WO2022001822A1 - 获取神经网络的方法和装置 - Google Patents

获取神经网络的方法和装置 Download PDF

Info

Publication number
WO2022001822A1
WO2022001822A1 PCT/CN2021/102079 CN2021102079W WO2022001822A1 WO 2022001822 A1 WO2022001822 A1 WO 2022001822A1 CN 2021102079 W CN2021102079 W CN 2021102079W WO 2022001822 A1 WO2022001822 A1 WO 2022001822A1
Authority
WO
WIPO (PCT)
Prior art keywords
neural network
originating
performance
target
output
Prior art date
Application number
PCT/CN2021/102079
Other languages
English (en)
French (fr)
Inventor
梁璟
毕晓艳
刘永
Original Assignee
华为技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 华为技术有限公司 filed Critical 华为技术有限公司
Priority to EP21833854.9A priority Critical patent/EP4163827A4/en
Publication of WO2022001822A1 publication Critical patent/WO2022001822A1/zh
Priority to US18/147,111 priority patent/US20230136416A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W24/00—Supervisory, monitoring or testing arrangements
    • H04W24/02—Arrangements for optimising operational condition
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F11/00—Error detection; Error correction; Monitoring
    • G06F11/30—Monitoring
    • G06F11/34—Recording or statistical evaluation of computer activity, e.g. of down time, of input/output operation ; Recording or statistical evaluation of user activity, e.g. usability assessment
    • G06F11/3466—Performance evaluation by tracing or monitoring
    • G06F11/3495—Performance evaluation by tracing or monitoring for systems
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/096—Transfer learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/0985—Hyperparameter optimisation; Meta-learning; Learning-to-learn
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/04—Network management architectures or arrangements
    • H04L41/044—Network management architectures or arrangements comprising hierarchical management structures
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/08—Configuration management of networks or network elements
    • H04L41/0803—Configuration setting
    • H04L41/0823—Configuration setting characterised by the purposes of a change of settings, e.g. optimising configuration for enhancing reliability
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/14—Network analysis or design
    • H04L41/145—Network analysis or design involving simulating, designing, planning or modelling of a network
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/16—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks using machine learning or artificial intelligence
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0495—Quantised networks; Sparse networks; Compressed networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/08—Configuration management of networks or network elements
    • H04L41/085—Retrieval of network configuration; Tracking network configuration history
    • H04L41/0853—Retrieval of network configuration; Tracking network configuration history by actively collecting configuration information or by backing up configuration information

Definitions

  • the present application relates to the field of artificial intelligence, and more particularly, to methods and apparatus for acquiring neural networks.
  • a transmitter is used to transmit signals, and a receiver is used to receive signals transmitted by the transmitter.
  • Neural networks can be applied to transmitters and receivers to jointly optimize the transceiver to improve the overall performance.
  • the transmitter can compress CSI based on a neural network to generate air interface information.
  • the receiver can parse the air interface information based on the neural network to reconstruct the channel.
  • the joint optimization of the algorithm can be realized by using the joint design of the neural network of the transmitter and the neural network of the receiver to obtain the optimal compression and reconstruction algorithm.
  • receivers and transmitters are often provided by different manufacturers.
  • the transmitter is provided by manufacturer A and the receiver is provided by manufacturer B.
  • an implementation method is that the transmitter and the receiver download the neural network structure and parameter information that match the current usage scenario, respectively, and then perform processing at the receiving and sending ends according to the downloaded neural network. This will cause huge overhead of the air interface, which is greatly limited in actual use.
  • the present application provides a method and apparatus for acquiring a neural network, in order to reduce the huge air interface overhead caused by the transmitter and the receiver downloading the neural network structure and parameter information.
  • a method for obtaining a neural network includes: based on a standardized reference data set, one or more reference neural networks and performance judgment criteria, determining whether the constructed target neural network satisfies a preset output condition; if the output condition is satisfied, outputting the The target neural network, the output target neural network is the originating target neural network applied to the transmitter or the receiving end target neural network applied to the receiver.
  • the acquisition device can evaluate the target neural network based on the existing standardized reference neural network.
  • the output is used only when the output conditions are met.
  • the output target neural network can be applied in the transmitter or receiver. Therefore, the overhead of downloading the neural network structure and parameter information by the transmitter or the receiver through the air interface can be avoided.
  • it can also ensure that the products of different manufacturers can be interconnected.
  • On the premise of ensuring performance it can also reflect the differentiated design and competitiveness of products of different manufacturers. Therefore, on the whole, it greatly improves the feasibility of dual network implementation. sex.
  • the method further includes: under the condition that the output condition is not satisfied, continue to optimize the target neural network until the obtained target neural network The network satisfies the output condition.
  • the target neural network can be optimized continuously.
  • the optimization may include, but is not limited to, adjusting the structural design of the neural network, adjusting the reference data set, adjusting the training method of the neural network, adjusting the cost function, the definition of the loss function or the objective function, changing the initialization method of the neural network, changing the parameters of the neural network. Constraints, and changing the definition of the activation function of the neural network, etc.
  • the obtained optimized target neural network can continue to be evaluated based on the above performance judgment criteria, and this cycle guides the optimized target neural network to satisfy the output conditions, and then outputs it.
  • the target neural network at the originating end is used for constellation modulation, channel coding, precoding, or channel information extraction and compression; the target neural network at the receiving end is used for extraction and compression of channel information; Used for constellation demodulation, channel decoding, detection reception, or reconstruction of channel information.
  • the target neural network can be applied to various stages of communication.
  • the originating target neural network and the terminating target neural network can be coupled to each other and work together.
  • the cooperating originating target neural network and terminating target neural network can be used for constellation modulation and demodulation, or channel coding and decoding, or precoding and detection reception, or channel information extraction and compression, and channel information reuse, respectively. structure.
  • the method further includes: determining a configuration suitable for the current application scenario of the target neural network from at least one set of standardized configurations;
  • the configuration includes at least one or more of the following: a standardized reference data set, a standardized reference neural network, a standardized data structure, a standardized performance measurement function, a standardized performance judgment criterion, and a standardized achievable capability measurement value.
  • the reference neural network includes a receiving end reference neural network applied in the receiver and/or a transmitting end reference neural network applied in the transmitter; the data structure is used for generating and/or parsing air interface information; the performance The metric function is used to generate a performance metric value, the performance judgment criterion includes a performance index and an output condition of the target neural network, and the performance judgment criterion is used to determine the target based on the performance metric value and the performance index Whether the neural network satisfies the output condition; the achievable capability metric includes a neural network used by the neural network to select a matching neural network and/or a matching performance index.
  • the matching neural network described herein may specifically refer to that the neural network can be applied in a transmitter or a receiver under given constraints, and the given constraints may include implementation complexity and/or or air interface overhead.
  • the matched performance index may specifically refer to the performance index used to evaluate the neural network in the stage of obtaining the neural network.
  • Different standardized configurations can be defined for different application scenarios.
  • the performance judgment criteria may be different; or, the reference neural network may be different, and so on.
  • the output target neural network can be matched with different application scenarios, which is beneficial to obtain better system performance.
  • the performance metric function includes one or more of the following: mean square error, normalized mean square error, mean absolute error, maximum absolute error, correlation coefficient, cross entropy, mutual information, bit error rate, or frame error rate.
  • performance measurement functions listed above are only examples and should not constitute any limitation to the present application.
  • the performance metric function is used to evaluate the target neural network, one or more of them can be used for evaluation at the same time. This application does not limit this.
  • the performance indicator includes one or more performance indicators
  • the performance judgment criterion includes determining whether the output condition is satisfied based on a comparison between the performance measurement value and the one or more performance indicators.
  • performance indicators may be one or more.
  • the upper bound and/or lower bound of the corresponding performance index may also be one or more.
  • the following shows several examples of evaluating the target neural network to determine whether it meets the output conditions in combination with the above performance judgment criteria.
  • the target neural network is an originating target neural network
  • the one or more reference neural networks are one or more sinking reference neural networks
  • the standardized reference data set, One or more reference neural networks and performance judgment criteria to determine whether the designed target neural network meets preset output conditions including: inputting the originating data obtained from the reference data set into the originating target neural network , the originating target neural network is used to process the originating data according to a predefined data structure to generate air interface information to be sent to the one or more receiving-end reference neural networks;
  • the receiving end reference neural network obtains one or more output results, and the one or more output results are obtained by the one or more receiving end reference neural networks respectively based on the received air interface information;
  • the output results and the originating data are used as the input of the performance measurement function to obtain one or more performance measurement values; according to the one or more performance measurement values and the performance judgment criterion, determine the target neural network Whether the network satisfies the output condition.
  • the target neural network at the receiving end is evaluated by cascading the target neural network at the sending end and the reference neural network at the receiving end.
  • the target neural network is an originating target neural network
  • the one or more reference neural networks are one or more originating reference neural networks
  • the standardized reference dataset, One or more reference neural networks and performance judgment criteria to determine whether the designed target neural network meets preset output conditions including: inputting the originating data obtained from the reference data set into the originating target neural network , the originating target neural network is used to process the originating data according to a predefined data structure to generate target air interface information; input the originating data into the one or more originating reference neural networks, the One or more originating reference neural networks are used to process the originating data according to a predefined data structure to generate one or more reference air interface information; the one or more reference air interface information and the target air interface information are combined As the input of the performance measurement function, to obtain one or more performance measurement values; according to the one or more performance measurement values and the performance judgment criterion, determine whether the originating target neural network satisfies the output condition .
  • the originating target neural network is evaluated using the originating reference neural network as a reference.
  • the target neural network is a destination target neural network
  • the one or more reference neural networks are one or more originating reference neural networks
  • the standardized reference data set is based on , one or more reference neural networks and performance judgment criteria, and determining whether the designed target neural network satisfies the preset output conditions, including: inputting the originating data obtained from the reference data set to the one or more In the originating reference neural network, the one or more originating reference neural networks are used to process the originating data according to a predefined data structure to obtain one or more air interfaces to be sent to the destination target neural network information; obtain one or more output results from the target neural network at the receiving end, where the one or more output results are the outputs generated by the target neural network at the receiving end according to the received one or more air interface information respectively result; using the one or more output results and the originating data as the input of the performance metric function to obtain one or more performance metric values; according to the one or more performance metric values, and the performance A judgment criterion is used
  • the end target neural network and the source reference neural network are cascaded to evaluate the end target neural network.
  • the target neural network is a target neural network at the receiving end
  • the one or more reference neural networks include one or more reference neural networks at the sending end and one or more reference neural networks at the receiving end;
  • determining whether the designed target neural network satisfies the preset output conditions based on the standardized reference data set, one or more reference neural networks and performance judgment criteria includes: obtaining from the reference data set The originating data is input into the one or more originating reference neural networks, and the one or more originating reference neural networks are used to process the originating data according to a predefined data structure, so as to obtain the data to be sent to the receiver.
  • One or more air interface information of the end target neural network and the one or more receiving end reference neural networks One or more target output results are obtained from the receiving end target neural network, the one or more target output results is the output result respectively generated by the target neural network at the receiving end according to the received information of the one or more air interfaces; obtains one or more reference output results from the one or more reference neural networks at the receiving end, and the one or more reference output results are output results generated by the one or more receiving end reference neural networks according to the received one or more air interface information; the one or more target output results and the One or more reference output results are used as the input of the performance metric function to obtain one or more performance metric values; according to the one or more performance metric values and the performance judgment criterion, determine the destination target Whether the neural network satisfies the output condition.
  • the target neural network at the receiving end and the reference neural network at the sending end are cascaded, and the target neural network at the receiving end is used as a reference to evaluate the target neural network at the receiving end.
  • the achievable capability metric value includes one or more of the following: the overhead of air interface information, the amount of computation required for neural network implementation, and the required amount of neural network implementation. The amount of parameters stored, and the computational precision required for the neural network implementation.
  • matching neural networks can be selected to work, or matching performance indicators can be selected for evaluation.
  • a communication method comprising: a first device determining a scene mode based on air interface transmission conditions with the second device, where the scene mode is used to determine a neural network suitable for current air interface transmission conditions , the neural network is determined from a plurality of pre-designed neural networks, each neural network in the plurality of neural networks is matched with one or more scene patterns; the first device reports to the The second device sends the indication information of the scene mode.
  • the first device can select an appropriate scene mode based on the current air interface transmission conditions with the second device, and indicate the scene mode to the second device, so that both the first device and the second device are based on the same Scenario mode to determine what kind of neural network to use. This will help to improve performance.
  • the method further includes: receiving, by the first device, confirmation information from the second device, where the confirmation information is used to indicate that the Successful reception of the indication of the scene mode.
