traditional joint training as a function of the number of training steps, and curve 930 shows the average test error rate of the neural network 130 obtained through traditional self-training as a function of the number of training steps. It is assumed that both traditional joint training and self-training are performed based on a randomly initialized parameter set.

[0136] By comparing curve 910 with curves 920 and 930, it can be seen that according to the example implementation of training disclosed in the present invention, neural network 130 can achieve a lower error rate at a faster speed (e.g., fewer training steps). In addition, the example implementation of training according to the present invention can achieve a maximum accuracy of 0.9922 after 50,000 sequence steps, which is greater than the accuracy of traditional joint training and traditional self-training.

[0137] Fig.10 A comparison chart of average test errors of two example trainings according to some example embodiments of the present disclosure is shown. Curve 1010 represents the variation of the average test error rate of the neural network 130 obtained by training based on simulated training data and real training data (such as the training shown in process 800) with the number of training steps in some embodiments of the present disclosure, and curve 1020 represents the variation of the average test error rate of the neural network 130 obtained by training only based on simulated training data with the number of training steps in some embodiments of the present disclosure. By comparison, it can be seen that by training based on simulated training data and real training data, the neural network 130 can achieve a lower error rate at a faster speed (for example, fewer training steps). For example, Fig.10As shown by the arrow, in the example of training based on simulated training data and real training data, the neural network 130 can obtain an error rate less than or equal to 0.01 after 551 training steps, and in the example of training based on simulated training data, the neural network 130 can obtain an error rate less than or equal to 0.01 after 35928 training steps.

[0138] Fig.11 A comparison chart of the average test error of another example training according to the present disclosure and traditional training is shown. Curve 1110 represents the average test error rate of the neural network 130 obtained by training based on simulated training data and real training data (such as the training shown in process 800) in some embodiments of the present disclosure, and curve 1120 shows the average test error rate of the neural network 130 obtained by traditional self-training. Changes with the number of training steps. It can be seen that according to the example implementation of training disclosed in the present disclosure, the neural network 130 can reach a lower error rate at a faster speed (for example, fewer training steps), and obtain an error rate of less than or equal to 0.01 after 551 training steps. Traditional self-training may require more than 200,000 training steps to reach an error rate of 0.01. The maximum accuracy rate that can be achieved by the example implementation of training according to the present disclosure is 0.9974, which is greater than the accuracy rate of traditional self-training 0.9898.

[0139] Fig.12 A flowchart of a method 1200 implemented at a first optical communication device according to some example embodiments of the present disclosure is shown. The first optical communication device may be, for example, Figure 3 OLT 110.

[0140] like Fig.12 As shown, in block 1210, the first optical communication device receives feedback information for training of a neural network from at least one second optical communication device (e.g., ONU 120). The neural network is configured to process a signal received from the first optical communication device. The feedback information includes at least a training performance indication of each training of the neural network at the at least one second optical communication device.

[0141] At block 1220, the first optical communication device updates a first initial parameter value set for the neural network maintained at the first optical communication device based on the feedback information to obtain a second initial parameter value set for the neural network. At block 1230, the first optical communication device transmits the second initial parameter value set to at least one other second optical communication device for training the neural network based on the second initial parameter value set at the at least one other second optical communication device.

[0142] In some embodiments, the method 1200 further includes: transmitting a first initial parameter value set for the neural network from the first optical communication device to at least one second optical communication device. In some embodiments, the feedback information indicates a training performance of the neural network trained based on the first initial parameter value set at the at least one second optical communication device.

[0143] In some embodiments, the first initial parameter value set or the second initial parameter value set is transmitted to the at least one second optical communication device or the at least one further second optical communication device, respectively, via at least one of: an activated communication channel, and forwarding by an auxiliary communication device, the auxiliary communication device being communicatively coupled to the at least one second optical communication device or the at least one further second optical communication device.

[0144] In some embodiments, the training performance indication includes at least one of the following: first indication information indicating the time consumption to complete the training of the neural network at each of the at least one second optical communication devices, and second indication information indicating the error rate of the neural network after the training is completed at each of the at least one second optical communication devices.