  • the second device may send confirmation information to the first device based on the successful reception of the above-mentioned indication information, so that the first device can perform subsequent procedures according to the reception of the indication information.
  • a neural network matching the scene pattern is used to work; in the case of unsuccessful reception, the indication information of the scene pattern is resent, and so on. This application does not limit this.
  • the method further includes: the first device communicates with the second device using a neural network matching the scene pattern.
  • the first device may use a neural network matching the scene pattern to work based on the successful reception of the above-mentioned indication information by the second device.
  • the second device can also determine and use a neural network matching the indicated scene pattern to work based on the successful reception of the above-mentioned indication information, so that two mutually coupled neural networks can work together, which is beneficial to improve system performance.
  • a communication method comprising: a second device receiving indication information of a scene mode from a first device, where the scene mode is used to determine a neural network suitable for current air interface transmission conditions, the neural network The network is determined from a plurality of pre-designed neural networks, each neural network of the plurality of neural networks is matched with one or more scene patterns; the second device is based on the indication of the scene patterns information, determine the scene mode; according to the scene mode, determine the neural network suitable for the air interface transmission condition.
  • the first device can select an appropriate scene mode based on the current air interface transmission conditions with the second device, and indicate the scene mode to the second device, so that both the first device and the second device are based on the same Scenario mode to determine what kind of neural network to use. This will help to improve performance.
  • the method further includes: the second device sends confirmation information to the first device, where the confirmation information is used to indicate that the scene Mode indicating successful reception of information.
  • the second device may send confirmation information to the first device based on the successful reception of the above-mentioned indication information, so that the first device can perform subsequent procedures according to the reception of the indication information.
  • a neural network matching the scene pattern is used to work; in the case of unsuccessful reception, the indication information of the scene pattern is resent, and so on. This application does not limit this.
  • the method further includes: the second device communicates with the first device by using a neural network matching the scene mode.
  • the first device may use a neural network matching the scene pattern to work based on the successful reception of the above-mentioned indication information by the second device.
  • the second device can also determine and use a neural network matching the indicated scene mode to work based on the successful reception of the above-mentioned indication information, so that two mutually coupled neural networks can work together, which is beneficial to improve system performance.
  • each neural network in the plurality of neural networks is obtained based on a standardized reference data set, a standardized reference neural network, and a standardized performance judgment criterion. neural network.
  • the neural network used by the first device and the second device may be the target neural network obtained based on the method described in the first aspect above.
  • the first device is a network device, and the second device is a terminal device; or, the first device is a terminal device, and the first device is a terminal device.
  • the second device is a network device.
  • the above-mentioned determination of the air interface transmission conditions and selection of the scene mode may be performed by the network device, or may be performed by the terminal device. This application does not limit this.
  • a neural network matching the scene mode is determined between the first device and the second device through the indication information of the scene mode.
  • the determination of the air interface transmission conditions and the selection of the scene mode may not be indicated by signaling interaction.
  • the network device and the terminal device may each select based on the preset air interface transmission conditions and scene modes. Adapt the neural network to the matching relationship of the neural network. This application also does not limit this.
  • an apparatus for acquiring a neural network including each module or unit for executing the method in any possible implementation manner of the first aspect.
  • an apparatus for acquiring a neural network including a processor.
  • the processor is coupled to the memory, and can be used to execute instructions or data in the memory, so as to implement the method in any possible implementation manner of the first aspect above.
  • the apparatus further includes a memory.
  • the apparatus further includes an input/output interface to which the processor is coupled.
  • a communication apparatus including each module or unit for executing the method in any possible implementation manner of the second aspect.
  • a communication apparatus including a processor.
  • the processor is coupled to the memory, and can be used to execute instructions or data in the memory, so as to implement the method in any possible implementation manner of the second aspect.
  • the communication device further includes a memory.
  • the communication device further includes a communication interface, and the processor is coupled to the communication interface.
  • the communication apparatus is a terminal device.
  • the communication interface may be a transceiver, or an input/output interface.
  • the communication device is a chip configured in the terminal device.
  • the communication interface may be an input/output interface.
  • the transceiver may be a transceiver circuit.
  • the input/output interface may be an input/output circuit.
  • a communication apparatus including each module or unit for executing the method in any possible implementation manner of the third aspect.
  • a communication apparatus including a processor.
  • the processor is coupled to the memory and can be used to execute instructions or data in the memory, so as to implement the method in any possible implementation manner of the third aspect.
  • the communication device further includes a memory.
  • the communication device further includes a communication interface, and the processor is coupled to the communication interface.
  • the communication apparatus is a network device.
  • the communication interface may be a transceiver, or an input/output interface.
  • the communication device is a chip configured in a network device.
  • the communication interface may be an input/output interface.
  • the transceiver may be a transceiver circuit.
  • the input/output interface may be an input/output circuit.
  • a processor comprising: an input circuit, an output circuit and a processing circuit.
  • the processing circuit is configured to receive a signal through the input circuit and transmit a signal through the output circuit, so that the processor executes the method in any one of the possible implementations of the first aspect to the third aspect.
  • the above-mentioned processor may be one or more chips
  • the input circuit may be input pins
  • the output circuit may be output pins
  • the processing circuit may be transistors, gate circuits, flip-flops and various logic circuits, etc. .
  • the input signal received by the input circuit may be received and input by, for example, but not limited to, a receiver
  • the signal output by the output circuit may be, for example, but not limited to, output to and transmitted by a transmitter
  • the circuit can be the same circuit that acts as an input circuit and an output circuit at different times.
  • the embodiments of the present application do not limit the specific implementation manners of the processor and various circuits.
  • a processing apparatus including a processor and a memory.
  • the processor is used to read the instructions stored in the memory, and can receive signals through the receiver and transmit signals through the transmitter, so as to execute the method in any possible implementation manner of the first aspect to the third aspect.
  • processors there are one or more processors and one or more memories.
  • the memory may be integrated with the processor, or the memory may be provided separately from the processor.
  • the memory can be a non-transitory memory, such as a read only memory (ROM), which can be integrated with the processor on the same chip, or can be separately set in different On the chip, the embodiment of the present application does not limit the type of the memory and the setting manner of the memory and the processor.
  • ROM read only memory
  • the relevant data interaction process such as sending indication information, may be a process of outputting indication information from the processor, and receiving capability information may be a process of receiving input capability information by the processor.
  • the data output by the processor can be output to the transmitter, and the input data received by the processor can be from the receiver.
  • the transmitter and the receiver may be collectively referred to as a transceiver.
  • the processing device in the eleventh aspect above may be one or more chips.
  • the processor in the processing device may be implemented by hardware or by software.
  • the processor can be a logic circuit, an integrated circuit, etc.; when implemented by software, the processor can be a general-purpose processor, implemented by reading software codes stored in a memory, which can Integrated in the processor, can be located outside the processor, independent existence.
  • a twelfth aspect provides a computer program product, the computer program product comprising: a computer program (also referred to as code, or instructions), which, when the computer program is executed, causes a computer to execute the above-mentioned first aspect to The method in any possible implementation manner of the third aspect.
  • a computer program also referred to as code, or instructions
  • a thirteenth aspect provides a computer-readable storage medium, where the computer-readable storage medium stores a computer program (also referred to as code, or instruction), when it runs on a computer, causing the computer to execute the above-mentioned first
  • a computer program also referred to as code, or instruction
  • a communication system including the aforementioned network device and terminal device, wherein the network device and the terminal device may be respectively configured with the aforementioned target neural network.
  • FIG. 1 is a communication system that can communicate using the neural network provided by the embodiments of the present application;
  • Fig. 2 is the schematic diagram of the signal processing process of the transmitter and the receiver configured with the neural network
  • FIG. 3 is a schematic diagram of a deep neural network (DNN).
  • DNN deep neural network
  • FIG. 4 is a schematic diagram of a convolutional neural network (CNN).
  • FIG. 5 is a schematic diagram of a recurrent neural network (RNN);
  • 6 to 14 are schematic flowcharts of a training method of a neural network in a communication system provided by an embodiment of the present application
  • FIG. 15 and FIG. 16 are schematic flowcharts of the communication method provided by the embodiment of the present application.
  • FIG. 17 and FIG. 18 are schematic block diagrams of an apparatus for acquiring a neural network provided by an embodiment of the present application.
  • FIG. 19 and FIG. 20 are schematic block diagrams of a communication apparatus provided by an embodiment of the present application.
  • 21 is a schematic structural diagram of a terminal device provided by an embodiment of the present application.
  • FIG. 22 is a schematic structural diagram of a network device provided by an embodiment of the present application.
  • the communication method provided by the present application can be applied to various communication systems, such as: Long Term Evolution (Long Term Evolution, LTE) system, LTE frequency division duplex (frequency division duplex, FDD) system, LTE time division duplex (time division duplex, TDD), universal mobile telecommunication system (UMTS), worldwide interoperability for microwave access (WiMAX) communication system, future fifth generation (5th Generation, 5G) mobile communication system or new wireless Access technology (new radio access technology, NR).
  • the 5G mobile communication system may include a non-standalone (NSA, NSA) and/or an independent network (standalone, SA).
  • the communication method provided by this application can also be applied to machine type communication (MTC), long term evolution-machine (LTE-M), device to device (device to device, D2D) networks , Machine to Machine (M2M) network, Internet of Things (IoT) network or other network.
  • the IoT network may include, for example, the Internet of Vehicles.
  • vehicle to X, V2X, X can represent anything
  • the V2X may include: vehicle to vehicle (vehicle to vehicle, V2V) communication, vehicle and vehicle Infrastructure (V2I) communication, vehicle to pedestrian (V2P) or vehicle to network (V2N) communication, etc.
  • the communication method provided in this application can also be applied to future communication systems, such as the sixth generation mobile communication system. This application does not limit this.
  • the network device may be any device with a wireless transceiver function.
  • Network equipment includes but is not limited to: evolved Node B (evolved Node B, eNB), radio network controller (radio network controller, RNC), Node B (Node B, NB), base station controller (base station controller, BSC) , base transceiver station (base transceiver station, BTS), home base station (for example, home evolved NodeB, or home Node B, HNB), baseband unit (baseband unit, BBU), wireless fidelity (wireless fidelity, WiFi) system Access point (AP), wireless relay node, wireless backhaul node, transmission point (TP) or transmission and reception point (TRP), etc.
  • eNB evolved Node B
  • RNC radio network controller
  • Node B Node B
  • BSC base station controller
  • base transceiver station base transceiver station
  • BTS home base station
  • home base station for example, home evolved NodeB, or home Node B, HNB
  • It can also be 5G, such as NR , a gNB in the system, or, a transmission point (TRP or TP), one or a group of (including multiple antenna panels) antenna panels of a base station in a 5G system, or, it can also be a network node that constitutes a gNB or a transmission point, Such as baseband unit (BBU), or distributed unit (distributed unit, DU) and so on.
  • BBU baseband unit
  • DU distributed unit
  • a gNB may include a centralized unit (CU) and a DU.
  • the gNB may also include an active antenna unit (AAU).
  • CU implements some functions of gNB
  • DU implements some functions of gNB.
  • CU is responsible for processing non-real-time protocols and services, implementing radio resource control (RRC), and packet data convergence protocol (PDCP) layer function.
  • RRC radio resource control
  • PDCP packet data convergence protocol
  • the DU is responsible for processing physical layer protocols and real-time services, and implementing the functions of the radio link control (RLC) layer, medium access control (MAC) layer, and physical (PHY) layer.
  • RLC radio link control
  • MAC medium access control
  • PHY physical layer.
  • AAU implements some physical layer processing functions, radio frequency processing and related functions of active antennas.
  • the higher-layer signaling such as the RRC layer signaling
  • the network device may be a device including one or more of a CU node, a DU node, and an AAU node.
  • the CU can be divided into network devices in an access network (radio access network, RAN), and the CU can also be divided into network devices in a core network (core network, CN), which is not limited in this application.
  • the network equipment provides services for the cell, and the terminal equipment communicates with the cell through transmission resources (for example, frequency domain resources, or spectrum resources) allocated by the network equipment, and the cell may belong to a macro base station (for example, a macro eNB or a macro gNB, etc.) , can also belong to the base station corresponding to the small cell, where the small cell can include: urban cell (metro cell), micro cell (micro cell), pico cell (pico cell), femto cell (femto cell), etc. , these small cells have the characteristics of small coverage and low transmission power, and are suitable for providing high-speed data transmission services.
  • a macro base station for example, a macro eNB or a macro gNB, etc.
  • the small cell can include: urban cell (metro cell), micro cell (micro cell), pico cell (pico cell), femto cell (femto cell), etc.
  • these small cells have the characteristics of small coverage and low transmission power, and are suitable for providing high-speed data transmission
  • a terminal device may also be referred to as user equipment (user equipment, UE), an access terminal, a subscriber unit, a subscriber station, a mobile station, a mobile station, a remote station, a remote terminal, a mobile device, a user terminal, Terminal, wireless communication device, user agent or user equipment.
  • user equipment user equipment
  • UE user equipment
  • an access terminal a subscriber unit, a subscriber station, a mobile station, a mobile station, a remote station, a remote terminal, a mobile device, a user terminal, Terminal, wireless communication device, user agent or user equipment.
  • the terminal device may be a device that provides voice/data connectivity to the user, such as a handheld device with a wireless connection function, a vehicle-mounted device, and the like.
  • some examples of terminals can be: mobile phone (mobile phone), tablet computer (pad), computer with wireless transceiver function (such as notebook computer, palmtop computer, etc.), mobile internet device (mobile internet device, MID), virtual reality (virtual reality, VR) equipment, augmented reality (augmented reality, AR) equipment, wireless terminals in industrial control (industrial control), wireless terminals in unmanned driving (self driving), wireless terminals in remote medical (remote medical) Terminal, wireless terminal in smart grid, wireless terminal in transportation safety, wireless terminal in smart city, wireless terminal in smart home, cellular phone, cordless Telephone, session initiation protocol (SIP) telephone, wireless local loop (WLL) station, personal digital assistant (PDA), handheld device, computing device or connection with wireless communication capabilities
  • wearable devices can also be called wearable smart devices, which is a general term for the intelligent design of daily wear and the development of wearable devices using wearable technology, such as glasses, gloves, watches, clothing and shoes.
  • a wearable device is a portable device that is worn directly on the body or integrated into the user's clothing or accessories.
  • Wearable device is not only a hardware device, but also realizes powerful functions through software support, data interaction, and cloud interaction.
  • wearable smart devices include full-featured, large-scale, complete or partial functions without relying on smart phones, such as smart watches or smart glasses, and only focus on a certain type of application function, which needs to cooperate with other devices such as smart phones. Use, such as various types of smart bracelets, smart jewelry, etc. for physical sign monitoring.
  • the terminal device may also be a terminal device in an internet of things (Internet of things, IoT) system.
  • IoT Internet of things
  • IoT is an important part of the development of information technology in the future. Its main technical feature is to connect items to the network through communication technology, so as to realize the intelligent network of human-machine interconnection and interconnection of things.
  • IoT technology can achieve massive connections, deep coverage, and terminal power saving through, for example, narrow band (NB) technology.
  • NB narrow band
  • terminal equipment can also include sensors such as smart printers, train detectors, and gas stations.
  • the main functions include collecting data (part of terminal equipment), receiving control information and downlink data of network equipment, and sending electromagnetic waves to transmit uplink data to network equipment. .
  • the communication system 100 may include at least one network device, such as the network device 101 shown in FIG. 1; the communication system 100 may also include at least one terminal device, such as the terminal device 102 to 107.
  • the terminal devices 102 to 107 may be mobile or stationary.
  • Each of the network device 101 and one or more of the end devices 102 to 107 may communicate over a wireless link.
  • Each network device can provide communication coverage for a specific geographic area and can communicate with terminal devices located within that coverage area.
  • the network device may send configuration information to the terminal device, and the terminal device may send uplink data to the network device based on the configuration information; for another example, the network device may send downlink data to the terminal device. Therefore, the network device 101 and the terminal devices 102 to 107 in FIG. 1 constitute a communication system.
  • D2D technology can be used to realize direct communication between terminal devices.
  • D2D technology can be used for direct communication between terminal devices 105 and 106 and between terminal devices 105 and 107 .
  • Terminal device 106 and terminal device 107 may communicate with terminal device 105 individually or simultaneously.
  • the terminal devices 105 to 107 can also communicate with the network device 101, respectively. For example, it can communicate directly with the network device 101. In the figure, the terminal devices 105 and 106 can communicate directly with the network device 101; it can also communicate with the network device 101 indirectly. In the figure, the terminal device 107 communicates with the network device via the terminal device 105. 101 Communications.
  • FIG. 1 exemplarily shows a network device, a plurality of terminal devices, and communication links between the communication devices.
  • the communication system 100 may include multiple network devices, and the coverage of each network device may include other numbers of terminal devices, such as more or less terminal devices. This application does not limit this.
  • Each of the above communication devices may be configured with multiple antennas.
  • the plurality of antennas may include at least one transmit antenna for transmitting signals and at least one receive antenna for receiving signals.
  • each communication device additionally includes a transmitter and a receiver, and those of ordinary skill in the art can understand that they can all include multiple components (eg, processors, modulators, multiplexers, demultiplexers, etc.) related to signal transmission and reception. modulator, demultiplexer or antenna, etc.). Therefore, the network device and the terminal device can communicate through the multi-antenna technology.
  • the wireless communication system 100 may further include other network entities such as a network controller, a mobility management entity, and the like, which are not limited in this embodiment of the present application.
  • network entities such as a network controller, a mobility management entity, and the like, which are not limited in this embodiment of the present application.
  • the transmitter in the network device can be used to modulate the signal, and the receiver in the terminal device can be used to demodulate the modulated signal; the transmitter in the network device can be used to encode the signal, and the receiver in the terminal device can be used to encode the signal.
  • the receiver in the terminal device can be used to decode the encoded signal; the transmitter in the network device can be used to precode the signal, and the receiver in the terminal device can be used to detect the precoded signal; the transmitter in the terminal device can be used to detect the precoded signal; It can be used to compress the channel information, and the receiver in the network device can be used to reconstruct the channel based on the compressed channel information.
  • both the network device and the terminal device may use a neural network to work.
  • the neural network adopted by the network device and the terminal device may be obtained based on the method for obtaining a neural network provided in the embodiments of the present application.
  • the originating neural network obtained by the method for acquiring a neural network provided by the embodiment of the present application is applied to the transmitter of the network device
  • the receiving end neural network obtained by the method for acquiring the neural network provided by the embodiment of the present application is applied.
  • the originating neural network obtained by the method for obtaining a neural network provided by the embodiment of the present application is applied in the receiver of the terminal device; and/or, the originating neural network obtained by the method for obtaining a neural network provided by the The end-end neural network obtained by the network method is applied to the transmitter of the network device.
  • the transmitter in the network device and the receiver in the terminal device can be jointly optimized based on the neural network, and the transmitter in the terminal device and the receiver in the network device can also be jointly optimized based on the neural network. Thereby, the performance of the communication system can be improved.
  • Fig. 2 shows the signal processing process of the transmitter and the receiver respectively configured with the neural network.
  • Fig. 2 shows the signal processing process of the transmitter using the neural network and the receiver using the neural network by taking the extraction and compression of the channel information and the reconstruction of the channel information as an example.
  • the originating neural network for extracting and compressing channel information may include a pre-processing module, an information extraction module and a post-processing module.
  • the originating neural network used for extracting and compressing channel information may also include an information extraction module and a post-processing module instead of a pre-processing module, or may include an information extraction module instead of a pre-processing module and a post-processing module.
  • the modules specifically included in the originating neural network may be defined based on different designs, which are not limited in this application.
  • the end-end neural network used to reconstruct the channel information may include an inverse preprocessing module, an inverse information extraction module, and an inverse postprocessing module. However, this should not constitute any limitation to this application.
  • the receiving-end neural network used to reconstruct the channel information may not include an inverse preprocessing module, but may include an inverse information extraction module and an inverse postprocessing module, or may not include an inverse preprocessing module and an inverse postprocessing module. , and includes the inverse information extraction module. It should be understood that the modules specifically included in the end-end neural network may be defined based on different designs, which are not limited in this application.
  • the first possible structure of the originating neural network for extracting and compressing channel information includes a pre-processing module, an information extraction module and a post-processing module.
  • the preprocessing module can be used to preprocess the original data input to the originating target neural network based on the standardized data structure.
  • the information extraction module can be used to extract information from the preprocessed data, so as to compress the preprocessed data to obtain the compressed data.
  • the post-processing module can be used to quantize and bit map the compressed data based on a standardized data structure to generate air interface information.
  • the second possible structure of the originating neural network for channel information extraction and compression is to include an information extraction module and a post-processing module.
  • the originating target neural network may not include a preprocessing module.
  • the preprocessing module can be configured in the transmitter and can be used to preprocess the raw data input to the transmitter based on a standardized data structure.
  • the preprocessed data is input into the originating neural network.
  • the information extraction module in the originating neural network can be used to extract information from the preprocessed data, so as to compress the preprocessed data to obtain the compressed data.
  • the post-processing module can be used to quantize and bit map the compressed data based on a standardized data structure to generate air interface information.
  • a third possible structure of the originating neural network for extraction and compression of channel information is to include an information extraction module.
  • the originating neural network may not include a pre-processing module and a post-processing module.
  • the pre-processing module and the post-processing module can be configured in the transmitter.
  • the preprocessing module can be used to preprocess the raw data input to the transmitter based on the standardized data structure.
  • the preprocessed data is input into the originating neural network.
  • the information extraction module in the originating neural network is used to extract information from the preprocessed data, so as to compress the preprocessed data to obtain the compressed data.
  • the compressed data can be input into a post-processing module, and the post-processing module can be used to perform quantization and bit mapping on the compressed data based on a standardized data structure to generate air interface information.
  • the first possible structure of the end-end neural network for channel information reconstruction is to include an inverse post-processing module, an inverse information extraction module and an inverse pre-processing module.
  • the inverse post-processing module can be used to parse the air interface information input to the target neural network at the receiving end based on a standardized data structure to obtain compressed data.
  • the inverse information extraction module can be used to process the compressed data to obtain the data before compression.
  • the inverse preprocessing module can be used to process the data before compression to obtain the original data.
  • the second possible structure of the end-end neural network used to reconstruct the channel information is to include an inverse information extraction module and an inverse post-processing module.
  • the end-end neural network may not include an inverse preprocessing module.
  • the inverse preprocessing module can be configured in the receiver.
  • the inverse post-processing module in the end-end neural network can be used to analyze the air interface information input to the end-end target neural network based on the standardized data structure to obtain compressed data.
  • the inverse information extraction module can be used to process the compressed data to obtain the data before compression.
  • the pre-compressed data can be input into an inverse pre-processing module, and the inverse pre-processing module can be used to process the pre-compressed data to obtain original data.
  • the third possible structure of the target neural network at the receiving end for reconstructing the channel information is to include an inverse information extraction module.
  • the end-end neural network may not include an inverse pre-processing module and an inverse post-processing module.
  • the inverse pre-processing module and the inverse post-processing module may be configured in the receiver.
  • the inverse post-processing module can be used to parse the air interface information input to the receiver based on the standardized data structure to obtain compressed data.
  • the compressed data can be input into the end-end neural network.
  • the inverse information extraction module in the end-end neural network can be used to process the compressed data to obtain the data before compression.
  • the compressed data can be obtained by processing the received air interface information based on a standardized data structure.
  • the data before compression can be input to an inverse preprocessing module, and the inverse preprocessing module can be used to process the data before compression to obtain original data.
  • the inverse preprocessing module can be regarded as the inverse processing module of the preprocessing module
  • the inverse information extraction module can be regarded as the inverse processing module of the information extraction module
  • the inverse postprocessing module can be regarded as the inverse processing module of the postprocessing module.
  • the behavior of the modules in the transmitter and the modules in the receiver are reciprocal.
  • the preprocessing module may also be referred to as a preprocessing module
  • the postprocessing module may also be referred to as a quantization module, and so on.
  • FIG. 2 is a schematic diagram of extraction and compression of channel information and reconstruction of channel information.
  • the channel information shown in FIG. 2 is channel H.
  • the transmitter obtains the channel H, it can first input the channel H into the pre-processing module, and the pre-processing module pre-processes the channel H to obtain the pre-processed channel H.
  • Channel characteristic information V For example, perform singular value decomposition on channel H to obtain eigenvectors, or perform space-frequency domain transformation on channel H to obtain an angle delay spectrum.
  • V can be understood as the channel after preprocessing.
  • the pre-processed channel feature information V is input to the information extraction module, and the information extraction module can perform information extraction on the channel feature information V based on a pre-designed algorithm to obtain a compressed channel V'.
  • the compressed channel feature information V' is obtained based on the function f en (V, ⁇ en ).
  • the compressed channel feature information V' is input to the post-processing module, and the post-processing module can perform quantization and bit mapping on the compressed channel feature information V' based on a predefined data structure to generate air interface information S.
  • the receiver After receiving the air interface information S, the receiver can input the air interface information S to an inverse post-processing module in the receiver.