[0145] In some embodiments, updating the first initial parameter value set based on the feedback information includes: determining whether training performance at at least one second optical communication device meets threshold performance based on the training performance indication; and updating the first initial parameter value set based on the training performance not meeting the threshold performance.

[0146] In some embodiments, the feedback information further includes at least a portion of training data used by the at least one second optical communication device to train the neural network. In some embodiments, updating the first set of initial parameter values ​​based on the training performance not satisfying the threshold performance includes: if it is determined that the training performance does not satisfy the threshold performance, training the neural network based on at least a portion of the training data and the first set of initial parameter values ​​to update the first set of initial parameter values ​​to a second set of initial parameter values.

[0147] In some embodiments, the feedback information further includes at least a portion of training data respectively used by the plurality of second optical communication devices to train the neural network. In some embodiments, training the neural network includes: training the neural network using a model-independent meta-learning (MAML) process based on at least a portion of the training data respectively used by the plurality of second optical communication devices to train the neural network.

[0148] In some embodiments, the method 1200 further includes: transmitting a plurality of pilot sequences to the at least one second optical communication device to be used as training data for the neural network at the at least one second optical communication device.

[0149] In some embodiments, transmitting the plurality of pilot sequences comprises: carrying the plurality of pilot sequences respectively in predetermined sections of payload parts of respective plurality of messages having a predetermined frame structure; and transmitting the plurality of messages to at least one second optical communication device.

[0150] In some embodiments, the neural network is configured to perform equalization processing on a signal received from the first optical communication device.

[0151] In some embodiments, the first optical communication device comprises an optical line terminal, and at least one of the at least one second optical communication device and the at least one further second optical communication device comprises an optical network unit.

[0152] Fig.13 A flowchart of a method 1300 implemented at a second optical communication device according to some example embodiments of the present disclosure is shown. The second optical communication device may be, for example, Figure 3 ONU 120.

[0153] like Fig.13 As shown, at block 1310, the second optical communication device trains a neural network, and the neural network is configured to process a signal received from the first optical communication device (e.g., OLT 110). At block 1320, the second optical communication device transmits feedback information for the training of the neural network to the first optical communication device, and the feedback information includes at least a training performance indication of the training of the neural network at the second optical communication device.

[0154] In some embodiments, the method 1300 further comprises: receiving a first initial parameter value set for the neural network from the first optical communication device. In some embodiments, training the neural network comprises: training the neural network based on the first initial parameter value set.

[0155] In some embodiments, the method 1300 further includes: receiving a second initial parameter value set for the neural network from the first optical communication device, the second initial parameter value set being determined by the first optical communication device after updating the first initial parameter value set for the neural network based on feedback information received from at least one other optical communication device. In some embodiments, training the neural network includes: training the neural network based on the second initial parameter value set.

[0156] In some embodiments, the first initial set of parameter values ​​or the second initial set of parameter values ​​are received via at least one of: an activated communication channel, and forwarding by an auxiliary communication device communicatively coupled to the second optical communication device.

[0157] In some embodiments, the training performance indication includes at least one of the following: first indication information indicating the time consumption to complete the training of the neural network at the second optical communication device, and second indication information indicating the error rate of the neural network after the training is completed at the second optical communication device.

[0158] In some embodiments, the feedback information also includes at least a portion of training data used by the second optical communication device to train the neural network.

[0159] In some embodiments, training the neural network includes: receiving a plurality of pilot sequences from the first optical communication device; and training the neural network by using the received plurality of pilot sequences as training data.

[0160] In some embodiments, receiving the plurality of pilot sequences includes: receiving a plurality of messages having a predetermined frame structure from the first optical communication device; and extracting the plurality of pilot sequences from predetermined sections of payload parts of the respective plurality of messages.

[0161] In some embodiments, the neural network is configured to perform equalization processing on a signal received from the first optical communication device.

[0162] In some embodiments, the first optical communication device comprises an optical line terminal and the second optical communication device comprises an optical network unit.