  • the inverse post-processing module can process the air interface information S based on the predefined data structure to recover the compressed channel feature information Compressed channel feature information It is input to the inverse information extraction module, and the inverse information extraction module can analyze the channel characteristic information based on the algorithm corresponding to the information extraction module in the transmitter.
  • the channel feature information is recovered based on the function f de (f en (V, ⁇ en ), ⁇ de )
  • Channel characteristic information is input to the inverse pre-processing module, and the inverse pre-processing module has the channel characteristic information
  • the channel can be recovered It should be understood that the channel thus recovered is the estimated value of channel H input into the transmitter above.
  • the pre-processing module and the reverse pre-processing module need to work based on consistent rules
  • the information extraction module and the reverse information extraction module need to work based on consistent rules
  • the post-processing module and the reverse post-processing module need to work.
  • Work based on consistent rules to ensure that the neural network on the transmitter side and the neural network on the receiver side can work together. Consistency here means that the two can be designed based on consistent rules, but it does not mean that both parties use the same algorithm, the same data structure, etc. to process data.
  • Working based on consistent rules aims to emphasize that the algorithms, data structures, etc. used by both parties when they are used to process data are corresponding.
  • the post-processing module in the transmitter performs quantization and bit mapping on the compressed channel V' based on a certain data structure to generate air interface information S; the inverse post-processing module in the receiver needs to be based on a corresponding data structure.
  • the air interface information S recovers the channel characteristic information to the greatest extent
  • the information extraction module in the transmitter compresses the channel characteristic information V based on a certain algorithm to obtain the compressed channel characteristic information V'; the inverse information extraction module in the receiver needs to compress the compressed channel characteristic information V' based on a corresponding algorithm.
  • the channel characteristic information V' is processed to restore the channel characteristic information before compression to the greatest extent.
  • the channel H illustrated in FIG. 2 is only an example for ease of understanding, and should not constitute any limitation to the present application.
  • the channel H and the channel characteristic information V listed above can be used as examples of the channel information. Due to the different designs of the modules included in the neural network, the channel information input to the originating neural network may be the channel H or the channel feature information V; the output of the originating neural network may be the compressed channel feature information V', It may also be air interface information S.
  • the input of the neural network at the receiving end may be the air interface information S, or it may be the compressed channel feature information recovered based on the air interface information S
  • the channel information output from the receiving end neural network may be the recovered channel feature information It may also be the recovered channel
  • the definitions of the channel information may also be different.
  • the channel information input to the sender neural network may be channel H
  • the channel information output by the receiver neural network may be the recovered channel feature information
  • X represents the channel information input to the originating neural network
  • Y represents the channel information output by the receiving neural network.
  • X can be either H or V
  • Y can be or any one of .
  • the originating neural network described above in conjunction with FIG. 2 is a neural network for extracting and compressing channel information, and the modules included therein are related to the process of extracting and compressing channel information.
  • the end-end neural network described above in conjunction with FIG. 2 is a neural network used for reconstruction of channel information, and the modules included therein are also related to the reconstruction process of channel information.
  • the modules included in the sender neural network and the receiver neural network may be different. For brevity, examples are not given here.
  • the transmitter and receiver may be provided by different manufacturers.
  • the transmitter or receiver may need to download the neural network structure and parameter information matching the current scene from the peer, or the transmitter and receiver may download the neural network structure matching the current scene from a certain network virtual processing node and parameter information. For example, download the rules for generating and parsing air interface information; another example, download the parameters and structure description of the neural network that the transmitter performs information extraction and compression on the channel H, etc. This may bring about a large air interface signaling overhead.
  • the present application provides a method for acquiring a neural network in a communication system, in order to reduce the huge air interface overhead caused by the transmitter and the receiver downloading the neural network structure and parameter information.
  • Neural network As an important branch of artificial intelligence, it is a network structure that imitates the behavioral characteristics of animal neural networks for information processing.
  • the structure of the neural network is composed of a large number of nodes (or neurons) connected to each other, and the purpose of processing information is achieved by learning and training the input information based on a specific operation model.
  • a neural network includes an input layer, a hidden layer and an output layer.
  • the input layer is responsible for receiving input signals
  • the output layer is responsible for outputting the calculation results of the neural network
  • the hidden layer is responsible for complex functions such as feature expression.
  • the function of the hidden layer is determined by the weight matrix and the corresponding activation function to represent.
  • a deep neural network is generally a multi-layer structure. Increasing the depth and width of a neural network can improve its expressive ability and provide more powerful information extraction and abstract modeling capabilities for complex systems.
  • DNN can be constructed in various ways, for example, it can include, but is not limited to, recurrent neural network (RNN), convolutional neural network (CNN), and fully connected neural network.
  • RNN recurrent neural network
  • CNN convolutional neural network
  • fully connected neural network recurrent neural network
  • FIG. 3 is a schematic diagram of a feedforward neural network (FNN). As shown, different boxes represent different layers. For example, from left to right, the first layer is the input layer, the second layer is the hidden layer, and the third layer is the output layer. DNN generally also contains multiple hidden layers, which are used to extract feature information to varying degrees. The output layer can be used to map the desired output information from the extracted feature information. Each black circle in the figure can represent a neuron, and neurons in different boxes represent neurons in different layers. The way the neurons in each layer are connected and the activation function used will determine the expression function of the neural network.
  • FNN feedforward neural network
  • CNN Convolutional Neural Network
  • each layer is composed of multiple two-dimensional planes, and each plane is composed of multiple independent neurons, and multiple neurons in each plane share weights , the number of parameters in the neural network can be reduced by weight sharing.
  • the convolution operation performed by the processor is usually to convert the convolution of the input signal feature and the weight into a matrix multiplication operation between the signal matrix and the weight matrix.
  • the signal matrix and the weight matrix are divided into blocks to obtain a plurality of fractal signal matrices and fractal weight matrices, and then matrix multiplication and accumulation operations are performed on the plurality of fractal signal matrices and fractal weight matrices.
  • Figure 4 is a schematic diagram of the hidden layers in a CNN.
  • the black box in the figure is the channel of each layer, or the feature map. Different channels can correspond to the same or different convolution kernels.
  • CNN not all neurons in the upper and lower layers can be directly connected, but through the "convolution kernel" as an intermediary.
  • the information to be processed can still retain the original positional relationship after the convolution operation.
  • the mth layer on the right is obtained.
  • FIG. 4 is only an example for ease of understanding, and for a more detailed description of the CNN, reference may be made to the prior art. For brevity, no detailed description is given here.
  • RNN Recurrent Neural Network
  • FIG. 5 is a schematic diagram of an RNN. Each black circle in the figure represents a neuron. It can be seen that the input of the neurons in the hidden layer 2 not only includes the output of the neurons in the hidden layer 1 at this moment (as shown by the straight line structure in the figure), but also includes its own output at the previous moment (as shown in the figure). shown by the semicircular arrow).
  • FIG. 5 is merely an example for ease of understanding. A more detailed description of RNN can be found in the prior art. For brevity, no detailed description is given here.
  • Performance judgment criterion In this embodiment of the present application, in order to obtain a target neural network with better performance, the pre-built target neural network can be evaluated through the performance judgment criterion.
  • the performance judgment criteria may include performance indicators and output conditions.
  • a standardized performance metric function can be used to obtain a performance metric value, and the performance metric value can be input into the performance judgment criterion and compared with one or more performance indexes to judge whether it satisfies the output condition.
  • performance metric functions may include, but are not limited to, one or more of the following: mean square error (MSE), normalized mean square error (NMSE), mean absolute Error (mean absolute error, MAE), maximum absolute error (also called absolute error bound), correlation coefficient, cross entropy, mutual information, bit error rate, or bit error rate, etc.
  • MSE mean square error
  • NMSE normalized mean square error
  • MAE mean absolute Error
  • MAE mean absolute error
  • maximum absolute error also called absolute error bound
  • MSE, NMSE, MAE, maximum absolute error, bit error rate, bit error rate, etc. can be used to characterize the degree of difference
  • correlation coefficient, cross entropy, mutual information, etc. can be used to characterize the degree of similarity.
  • the performance metrics may include upper bounds for one or more performance metrics and/or lower bounds for one or more performance metrics.
  • the performance indicator can be characterized, for example, by one or more of the items listed above in the performance measurement function. For different performance measurement functions, different performance indicators can be defined.
  • the upper bound of the performance index can be defined as 0.99999, and the lower bound of the performance index can be defined as 0.9.
  • the lower bound of the performance index can be defined as 0.1.
  • the performance index is an index set for judging whether the target neural network satisfies the output conditions, in other words, the target neural network that achieves the performance index can achieve acceptable performance after being applied to the transmitter or receiver. Scope. For example, taking the system throughput as an example, if the similarity coefficient is less than 0.9, it can be considered that the performance is too poor; if the correlation coefficient is greater than or equal to 0.9, the performance can be considered acceptable; if the correlation coefficient is greater than 0.99999, it can be considered that the performance has no greater performance gain, or very little gain. Therefore the acceptable range of the correlation coefficient is between [0.9, 0.99999]. Therefore, for the correlation coefficient, reaching the lower bound of the performance index may mean greater than or equal to 0.9, and not exceeding the upper bound of the performance index may mean less than or equal to 0.99999.
  • the bit error rate if the MSE is greater than 0.1, it can be considered that the performance is too poor; if the MSE is less than or equal to 0.1, the performance can be considered acceptable; if the MSE is less than 0.001, it can be considered that there is no greater gain in performance , so the acceptable range of MSE can be between [0.001, 0.1]. Therefore, for MSE, reaching the lower bound of the performance index may mean less than or equal to 0.1, and not exceeding the upper bound of the performance index may mean greater than or equal to 0.001.
  • the performance measurement function is used to represent the degree of similarity, the higher the performance measurement value output by the performance measurement function, the higher the degree of similarity, and the higher the performance of the target neural network.
  • the upper bound of the performance metric is smaller than the lower bound of the performance metric.
  • the upper and lower bounds of performance indicators are defined based on good or bad performance. Therefore, for different performance indicators, the relationship between the upper or lower bounds and specific values is not static.
  • the degree of similarity when it comes to the relevant description of "reaching the lower bound of the performance index”, it may mean that the degree of similarity is greater than or equal to a certain threshold (for the convenience of distinction and description, for example, it is denoted as the first threshold); It may mean that the degree of difference is less than or equal to a certain threshold (for the convenience of distinction and description, for example, it is denoted as a second threshold).
  • the degree of similarity is less than or equal to a certain threshold (for the convenience of distinction and description, for example, it is recorded as the third threshold); it can also mean that the degree of difference is greater than or equal to a certain threshold.
  • the fourth threshold is or equal to a certain threshold (for the convenience of distinction and description, for example, it is recorded as the fourth threshold).
  • the corresponding first threshold is smaller than the third threshold; for the degree of difference, the corresponding second threshold is greater than the fourth threshold.
  • the one or more performance indicators may include one or more performance indicators given for a performance measurement function, such as including upper bounds and/or lower bounds of the performance indicators; the one or more performance indicators may also include Each performance measurement function in the different performance measurement functions is given one or more performance indicators respectively. This application does not limit this.
  • whether the target neural network satisfies the output condition is determined based on the correlation coefficient.
  • the protocol can define the upper and lower bounds of the correlation coefficient, and can define the output condition as: the performance measurement values all reach the lower bound of the performance index, and do not exceed the upper bound of the performance index. As in the previous example, the lower bound is 0.9 and the upper bound is 0.99999.
  • the results output by the target neural network and the reference neural network are input to the performance measurement function, and one or more performance measurement values can be obtained. Compare the one or more performance metrics to the upper and lower bounds, respectively, to see if they all fall within the range [0.9, 0.99999]. If so, the output condition is satisfied; if not, the output condition is not satisfied.
  • the protocol can define the upper and lower bounds of the correlation coefficient, and can define the output condition as follows: all the performance measurement values reach the lower bound of the performance index, and 90% of the performance measurement values do not exceed the upper bound of the performance index. As in the previous example, the lower bound is 0.9 and the upper bound is 0.99999. Then, one or more performance measurement values output by the performance measurement function can be compared with the upper bound and the lower bound, respectively, to see whether all reach the lower bound of 0.9, and whether more than 90% of the performance measurement values do not exceed the upper bound of 0.99999. If so, the output condition is satisfied; if not, the output condition is not satisfied.
  • whether the target neural network satisfies the output condition is determined based on the correlation coefficient and the bit error rate.
  • the protocol can define the upper bound of the performance index, including the upper and lower bounds of the correlation coefficient, and the lower bound of the bit error rate; and the output condition can be defined as: the performance measurement value reaches the lower bound of the correlation coefficient and the lower bound of the bit error rate respectively, and does not exceed the lower bound of the bit error rate.
  • the upper bound of the correlation coefficient As in the previous example, the lower bound of the correlation coefficient is 0.9, the upper bound is 0.99999, and the lower bound of the bit error rate is 0.1.
  • one or more performance metric values representing the correlation coefficient output by the performance metric function can be compared with the upper and lower bounds of the correlation coefficient to see if both reach the lower bound of 0.9, and none of them exceed the upper bound of 0.99999; Whether one or more of the output performance metrics that characterize the bit error rate reach a lower bound.
  • the performance judgment criteria may further include upper bounds and/or lower levels of performance indicators set for different service priorities, and output conditions.
  • the corresponding performance indicators may be higher, or the output conditions may be more stringent; for services such as short messages, the corresponding performance indicators may be lower, or the output conditions may be looser.
  • the lower bound of the correlation coefficient is 0.9, and the performance index of the former is higher than that of the lower bound of the correlation coefficient is 0.8.
  • the lower bound of MSE is 0.1 compared to the lower bound of MSE of 0.01, the latter has a higher performance index.
  • the output condition is that all the performance measurement values reach the lower bound of the performance index, and the former is more severe than that some performance measurement values reach the lower bound of the performance index.
  • the corresponding performance indicators and output conditions can be used to evaluate it according to the priority of the applied business.
  • Dual network It is used to describe the neural network structure jointly optimized by both transceiver ends (such as the base station side and the terminal side) in the communication network.
  • the dual network may specifically include an originating neural network and a sinking neural network.
  • the dual network includes, but is not limited to, an autoencoder (Autoencoder) structure or various variations of the autoencoder structure, or a receiving and transmitting neural network jointly constructed by a combination of other neural network structures.
  • the originating neural network is an encoder neural network
  • the receiving neural network is a decoder neural network. The two are mutually restrained and work together.
  • encoders and decoders described here are different from the encoders and decoders described above for channel encoding and decoding.
  • the originating neural network and the terminating neural network may be one-to-one, one-to-many, and many-to-one combinations to meet various system performance and implementation complexity requirements.
  • the method may include a process of acquiring an originating neural network applied in a transmitter and a process of acquiring a receiving neural network applied in a receiver.
  • the following describes the process of acquiring the originating neural network and the sinking neural network in conjunction with different embodiments.
  • the neural network to be acquired hereinafter is referred to as the target neural network.
  • the target neural network applied to the transmitter is called the originating target neural network
  • the target neural network applied to the receiver is called the receiving target neural network.
  • the acquisition of the originating target neural network and the terminating target neural network may be done offline, for example.
  • the originating target neural network and the terminating target neural network may, for example, be obtained from the same device, or may be obtained from different devices, which are not limited in this application.
  • the device for acquiring the target neural network may be, for example, a communication device, or a computing device different from the communication device. This application does not limit this.
  • the device for acquiring the target neural network is referred to as an acquisition device.
  • FIG. 6 is a schematic flowchart of a method 600 for acquiring a neural network provided by an embodiment of the present application. As shown in FIG. 6 , the method 600 includes steps 610 to 630 . Each step in the method 600 is described in detail below.
  • step 610 based on the standardized reference data set, the standardized one or more reference neural networks and the standardized performance judgment criteria, it is determined whether the constructed target neural network satisfies the preset output conditions.
  • the target neural network can be a dedicated neural network for an application.
  • the target neural network can be applied to each signal processing stage in the communication system. For example, coding and decoding, constellation modulation and demodulation, precoding and detection reception, extraction and compression of channel information and reconstruction of channel information.
  • the originating neural network can be used for encoding, and the receiving neural network can be used for decoding;
  • the originating neural network can be used for constellation modulation, and the receiving neural network can be used for demodulation;
  • the transmitting neural network can be used for precoding, and the receiving neural network can be used for Detection and reception;
  • the neural network at the sending end can be used to extract and compress the channel information, and the receiving end can be used to reconstruct the channel information.
  • target neural network obtained based on the methods provided in the embodiments of the present application may also be used as other functions, which are not limited in the present application.
  • the target neural network can be built on the existing neural network structure, for example, the design can be extended on the basic network structure such as CNN and RNN; it can also be designed with a new network structure, based on standardized reference data sets and/or related to the application.
  • the target neural network can be obtained by training it on other datasets.
  • the acquisition device can train the target neural network.
  • the weight vector of each layer of neural network can be updated according to the difference between the predicted value of the current network and the actual desired target value. For example, if the predicted value of the network is high, the weight vector can be adjusted to make the prediction lower, and through continuous adjustment, the neural network can predict the real desired target value. Therefore, it is necessary to pre-define "how to compare the difference between the predicted value and the target value", that is, the loss function (loss function), cost function (cost function) or objective function (objective function), which are used to measure the predicted value. important equation for the difference from the target value. Among them, taking the loss function as an example, the higher the output value of the loss function (loss), the greater the difference, then the training of the neural network becomes the process of reducing this output value as much as possible.
  • the above-mentioned training of the target neural network can be implemented based on a standardized reference dataset and/or other datasets related to the application.
  • the standardized reference data set may specifically include: a reference data set for training the target neural network, a reference data set for evaluating the performance of the target neural network, and the like.
  • Standardization may refer to protocol definitions. Vendors can build a target neural network based on the above method and evaluate the target neural network based on a standardized reference data set.
  • the target neural network is a neural network applied in a transmitter or a receiver.
  • the target neural network can be evaluated based on the standardized reference data set, the standardized reference neural network and the standardized performance judgment criteria to determine whether it satisfies the output conditions.
  • the reference dataset, reference neural network, and performance judgment criteria can all be standardized, or pre-defined by the protocol.
  • the protocol can predefine multiple sets of different standardized configurations.
  • a set of standardized configurations that match the applied scenarios can be used for evaluation.
  • a matching reference data set, reference neural network, performance measurement function, and performance judgment criteria can be used to evaluate the target neural network to obtain target neural networks applied to different scenarios.
  • the application scenarios mentioned here may specifically refer to scenarios differentiated based on different air interface transmission conditions.
  • the above application scenarios may include, but are not limited to, different channel scenarios such as outdoor dense urban areas, outdoor ordinary urban areas, outdoor villages, outdoor mountainous areas, indoor offices, indoor workshops, etc. Interfering signal types, different SNR conditions with high SNR in the near point and low SNR in the far point, different moving conditions such as low walking speed, medium vehicle speed in urban areas, and ultra-high speed on high-speed rail, etc.
  • the standardized configuration that matches the application scenario may include one or more of the following: a standardized reference data set, a standardized reference neural network, a standardized data structure, a standardized performance measurement function, and a standardized performance judgment criterion.
  • the reference neural network may include a receiving end reference neural network and/or a sending end reference neural network. Moreover, when the reference neural network includes a receiving end reference neural network and a sending end reference neural network, the receiving end reference neural network and the sending end reference neural network are one-to-one coupled.
  • Data structures can be used to generate and/or parse air interface information. More specifically, the data structure can be used for the originating neural network to generate air interface information, and can also be used for the receiving neural network to parse the air interface information.
  • the data structure can also be understood as a rule for generating air interface information, or a rule for interpreting air interface information.
  • a data structure is a high-level generalization that can be used to guide the processing of data at various stages of processing the data.
  • the originator can pre-process the input original data based on a certain data structure (for example, denoted as data structure 1), such as rearranging the input original data; Information extraction is performed on the rearranged data, for example, to determine which tensor expression to use; air interface information can also be generated based on another data structure (for example, denoted as data structure 3), for example, to determine how to perform quantization and bit mapping.
  • the receiving end can parse the air interface information based on the data structure corresponding to the data structure 1, and can process the parsed data based on the data structure corresponding to the data structure 2, and also based on the data structure corresponding to the data structure 3. , to get the raw data.
  • This application does not limit the specific content of the data structure.
  • the standardized achievable capability metric value can be understood as the capability parameter required by the neural network to implement a certain application, and can be used to select a matching neural network and/or a matching performance index.
  • the matched neural network described here may specifically refer to that the neural network can be applied to a transmitter or a receiver under given constraints, where the given constraints may include implementation complexity and/or air interface overhead .
  • the matched performance index may specifically refer to the performance index used to evaluate the neural network in the stage of obtaining the neural network.
  • the achievable capability metric value may include, but is not limited to, one or more of the following: the overhead of air interface information, the amount of computation required for neural network implementation, the amount of parameters stored in neural network implementation, and the computational accuracy required for neural network implementation .
  • the overhead of the air interface information may refer to, for example, the bit overhead; the amount of computation required by the neural network implementation may, for example, refer to the number of floating-point operations per second (Flops); the computational precision required by the neural network implementation For example, it can be 2 to specific points, or 4 to specific points, or 8 to specific points, or 16 to specific points, or 16-bit floating point, or 32-bit floating point, or 64-bit floating point. This application does not limit this.
  • Each neural network can be characterized by parameters corresponding to those listed above. Since the implementation complexity of different applications may be different, the requirements for neural network capabilities are also different.
  • the acquisition device can select a matching neural network based on the current application's ability to support the implementation complexity.
  • the upper and/or lower limits of the performance that different neural networks can achieve may be different.
  • the same performance metric may have one or more performance metric thresholds (eg, including upper and/or lower bounds), corresponding to one or more different capabilities.
  • corresponding performance indicators may be selected for evaluation according to the selected achievable capability metric value.
  • standardized configurations matched with different application scenarios may be different from each other, or may be partially the same.
  • standardized configurations that match different application scenarios can share the same reference data set, or share the same data structure, and so on. This application does not limit this.
  • manufacturers can cascade the constructed target neural network with one or more standardized reference neural networks based on a standardized reference data set, and evaluate the cascaded output results.
  • the performance judgment criterion determines whether the target neural network satisfies the output conditions. If the output conditions are met, step 620 may be executed to output the target neural network; if the output conditions are not met, step 630 may be executed to optimize the target neural network until the obtained target neural network satisfies Output condition, until it is output.
  • each manufacturer can input the same one or more data from the standardized reference data set into the constructed target neural network and one or more standardized reference neural networks, and output the target neural network
  • the result of the target neural network is compared with the output result of the reference neural network, and it is determined whether the target neural network meets the output conditions according to the standardized performance judgment criteria. If the output conditions are met, step 620 may be executed to output the target neural network; if the output conditions are not met, step 630 may be executed to optimize the target neural network until the obtained target neural network satisfies Output condition, until it is output.
  • the originating neural network may include a preprocessing module, an information extraction module and a postprocessing module
  • the sinking neural network may include an inverse postprocessing module, an inverse information processing module and a postprocessing module. Inverse preprocessing module.
  • FIG. 7 and FIG. 8 are schematic diagrams of a method for acquiring a neural network provided by an embodiment of the present application.
  • the methods for obtaining a neural network shown in FIGS. 7 and 8 are examples of cascading a target neural network and a standardized reference neural network, and evaluating the cascaded output results.
  • the target neural network shown in FIG. 7 can be applied to the transmitter, and the target neural network is the reference neural network at the origin; the reference neural network can be applied to the receiver, and the reference neural network is the reference neural network at the receiving end.
  • the target neural network shown in FIG. 8 can be applied to the receiver, the target neural network is the target neural network at the receiving end, and the reference neural network is applied to the transmitter, and the reference neural network is the reference neural network at the originating end.
  • the obtaining device may obtain one or more originating data from the reference data set. For example, denote one of the originating data as X, and X is input to the originating target neural network.
  • the compressed data X' can be obtained, and the originating target neural network can further process the compressed data X' according to the standardized data structure to generate an air interface Information S.
  • the air interface information S can be input into the reference neural network of the receiving end.
  • the receiving end neural network can recover the compressed data The receiving end refers to the neural network, which can be further based on the compressed data Get the output result Y.
  • the originating target neural network can be applied in different processing stages of the signal. Take the extraction and compression of channel information and the reconstruction of channel information as an example.
  • the originating data X may be channel H, for example.
  • the originating neural network may perform pre-processing, information extraction and post-processing on the channel H successively based on the operation flow of the originating neural network enumerated in FIG. 2 to generate air interface information S.
  • the receiving end reference neural network can reconstruct the channel H based on the received air interface information S.
  • the reference neural network at the receiving end can perform inverse post-processing, inverse information extraction and inverse pre-processing on the air interface information S successively based on the operation flow of the receiving end neural network listed in FIG. 2 to recover the channel.
  • the output of the reference neural network at the receiving end is the estimated value of channel H It should be understood that the estimated value is also the output result.