[0163] Fig.14 1 is a simplified block diagram of a device 1400 suitable for implementing embodiments of the present disclosure. The device 1400 may be provided to implement a communication device, such as Figure 1 and Figure 3 An OLT 110 and an ONU 120 are shown. As shown, the device 1400 includes one or more processors 1410, one or more memories 1440 coupled to the processor 1410, and one or more transmitters and / or receivers (TX / RX) 1440 coupled to the processor 1410.

[0164] TX / RX 1440 is used for bidirectional communication. TX / RX 1440 has at least one antenna to facilitate communication. The communication interface may represent any interface necessary to communicate with other network elements.

[0165] Processor 1410 may be of any type suitable for the local technology network and may include, but is not limited to, one or more of a general purpose computer, a special purpose computer, a microcontroller, a digital signal controller (DSP), and a controller-based multi-core controller architecture. Device 1400 may have multiple processors, such as application specific integrated circuit chips, which are time-slaved to a clock synchronized with a main processor.

[0166] The memory 1420 may include one or more non-volatile memories and one or more volatile memories. Examples of non-volatile memories include, but are not limited to, read-only memory (ROM) 1424, erasable programmable read-only memory (EPROM), flash memory, hard disk, compact disk (CD), digital video disk (DVD), and other magnetic storage and / or optical storage. Examples of volatile memories include, but are not limited to, random access memory (RAM) 1422 and other volatile memories that do not persist over the duration of a power outage.

[0167] Computer program 1430 includes computer executable instructions executed by associated processor 1410. Program 1430 may be stored in ROM 1420. Processor 1410 may perform any suitable actions and processes by loading program 1430 into RAM 1420.

[0168] The embodiment of the present disclosure can be implemented with the help of program 1430, so that the device 1400 can execute the Figures 3 to 13 Any process of the present disclosure discussed. The embodiments of the present disclosure may also be implemented by hardware or by a combination of software and hardware.

[0169] In some example embodiments, the program 1430 may be tangibly embodied in a computer-readable medium that may be included in the device 1400 (such as in the memory 1420) or other storage device accessible by the device 1400. The program 1430 may be loaded from the computer-readable medium to the RAM 1422 for execution. The computer-readable medium may include any type of tangible non-volatile memory, such as ROM, EPROM, flash memory, hard disk, CD, DVD, etc. Fig.15 An example of a computer readable medium 700 in the form of a CD or DVD is shown. The computer readable medium has a program 1430 stored thereon.

[0170] In general, various embodiments of the present disclosure may be implemented in hardware or dedicated circuits, software, logic, or any combination thereof. Some aspects may be implemented in hardware, while other aspects may be implemented in firmware or software, which may be performed by a controller, microprocessor, or other computing device. Although various aspects of the embodiments of the present disclosure are shown and described as block diagrams, flow charts, or using some other graphical representation, it should be understood that the blocks, devices, systems, techniques, or methods described herein may be implemented as, by way of non-limiting example, hardware, software, firmware, dedicated circuits or logic, general purpose hardware or a controller or other computing device, or some combination thereof.

[0171] The present disclosure also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer executable instructions, such as instructions included in a program module, which are executed in a device on a real or virtual processor of a target to perform the above-referenced Figure 3 and Figure 4 Methods 300 and 400 are described. Generally, program modules include routines, programs, libraries, objects, classes, components, data structures, etc. that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided between program modules as needed. Machine executable instructions for program modules can be executed in local or distributed devices. In distributed devices, program modules can be located in local and remote storage media.

[0172] The computer program code for realizing the method of the present disclosure can be written in one or more programming languages. These computer program codes can be provided to the processor of general-purpose computer, special-purpose computer or other programmable data processing device, so that the program code, when being executed by computer or other programmable data processing device, causes the function / operation specified in flow chart and / or block diagram to be implemented. The program code can be executed completely on computer, partly on computer, as independent software package, partly on computer and partly on remote computer or completely on remote computer or server.

[0173] In the context of the present disclosure, computer program codes or related data may be carried by any appropriate carrier to enable a device, apparatus or processor to perform the various processes and operations described above. Examples of carriers include signals, computer readable media, and the like. Examples of signals may include electrical, optical, radio, acoustic or other forms of propagation signals, such as carrier waves, infrared signals, and the like.