  • the acquiring device may input the output result Y and the input originating data X into the performance measurement function to obtain the performance measurement value. It should be understood that the obtaining device may input the output result Y and the originating data X into one or more performance measurement functions to obtain one or more performance measurement values.
  • the obtaining device can also process one or more originating data obtained from the reference data set according to the above process, and can obtain multiple pairs of output results and originating data.
  • Each pair of output result and originating data can be input into one or more performance metric functions to obtain one or more performance metric values.
  • Multiple pairs of output results and originating data can obtain multiple performance metrics.
  • the obtaining device may further compare the output performance metric value with the standardized performance index according to the standardized performance judgment criterion, and then determine whether the originating target neural network satisfies the preset output condition according to the comparison result.
  • the output condition may be that the performance metrics all reach the lower bound of the performance metrics.
  • the performance measurement function is, for example, MSE
  • the lower bound of the performance indicator can be, for example, the lower bound of MSE.
  • One or more MSE values can be obtained by inputting one or more pairs of output result and origination data into the performance metric function.
  • output The MSE of the two can be obtained by inputting the data H and the originating data into the performance measurement function, namely, The acquisition merge can further determine whether the acquired one or more performance metric values reach the lower bound of the normalized MSE. If all are met, the output conditions are met; if some or all of them are not met, the output conditions are not met.
  • the above performance index can be evaluated based on the performance index corresponding to the achievable capability metric value.
  • the input and output of the originating neural network and the input and output of the terminating neural network are not necessarily the same as the above examples.
  • the input of the originating neural network may be the channel H
  • the output of the terminating neural network may be the feature vector (it should be understood that the feature vector may be understood as an example of the channel feature information V described above).
  • the eigenvector output by the receiving end is the restored eigenvector, for example, denoted as But this does not affect the use of performance judgment criteria.
  • U eig(H) (eig() represents the function used to solve the eigenvector), which can be determined by the eigenvector Further recovery out of the channel Alternatively, the eigenvector U can be further determined from the channel H, and then the MSE of the two can be determined. which is, may also appear as Or, and are interchangeable.
  • the description of the performance measurement function is involved in many places below. For the sake of brevity, the description of the same or similar situations is omitted in the following.
  • the output condition may be that 90% of the performance metrics reach the lower bound of the performance metrics.
  • the performance measurement function is, for example, a calculation function of the correlation coefficient
  • the lower bound of the performance index is, for example, the lower bound of the correlation coefficient.
  • One or more correlation coefficients may be obtained by inputting one or more pairs of output results and originating data into the performance metric function.
  • the correlation coefficient with the originating data H can be expressed as or Corr() represents a function for finding the correlation coefficient.
  • the obtaining device may further judge whether the obtained one or more performance metric values reach the lower bound of the standardized correlation coefficient. If more than 90% of the performance metric values reach the lower bound of the correlation coefficient, the output condition is satisfied; if the number of performance metric values that reach the lower bound of the correlation coefficient is less than 90%, the output condition is not satisfied.
  • the MSE and the correlation coefficient are only examples, and the present application does not limit the performance indicators used by the acquisition device to determine whether the target neural network satisfies the output conditions and its related performance measurement functions. It should also be understood that reaching the lower bound of the performance index can be used as an example of the output condition of the target neural network, but should not constitute any limitation to the present application.
  • the obtaining device may obtain one or more originating data from the reference data set. For example, denote one of the originating data as X, and X is input to the originating reference neural network.
  • the compressed data X' can be obtained, and the originating reference neural network can further process the compressed data X' according to the standardized data structure to generate an air interface Information S.
  • the air interface information S can be input into the target neural network of the receiving end.
  • the target neural network at the receiving end can parse the air interface information S based on the standardized data structure, and recover the compressed data
  • the end-target neural network can be further based on the compressed data Get the output result Y.
  • the extraction and compression of channel information and the reconstruction of channel information are still taken as an example.
  • the originating data X may be channel H.
  • the originating reference neural network may sequentially perform preprocessing, information extraction and postprocessing on the channel H based on the operation flow of the originating neural network enumerated in FIG. 2 to generate air interface information S.
  • the target neural network at the receiving end can reconstruct the channel H based on the received air interface information S.
  • the target neural network at the receiving end can perform inverse post-processing, inverse information extraction and inverse pre-processing on the air interface information line successively based on the operation flow of the receiving end neural network listed in Figure 2 to recover the channel.
  • the output of the target neural network at the end is the estimated value of channel H It should be understood that the estimated value is also the output value.
  • the acquiring device may input the output result Y and the input originating data X into the performance measurement function to obtain the performance measurement value. It should be understood that the obtaining device may input the output result Y and the originating data X into one or more performance measurement functions to obtain one or more performance measurement values.
  • the obtaining device may also process one or more originating data obtained from the reference data set according to the above process, and multiple pairs of output results and originating data may be obtained.
  • Each pair of output result and originating data can be input into one or more performance metric functions to obtain one or more performance metric values.
  • Multiple pairs of output results and originating data can obtain multiple performance metrics.
  • the obtaining device may further compare the output performance metric value with the standardized performance index according to the standardized performance judgment criterion, and then determine whether the originating target neural network satisfies the preset output condition according to the comparison result.
  • FIG. 9 and FIG. 10 are schematic diagrams of a method for acquiring a neural network provided by an embodiment of the present application.
  • the neural network training methods shown in Figure 9 and Figure 10 use a standardized reference neural network as a reference to evaluate the training target neural network.
  • the target neural network and the reference neural network shown in FIG. 9 can be applied to the transmitter, the target neural network is the originating target neural network, and the reference neural network is the originating reference neural network.
  • the target neural network shown in FIG. 10 and the reference neural network can be applied to the receiver, the target neural network being the end target neural network and the reference neural network being the end reference neural network.
  • the obtaining device may obtain one or more originating data from the reference data set.
  • Each originating data may be input into the originating target neural network and the originating reference neural network to obtain target air interface information output from the originating target neural network and reference air interface information generated from the originating reference neural network.
  • the originating data X is input into the originating target neural network and the originating reference neural network.
  • Reference may be based on neural networks originating predesigned algorithm information extraction and preprocessing the originating data X, to obtain compressed data X 0 ', the neural network may be based on the originating reference standardized data structure of the compressed data of X 0 ' to process to obtain the reference air interface information S 0 generated by the originating reference neural network.
  • the originating target neural network can perform preprocessing and information extraction on the originating data X based on an algorithm corresponding to the receiving neural network to obtain the compressed data X 1 ′, and the originating target neural network can perform the compressed data X 1 ′ based on the standardized data structure.
  • the data X 1 ′ is processed to obtain target air interface information S 1 generated by the originating target neural network.
  • the preprocessing module used for preprocessing the originating data X may be a module shared by the originating reference neural network and the originating target neural network, that is, the process of preprocessing the originating data X is performed by the same preprocessing module .
  • the preprocessing module for preprocessing the originating data X may be a module included in the originating reference neural network and the originating target neural network respectively, that is, the process of preprocessing the originating data X is determined by the originating reference neural network and the originating target neural network.
  • the originating target neural network is executed separately.
  • the pre-processing module may also not be included in the originating reference neural network and the originating target neural network, and exists separately in the transmitter as a pre-processing module.
  • the preprocessing module can respectively input the preprocessed data to the originating reference neural network and the originating target neural network.
  • the originating data X is input to the originating target neural network and the originating reference neural network after being processed by the preprocessing module.
  • the extraction and compression of channel information and the reconstruction of channel information are still taken as an example.
  • the originating data X may be channel H.
  • the channel H may be input into the originating reference neural network and the originating target neural network respectively, and the originating reference neural network and the originating target neural network may follow the operation flow of the originating neural network as listed in FIG. 2 , respectively. , perform pre-processing, information extraction and post-processing on the channel H successively to obtain reference air interface information S 0 and target air interface information S 1 respectively .
  • the obtaining device may input the reference air interface information S 0 and the target air interface information S 1 into the performance measurement function to obtain the performance measurement value. It should be understood that the obtaining device may input the reference air interface information S 0 and the target air interface information S 1 into one or more performance measurement functions to obtain one or more performance measurement values.
  • the obtaining device may also process one or more originating data obtained from the reference data set according to the above process, and obtain multiple pairs of reference air interface information and target air interface information.
  • a pair of reference air interface information and target air interface information obtained from the same originating data may be input into one or more performance measurement functions to obtain one or more performance measurement values.
  • Multiple pairs of reference air interface information and target air interface information can obtain multiple performance metrics.
  • the obtaining device may further compare the output performance metric value with the standardized performance index according to the standardized performance judgment criterion, and then determine whether the originating target neural network satisfies the preset output condition according to the comparison result.
  • the acquisition device may input the same one or more air interface information to the reference neural network at the receiving end and the target neural network at the receiving end.
  • the one or more air interface information may be obtained directly from the reference data set, or may be air interface information obtained from a transmitter after the transmitter data obtained from the reference data set is input to the transmitter. This application does not limit this. As long as the air interface information input to the reference neural network at the receiving end and the target neural network at the receiving end each time is the same air interface information.
  • the air interface information input to the reference neural network of the receiving end and the target neural network of the receiving end may be obtained by the following methods: inputting the data of the transmitting end obtained in the reference data set into a reference neural network of the transmitting end,
  • the one originating reference neural network can process the input originating data according to a predefined data structure, and can obtain air interface information to be sent to the destination neural network of the destination and the reference neural network of the destination.
  • the air interface information output by the reference neural network at the originating end may be input to the reference neural network at the receiving end coupled to it.
  • the target neural network at the receiving end can be evaluated by a pair of the reference neural network at the sending end and the reference neural network at the receiving end that are coupled to each other.
  • the air interface information S can be input into the target neural network at the end and the reference neural network at the end.
  • the reference neural network at the receiving end can parse the air interface information S based on a standardized data structure, and can further process the analysis result based on a pre-designed algorithm to obtain the reference output result Y 0 output by the reference neural network at the receiving end.
  • the target neural network at the receiving end can parse the air interface information S based on the standardized data structure, and can further process the analysis results based on the corresponding algorithm of the reference neural network at the receiving end, so as to obtain the target output result Y output by the target neural network at the receiving end 1 .
  • the extraction and compression of channel information and the reconstruction of channel information are still taken as an example.
  • the air interface information input to the target neural network at the receiving end and the reference neural network at the receiving end are both S.
  • the reference neural network at the receiving end can perform inverse post-processing, inverse information extraction, and inverse pre-processing on the air interface information S successively based on the operation flow of the receiving end neural network listed in FIG. 2 to obtain the channel Estimated value of H
  • the target neural network at the receiving end can perform inverse post-processing, inverse information extraction, and inverse pre-processing successively on the air interface information S based on the operation flow of the receiving end neural network listed in FIG. 2 to obtain the channel Estimated value of H
  • the target neural network at the receiving end can perform inverse post-processing, inverse information extraction, and inverse pre-processing successively on the air interface information S based on the operation flow of the receiving end neural network listed in FIG. 2 to obtain the channel Estimated value of H
  • the acquisition device can output the reference to the result with the target output Enter into a performance measure function to obtain a performance measure. It should be understood that the acquisition device can output the reference with the target output Input to one or more performance measurement functions to obtain one or more performance measurement values.
  • the obtaining device may process one or more air interface information according to the above process to obtain multiple target output results output from the destination target neural network and multiple reference output results output from the reference neural network.
  • a pair of target output results and reference output results obtained from the same air interface information can be input into one or more performance measurement functions to obtain one or more performance measurement values. Multiple pairs of target output and reference output result in multiple performance metrics.
  • the obtaining device may further compare the output performance metric value with the standardized performance index according to the standardized performance judgment criterion, and then determine whether the originating target neural network satisfies the preset output condition according to the comparison result.
  • the acquisition device may also evaluate a target neural network using two different reference neural networks.
  • the two reference neural networks can generate two different output results based on the same input.
  • the two output results may be the output results of two reference neural networks designed to obtain different lower performance limits, respectively, or the two output results may also be two output results designed to obtain the upper and lower performance limits, respectively.
  • the output of a reference neural network Therefore, the two output results can be used to judge whether the target neural network satisfies the output conditions.
  • the two output results are respectively the output results of two different reference neural networks designed to obtain different performance lower bounds
  • the two output results can be respectively input to different performance metrics with the originating data. function in order to obtain different performance metric values, which can be compared with the lower bounds of the corresponding performance indicators.
  • both of the two output results exceed the lower bound of the performance index, it is determined that the target neural network satisfies the output condition. Therefore, it can be ensured that the performance of the target neural network is not lower than the lower limit when it is used.
  • the two output results are the output results of the two reference neural networks designed from the upper limit and the lower limit of the acquisition performance
  • the output of the reference neural network designed from the upper limit of the acquisition performance The result and the originating data can be input into the performance measurement function to obtain the performance measurement value, which can be used to compare with the upper bound of the performance index, so as to avoid the performance exceeding the upper limit when the target neural network is in use; obtain the performance from
  • the output result of the reference neural network designed starting from the lower bound of the performance index can be input into the performance metric function with the originating data to obtain another performance metric value, which can be used to compare with the lower bound of the performance index, so as to ensure the target The performance of the neural network does not fall below the lower limit when used.
  • 11 to 14 below illustrate the specific process of evaluating the target neural network by the acquisition device in combination with a plurality of standardized originating reference neural networks and/or terminating reference neural networks.
  • Fig. 11 shows the specific process of the acquisition device combining two standardized reference neural networks to determine whether the target neural network satisfies the output condition.
  • the target neural network can be cascaded with the reference neural network, the target neural network can be used in the transmitter, and is the target neural network at the origin; the reference neural network can be used in the receiver, and it is the reference neural network at the receiving end.
  • One of the two reference neural networks shown in FIG. 11 can be a neural network designed to obtain the lower limit of a certain performance (such as MSE), and the other (for example, the end reference neural network 2) in the figure is a neural network designed to obtain the lower limit of another performance (eg, correlation coefficient).
  • the air interface information S output by the originating target neural network is input to the receiving end reference neural network 1 and the receiving end reference neural network 2 respectively.
  • the receiving-end reference neural network 1 and the receiving-end reference neural network 2 can obtain output results based on the operation process of the receiving-end neural network listed in FIG.
  • the output result Y 01 can be obtained from the receiving-end reference neural network 1
  • the output result Y 02 can be obtained by referring to the neural network 2 at the receiving end.
  • the acquisition device can input the originating data X input to the originating target neural network and the output results Y 01 and Y 02 generated by the destination reference neural network into the performance measurement function, respectively, to obtain the performance metric value , and further according to the performance judgment criterion and the performance measurement value, it is determined whether the originating target neural network satisfies the output condition.
  • the performance metric function shown in Figure 11 includes MSE and correlation coefficient.
  • the performance metric value MSE(X, Y 01 ) can be obtained from the originating data X and the output result Y 01 ;
  • the performance metric value correlation coefficient (X, Y 02 ) can be obtained from the originating data X and the output result Y 02 .
  • the acquisition device further combines the lower bound of the standardized MSE and the lower bound of the correlation coefficient to determine whether the originating target neural network satisfies the output condition.
  • the output conditions are: the output result Y 01 generated by the receiving end reference neural network 1 and the performance metric value of the transmitting end data reach the lower bound of the MSE, and the output result Y 02 generated by the receiving end reference neural network 2 and the transmitting end.
  • the performance measure of the data reaches the lower bound of the correlation coefficient. Since the two reference neural networks at the receiving end outputting the above two output results are designed based on different performance lower bounds, respectively, the obtained performance metric values can be compared with the lower bounds of different performance indicators.
  • the originating target neural network satisfies the output condition.
  • the target neural network does not satisfy the output conditions. For example, for output condition 1, if MSE(X, Y 01 ) does not reach the lower bound of MSE, or the correlation coefficient (X, Y 02 ) does not reach the lower bound of the correlation coefficient, it can be determined that the originating target neural network does not meet the output condition .
  • the lower bound of the MSE and the lower bound of the correlation coefficient above may be a lower bound that matches the capability satisfied by the target neural network.
  • the acquisition device may select the corresponding performance index according to the achievable capability metric value satisfied by the selected target neural network to evaluate the target neural network.
  • the above two output results Y 01 and Y 02 may be obtained by processing the same air interface information by two different reference neural networks, for example, one is a high-complexity deep convolutional network, and the other is a low-complexity light-weight volume. network, which is not limited in this application.
  • more sinking reference neural networks are cascaded with the originating target neural network.
  • a part of these end-end reference neural networks (for example, denoted as end-end reference neural network set 1) may be a neural network designed to obtain the lower limit of performance, and the other part (such as denoted as end-end reference neural network set 2) may be It is a neural network designed to obtain the upper limit of performance.
  • end-end reference neural network set 1 A part of these end-end reference neural networks (for example, denoted as end-end reference neural network set 1) may be a neural network designed to obtain the lower limit of performance
  • the other part such as denoted as end-end reference neural network set 2 may be It is a neural network designed to obtain the upper limit of performance.
  • more output results of the end-end neural network can be obtained. These output results can also be respectively input into the performance measurement function together with the originating data input to the originating target neural network to obtain a plurality of performance metric values.
  • the output condition may be, for example: the result output by the reference neural network set 1 at the receiving end and the performance metric value of the data at the sending end both reach the lower bound of the performance index, and the result output by the reference neural network set 2 at the receiving end and the data at the sending end both reach the lower bound of the performance index. None of the performance metrics exceeded the upper bound of the performance metric.
  • the output condition may also be: the result output by the receiving end reference neural network set 1 and the performance metric value of the data at the sending end both reach the lower bound of the performance index, and the result output by the receiving end reference neural network set 2 and the performance of the data at the sending end More than 90% of the values in the metric do not exceed the upper bound of the performance metric, etc. For brevity, they are not listed here. It should be understood that the present application does not limit the specific content of the output conditions.
  • Fig. 12 shows the specific process of the acquisition device combining two standardized reference neural networks to determine whether the target neural network satisfies the output condition.
  • the target neural network in Figure 12 can be cascaded with the reference neural network.
  • the target neural network is applied to the receiver and is the target neural network at the receiving end;
  • the reference neural network is applied to the transmitter and is the reference neural network at the originating end.
  • One of the two reference neural networks shown in FIG. 12 (for example, the reference neural network 1 in the figure) may be a neural network set from the perspective of obtaining the lower limit of a certain performance (for example, MSE), and the other (for example, the reference neural network in the figure)
  • Neural network 2 is a neural network that is set up to obtain the lower limit of another performance (eg, correlation coefficient).
  • the air interface information S 01 can be obtained by the originating reference neural network 1
  • the air interface information S 02 can be obtained by the originating reference neural network 2
  • the target neural network at the receiving end can obtain output results Y 1 and Y 2 based on the air interface information S 01 and S 02 based on the operation flow of the receiving end neural network shown in FIG. 2 .
  • the acquisition device may be input to the neural network of the originating reference to the originating and terminating certain data X based on the output result of the neural network air interface information S 01 and S 02 respectively of Y 1 and Y 2 input performance metric function in order to obtain the performance measurement value, and further determine whether the target neural network satisfies the output condition according to the performance judgment criterion and the performance measurement value.
  • the performance metric function shown in Figure 12 includes MSE and correlation coefficient.
  • the performance metric value correlation coefficient (X, Y 2 ) can be obtained from the originating data X and the output result Y 2 .
  • the acquisition device further combines the lower bound of the standardized MSE and the lower bound of the correlation coefficient to determine whether the target neural network at the receiving end satisfies the output condition. Since the determination process has been described in detail above with reference to the performance determination criterion, for brevity, details are not repeated here.
  • Judging whether the target neural network satisfies the output conditions through the upper and lower bounds of the performance index is not limited to the method of cascading the target neural network and a reference neural network listed above.
  • the training device can also input the same one or more originating data into the originating target neural network and the originating reference neural network, and compare the output result of the originating target neural network with the output result of the originating reference neural network to determine the originating target Whether the neural network satisfies the output conditions.
  • Fig. 13 shows the specific process of the acquisition device combining two standardized reference neural networks to determine whether the target neural network satisfies the output condition.
  • Both the target neural network and the reference neural network can be applied to the transmitter, which are the originating target neural network and the originating reference neural network respectively.
  • One of the two originating reference neural networks shown in FIG. 13 may be a neural network set up to obtain a lower limit of a certain performance (such as MSE), and the other (for example, FIG.
  • the originating reference neural network in 2) is a neural network set from the lower limit of obtaining another performance (such as the correlation coefficient).
  • the same originating data X is input to originating reference neural network 1, originating reference neural network 2, and originating target neural network, respectively.
  • the originating reference neural network 1, the originating reference neural network 2, and the originating target neural network can process the originating data X respectively, for example, according to the operation flow of the originating neural network enumerated in FIG. air interface information.
  • Reference air interface information S 01 may be generated by the originating reference neural network 1
  • reference air interface information S 02 may be generated by the originating reference neural network 2
  • target air interface information S 1 may be generated by the originating target neural network.
  • the acquisition device inputs the target air interface information S 1 , the reference air interface information S 01 and S 02 into the performance measurement function to obtain the performance measurement value, and further according to the performance judgment criterion and the performance measurement value, Determine whether the target neural network satisfies the output condition. Since the determination process has been described in detail above with reference to the performance determination criterion, for brevity, details are not repeated here.
  • Fig. 14 shows the specific process of the acquisition device combining two standardized reference neural networks to determine whether the target neural network satisfies the output condition.
  • the reference neural network in FIG. 14 includes a reference neural network at the transmitter and a reference neural network at the receiver.
  • the target neural network is applied to the receiver and is the target neural network at the receiver.
  • the acquisition device may input the originating data acquired from the reference dataset to the originating reference neural network.
  • the originating reference neural network can generate air interface information S based on a standardized data interface.
  • the air interface information S can be input into the destination target neural network and the two destination reference neural networks.
  • one of the two reference neural networks at the receiving end may be a neural network designed to obtain the lower limit of a certain performance (such as MSE), and the other (for example, the reference neural network 1 in the figure)
  • the reference neural network at the end in 2) can be a neural network designed to obtain the lower limit of another performance (such as a correlation coefficient).
  • the receiving-end reference neural network 1, the receiving-end reference neural network 2, and the sending-end target neural network may analyze the air interface information S based on the standardized data structure, respectively, to obtain different output results.
  • the reference output result Y 01 can be obtained from the receiving end reference neural network 1
  • the reference output result Y 02 can be obtained from the receiving end reference neural network 2
  • the target output result Y 1 can be obtained from the receiving end target neural network.
  • the acquisition device can input the target output result Y 1 , the reference output results Y 01 and Y 02 into the performance measurement function to obtain the performance measurement value, and further according to the performance judgment criteria and the performance measurement value , to determine whether the target neural network at the receiving end satisfies the output condition. Since the determination process has been described in detail above with reference to the performance determination criterion, for brevity, details are not repeated here.
  • FIG. 7 to FIG. 14 mainly describe the specific process of step 610 by taking the modules included in the originating neural network and the sinking neural network shown in FIG. 2 as an example.
  • the input and output of the originating neural network may be different from those shown in the figure, and the input and output of the sinking neural network may also be the same as those shown in the figure. shown differently.
  • the acquisition device may perform one of step 620 or step 630 based on whether the target neural network satisfies the output condition. If the target neural network satisfies the output conditions, step 620 can be executed to output the target neural network; if the target neural network does not meet the output conditions, step 630 can be executed to continue to optimize the target neural network.
  • the process of optimizing the target neural network by the acquisition device may specifically include, but is not limited to, adjusting the neural network structure design, adjusting the reference data set, adjusting the neural network training method, adjusting the cost function, the definition of the loss function or the objective function, changing the neural network
  • the initialization method of the neural network the constraints of changing the parameters of the neural network, and the definition of the activation function of the neural network, etc.
  • the acquisition device can obtain an optimized target neural network.
  • the acquisition device can further evaluate the optimized target neural network to determine whether it satisfies the output conditions.
  • the relevant description in step 610 above which is not repeated here for brevity.
  • the acquisition device can evaluate the target neural network based on the existing standardized reference neural network.
  • the target neural network can be optimized until the target neural network satisfies the output conditions. Therefore, the overhead of downloading the neural network structure and parameter information through the air interface can be avoided.
  • it can also ensure that the products of different manufacturers can be interconnected.
  • it can also reflect the differentiated design and competitiveness of products of different manufacturers. Therefore, on the whole, it greatly improves the feasibility of dual network implementation. sex.
  • the target neural network obtained based on the method described above can be applied to a communication network.
  • the originating target neural network can be applied to the network device, and the receiving end target neural network can be applied to the terminal device.
  • the network device may comprise a transmitter as described above, and the terminal device may comprise a receiver as described above.
  • the originating target neural network can be applied to the terminal device, and the receiving end target neural network can be applied to the network device.
  • the terminal device may comprise a transmitter as described above, and the network device may comprise a receiver as described above. This application does not limit this.
  • the matching scene mode can be selected according to the current air interface transmission conditions.
  • the air interface transmission conditions may include, but are not limited to, channel status, interference status, and the like.