[0174] A computer readable medium may be any tangible medium that contains or stores a program for or related to an instruction execution system, apparatus, or device. A computer readable medium may be a computer readable signal medium or a computer readable storage medium. A computer readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination thereof. More detailed examples of computer readable storage media include an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0175] In addition, although the operation of the method of the present disclosure is described in a particular order in the accompanying drawings, this does not require or imply that these operations must be performed in this particular order, or that all the operations shown must be performed to achieve the desired result. On the contrary, the steps depicted in the flow chart can change the order of execution. Additionally or alternatively, some steps can be omitted, multiple steps can be combined into one step for execution, and / or one step can be decomposed into multiple steps for execution. It should also be noted that the features and functions of two or more devices according to the present disclosure can be embodied in one device. Conversely, the features and functions of a device described above can be further divided into being embodied by multiple devices.

[0176] Although the present disclosure has been described with reference to several specific embodiments, it should be understood that the present disclosure is not limited to the specific embodiments disclosed. The present disclosure is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A method for optical communication, comprising: receiving, at a first optical communication device, feedback information for training of a neural network from at least one second optical communication device, the neural network being configured to process a signal received from the first optical communication device, the feedback information comprising at least a training performance indication of training the neural network at each of the at least one second optical communication device; updating a first set of initial parameter values ​​for the neural network maintained at the first optical communication device based on the feedback information to obtain a second set of initial parameter values ​​for the neural network; as well as The second initial parameter value set is transmitted to at least one further second optical communication device for use in training the neural network based on the second initial parameter value set at the at least one further second optical communication device.

2. The method according to claim 1, further comprising: transmitting the first set of initial parameter values ​​for the neural network from the first optical communication device to the at least one second optical communication device, and The feedback information indicates training performance of the neural network trained based on the first initial parameter value set at the at least one second optical communication device.

3. The method according to claim 2, wherein the first initial parameter value set or the second initial parameter value set is transmitted to the at least one second optical communication device or the at least one further second optical communication device, respectively, via at least one of the following: The activated communication channels, and Relayed by an auxiliary communication device, the auxiliary communication device being communicatively coupled to the at least one second optical communication device or the at least one further second optical communication device.

4. The method of claim 1, wherein the training performance indicator comprises at least one of: first indication information indicating the time consumption for each of the at least one second optical communication device to complete the training of the neural network, and The second indication information indicates the error rate of the neural network after the training is completed at each of the at least one second optical communication devices.

5. The method according to claim 1, wherein updating the first set of initial parameter values ​​based on the feedback information comprises: determining whether training performance at the at least one second optical communication device satisfies a threshold performance based on the training performance indication; as well as Based on the training performance not satisfying the threshold performance, the first initial parameter value set is updated.

6. The method of claim 5, wherein the feedback information further comprises at least a portion of training data used by the at least one second optical communication device to train the neural network, and updating the first set of initial parameter values ​​based on the training performance not satisfying the threshold performance comprises: If it is determined that the training performance does not meet the threshold performance, the neural network is trained based on the at least a portion of the training data and the first initial parameter value set to update the first initial parameter value set to the second initial parameter value set.

7. The method according to claim 6, wherein the feedback information further comprises at least a portion of training data respectively used by a plurality of second optical communication devices to train the neural network, and training the neural network comprises: The neural network is trained using a model-independent meta-learning (MAML) process based on at least a portion of the training data respectively used by a plurality of second optical communication devices to train the neural network.

8. The method according to claim 1, further comprising: A plurality of pilot sequences are transmitted to the at least one second optical communication device to be used as training data for the neural network at the at least one second optical communication device.

9. The method of claim 8, wherein transmitting the plurality of pilot sequences comprises: Carrying the plurality of pilot sequences in predetermined sections of respective payload parts of a plurality of messages having a predetermined frame structure respectively; as well as The plurality of messages are transmitted to the at least one second optical communication device.

10. The method according to any one of claims 1 to 9, wherein the neural network is configured to perform equalization processing on a signal received from the first optical communication device.

11. The method according to any one of claims 1 to 9, wherein the first optical communication device comprises an optical line terminal, and at least one of the at least one second optical communication device and the at least one further second optical communication device comprises an optical network unit.

12. A method for optical communication, comprising: training a neural network at a second optical communication device, the neural network configured to process a signal received from the first optical communication device; as well as Transmitting feedback information for the training of the neural network to the first optical communication device, the feedback information at least including a training performance indication of the training of the neural network at the second optical communication device.

13. The method according to claim 12, further comprising: receiving a first set of initial parameter values ​​for the neural network from the first optical communication device, and in Training the neural network includes training the neural network based on the first set of initial parameter values.

14. The method according to claim 13, further comprising: receiving a second initial parameter value set for the neural network from the first optical communication device, the second initial parameter value set being determined by the first optical communication device after updating a first initial parameter value set for the neural network based on feedback information received from at least one other optical communication device, and Wherein training the neural network comprises: training the neural network based on the second initial parameter value set.

15. The method of claim 14, wherein the first initial parameter value set or the second initial parameter value set is received via at least one of: The activated communication channels, and forwarded by a secondary communication device communicatively coupled to the second optical communication device.

16. The method of claim 12, wherein the training performance indicator comprises at least one of: first indication information indicating the time consumption for completing the training of the neural network at each of the second optical communication devices, and The second indication information indicates the error rate of the neural network after the training is completed at the second optical communication device.

17. The method of claim 12, wherein the feedback information further comprises at least a portion of training data used by the second optical communication device to train the neural network.

18. The method of claim 12, wherein training the neural network comprises: receiving a plurality of pilot sequences from the first optical communication device; as well as The neural network is trained by using the received multiple pilot sequences as training data.

19. The method of claim 18, wherein receiving the plurality of pilot sequences comprises: receiving a plurality of messages having a predetermined frame structure from the first optical communication device; as well as The plurality of pilot sequences are extracted in predetermined sections of payload parts of respective ones of the plurality of messages.

20. The method of claim 12, wherein the neural network is configured to perform equalization processing on a signal received from the first optical communication device.

21. The method of claim 12, wherein the first optical communication device comprises an optical line terminal, and the second optical communication device comprises an optical network unit.

22. A first optical communication device, comprising: at least one processor; as well as at least one memory coupled to the at least one processor, the at least one memory including instructions stored therein; The at least one memory and the instructions are further configured to, together with the at least one processor, cause the first optical communication device to perform the method according to any one of claims 1 to 11.

23. A second optical communication device, comprising: at least one processor; as well as at least one memory coupled to the at least one processor, the at least one memory including instructions stored therein; The at least one memory and the instructions are further configured to, together with the at least one processor, cause the second optical communication device to perform the method according to any one of claims 12 to 21.

24. A first optical communication device, comprising: for receiving feedback information for training of a neural network from at least one second optical communication device, the neural network being configured to process a signal received from the first optical communication device, the feedback information comprising at least a training performance indication of training the neural network at each of the at least one second optical communication device; means for updating a first set of initial parameter values ​​for the neural network maintained at the first optical communication device based on the feedback information to obtain a second set of initial parameter values ​​for the neural network; as well as Means for transmitting the second set of initial parameter values ​​to at least one further second optical communication device for training the neural network based on the second set of initial parameter values ​​at the at least one further second optical communication device.

25. A second optical communication device, comprising: means for training a neural network configured to process signals received from a first optical communication device; as well as means for transmitting feedback information for the training of the neural network to the first optical communication device, the feedback information comprising at least an indication of training performance of the training of the neural network at the second optical communication device.

26. A computer-readable medium having instructions stored thereon, which, when executed by at least one processing unit, causes the at least one processing unit to be configured to perform the method according to any one of claims 1 to 11.

27. A computer-readable medium having instructions stored thereon, which, when executed by at least one processing unit, causes the at least one processing unit to be configured to perform the method according to any one of claims 12 to 21.