  • One of the network device and the terminal device may determine the scene mode based on the air interface transmission conditions in order to determine the target neural network for use with the current air interface transmission conditions.
  • FIG. 15 is a schematic flowchart of a communication method provided by an embodiment of the present application.
  • the method 700 shown in FIG. 15 may include steps 710 to 750 . Each step in the method 700 is described in detail below.
  • step 710 the terminal device sends capability information of the terminal device. Accordingly, the network device receives capability information of the terminal device.
  • the capability information may be used to indicate the achievable capability metric value.
  • the achievable capability metric value may specifically refer to the achievable capability metric value of the neural network that can be used by the terminal device.
  • the network device can select a matching neural network according to the achievable capability metric value reported by the terminal device and the scene mode determined by the air interface transmission condition with the terminal device.
  • step 720 the network device determines the scene mode based on the air interface transmission conditions with the terminal device.
  • the network device may obtain the air interface transmission conditions with the terminal device continuously or periodically, for example, by measuring the channel, measuring the interference, etc. to obtain the air interface transmission information.
  • the specific duration of the period may be a predefined value. This application does not limit the actual behavior of the network device to obtain the transmission conditions of the air interface.
  • the network device may determine the current scene mode based on air interface transmission conditions with the terminal device, such as channel state, interference state, and the like.
  • the scene mode may include, but is not limited to, different channel scenarios such as outdoor dense urban areas, outdoor ordinary urban areas, outdoor villages, outdoor mountainous areas, indoor offices, indoor workshops, etc., single interference source, multiple interference sources, and multiple interferences with different strengths Signal type, different signal-to-noise ratio conditions of high signal-to-noise ratio at mid-near point and low signal-to-noise ratio at far point, different moving conditions such as low walking speed, urban vehicle-mounted medium speed, high-speed rail ultra-high speed, etc.
  • the network device sends the indication information of the scene mode to the terminal device.
  • the terminal device receives the indication information of the scene mode from the network device.
  • the network device After the network device determines the scene mode, it can indicate the indication information of the scene mode to the terminal device through signaling, so that the terminal device can select a neural network that is one-to-one coupled with the network device to work together.
  • the network device may indicate the scene mode.
  • the correspondence between multiple indication bits and multiple scene modes may be predefined, and different scene modes are indicated by different indication bits.
  • the indication information of the scene mode is carried in higher layer signaling.
  • the higher layer signaling is, for example, a radio resource control (radio resource control, RRC) message or a medium access control (medium access control, MAC) control element (control element, CE).
  • RRC radio resource control
  • MAC medium access control
  • CE control element
  • the scene mode may be statically configured, or semi-statically configured.
  • the indication information of the scene mode is carried in physical layer signaling.
  • the physical layer signaling is, for example, downlink control information (downlink control information, DCI).
  • the scene mode may be dynamically configured.
  • step 740 the network device receives confirmation information from the terminal device, where the confirmation information is used to indicate the successful reception of the indication information of the scene mode.
  • the terminal device may send confirmation information to the network device.
  • the terminal device may send the confirmation information in the case of successful reception, and not transmit the confirmation information in the case of unsuccessful reception.
  • the network device may start a timer after sending the above-mentioned indication information of the scene mode, and the duration of the timer may be a predefined value. If the confirmation information is received before the timer times out, it can be determined that the terminal device has successfully received the indication information of the scene mode; otherwise, it is considered that the terminal device has not successfully received the indication information of the scene mode. In other words, step 740 is an optional step. If the terminal device fails to receive the indication information, the network device may fail to receive the confirmation information from the terminal device.
  • the confirmation information may be an indication bit. For example, “0” or “null” indicates unsuccessful reception; “1” indicates successful reception.
  • confirmation information may be various, which is not limited in this application.
  • the network device communicates with the terminal device using a neural network matching the scene mode and the capability information of the terminal device.
  • the terminal device also communicates with the network device using a neural network matching the scene mode and the capability information of the terminal device.
  • the network device may determine that the terminal device has successfully received the above-mentioned indication information. From now on, the network device and the terminal device can work with the originating neural network and the sinking neural network matching the scenario mode and the capability information of the terminal device.
  • the network device can work by using an originating neural network that matches the scenario mode and the capability information of the terminal device, and the terminal device can work using a receiving neural network that matches the scenario mode and capability information of the terminal device; for another example, the network The device can work with the receiving end neural network that matches the scenario mode and the capability information of the terminal device, and the terminal device can work using the originating neural network that matches the scenario mode and the capability information of the terminal device; for another example, the network device can use the While the originating neural network that matches a certain scene mode (such as scene mode 1) and the capability information of the terminal device works, it can also use another scene mode (such as scene mode 2) and the capability information of the terminal device.
  • a certain scene mode such as scene mode 1
  • the capability information of the terminal device works, it can also use another scene mode (such as scene mode 2) and the capability information of the terminal device.
  • the terminal device can use the end-end neural network that matches the scene mode 1 and the capability information of the terminal device to work, and can also use the end-end neural network that matches the scene mode 2 and the capability information of the terminal device. Network work. This application does not limit this.
  • each originating neural network has at least one terminating neural network that can cooperate with it; and each terminating neural network has at least one originating neural network that can cooperate with it.
  • both the originating neural network and the terminating neural network may be obtained based on the method in the above method 600 .
  • the network device may resend the indication information, for example, may re-execute the above-mentioned process until the confirmation information of the terminal device is received.
  • the network device may re-determine the scene mode when it finds that the transmission air interface conditions have changed, and in the case of a change in the scene mode, indicate the re-determined scene mode to the terminal device, so that the network device and the terminal device based on The newly identified scene patterns are worked with matching neural networks. That is, the above steps 720 to 750 can be repeatedly performed.
  • the network device can select an appropriate scene mode based on the current air interface transmission conditions with the terminal device, and indicate the scene mode to the terminal device, so that both the network device and the terminal device are based on the same scene mode and the terminal device.
  • FIG. 16 is a schematic flowchart of a communication method provided by another embodiment of the present application.
  • the method 800 shown in FIG. 16 may include steps 810 to 850 . Each step in the method 800 is described in detail below.
  • step 810 the terminal device reports capability information of the terminal device. Accordingly, the network device receives capability information of the terminal device.
  • the terminal device determines the scene mode based on the air interface transmission conditions with the network device.
  • step 830 the terminal device sends the indication information of the scene mode to the network device.
  • the network device receives the indication information of the scene mode of the terminal device.
  • step 840 the terminal device receives confirmation information from the network device, where the confirmation information is used to indicate the successful reception of the indication information of the scene mode.
  • the terminal device communicates with the network device using a neural network matching the scene mode and the capability information of the terminal device.
  • the network device communicates with the terminal device using a neural network matching the scene mode and the capability information of the terminal device.
  • step 820 to step 850 is similar to the specific process from step 720 to step 750 in the above method 700, the difference is that in method 800, the terminal device determines the scene mode according to the air interface transmission condition, and indicate the scene mode to the network device. The network device replies with confirmation information based on the successful reception of the indication information. After that, both the network device and the terminal device can work with the neural network matching the scene mode and the capability information of the terminal device.
  • step 720 to step 750 Since the specific processes from step 720 to step 750 have been described in detail in the method 700 above, for brevity, the detailed description is omitted here.
  • the terminal device may re-determine the scene mode when discovering that the transmission air interface conditions have changed, and indicate the re-determined scene mode to the network device when the scene mode changes, so as to facilitate the network device and the terminal device based on the new scenario mode.
  • the determined scene patterns are worked with matching neural networks. That is, the above steps 810 to 840 may also be repeatedly performed.
  • the terminal device can select an appropriate scene mode based on the current air interface transmission conditions with the network device, and indicate the scene mode to the network device, so that both the network device and the terminal device are based on the same scene mode and the terminal device.
  • the training method of the neural network in the communication network provided by the embodiment of the present application is described in detail with reference to FIG. 6 to FIG. 16 .
  • the apparatus for training the neural network in the communication network provided by the embodiments of the present application will be described in detail with reference to FIG. 17 to FIG. 20 .
  • FIG. 17 is a schematic block diagram of an apparatus 1000 for acquiring a neural network provided by an embodiment of the present application.
  • the training device 1000 includes a processing unit 1100 and an input and output unit 1200 .
  • the processing unit 1100 can be used to determine whether the constructed target neural network satisfies the preset output conditions based on the standardized reference data set, one or more reference neural networks and performance judgment criteria; the input and output unit 1200 can be used to satisfy the In the case of the output condition, the target neural network is output, and the output target neural network is the transmitter neural network applied to the transmitter or the receiver neural network applied to the receiver.
  • the apparatus 1000 may correspond to the acquisition device in the embodiments shown in FIGS. 6 to 14 , and may include a unit for executing the method performed by the acquisition device in the embodiment of the method 600 in FIGS. 6 to 14 .
  • the processing unit 1100 can be used to perform steps 610 and 620 in the above method 600
  • the input and output unit 1200 can be used to perform step 630 in the above method 600 .
  • the steps executed by the processing unit 1100 may be implemented by one or more processors executing corresponding programs.
  • the training of the target neural network can be implemented by a processor dedicated to training the neural network.
  • the processing unit 1100 may correspond to the processor 2010 shown in FIG. 18 .
  • the steps performed by the input-output unit 1200 may be implemented by, for example, input-output interfaces, circuits, and the like.
  • the input/output unit 1200 may correspond to, for example, the input/output interface 2020 shown in FIG. 18 .
  • the apparatus 1000 may be deployed on a chip.
  • FIG. 18 is a schematic block diagram of an apparatus 2000 for acquiring a neural network provided by an embodiment of the present application.
  • the apparatus 2000 includes a processor 2100 , an input/output interface 2200 and a memory 2030 .
  • the processor 2010, the input/output interface 2020 and the memory 2030 communicate with each other through an internal connection path, the memory 2030 is used to store instructions, and the processor 2010 is used to execute the instructions stored in the memory 2030 to control the input/output interface 2020 to send signal and/or receive signal.
  • the apparatus 2000 may correspond to the acquisition device in the embodiments shown in FIGS. 6 to 14 , and may include a unit for executing the method performed by the acquisition device in the embodiment of the method 600 in FIGS. 6 to 14 .
  • the memory 2030 may include read only memory and random access memory and provide instructions and data to the processor. A portion of the memory may also include non-volatile random access memory.
  • the memory 2030 may be a separate device or may be integrated in the processor 2010 .
  • the processor 2010 may be configured to execute the instructions stored in the memory 2030, and when the processor 2010 executes the instructions stored in the memory, the processor 2010 is configured to execute various steps and/or steps of the above-mentioned method embodiments corresponding to the acquisition device or process.
  • FIG. 19 is a schematic block diagram of a communication apparatus provided by an embodiment of the present application.
  • the communication apparatus 3000 may include a processing unit 3100 and a transceiver unit 3200 .
  • the communication apparatus 3000 may correspond to the terminal device in the above method embodiments, for example, may be a terminal device, or a component (such as a circuit, a chip or a chip system, etc.) configured in the terminal device.
  • the communication apparatus 3000 may correspond to the terminal device in the method 700 or the method 800 according to the embodiment of the present application, and the communication apparatus 3000 may include a terminal for executing the method 700 in FIG. 15 or the method 800 in FIG. 16 .
  • each unit in the communication apparatus 3000 and the above-mentioned other operations and/or functions are respectively to implement the corresponding flow of the method 700 in FIG. 15 or the method 800 in FIG. 16 .
  • the processing unit 3100 can be used to execute step 750 of the method 700
  • the transceiver unit 3200 can be used to execute steps 710 , 730 to 750 of the method 700 . It should be understood that the specific process of each unit performing the above-mentioned corresponding steps has been described in detail in the above-mentioned method embodiments, and for the sake of brevity, it will not be repeated here.
  • the processing unit 3100 can be used to execute the steps 820 and 850 of the method 800
  • the transceiver unit 3200 can be used to execute the steps 810 , 830 to 850 of the method 800 .
  • the transceiver unit 3200 in the communication apparatus 3000 may be implemented by a transceiver, for example, may correspond to the transceiver 4020 in the communication apparatus 4000 shown in FIG. 20 or the transceiver 4020 in FIG. 21 .
  • the transceiver 5020 in the illustrated terminal device 5000, the processing unit 3100 in the communication apparatus 3000 may be implemented by at least one processor, for example, may correspond to the processor 4010 in the communication apparatus 4000 shown in FIG. 20 or the processor 4010 in FIG. 21
  • the transceiver unit 3200 in the communication apparatus 3000 may be implemented through input/output interfaces, circuits, etc., and the processing unit 3100 in the communication apparatus 3000 It can be implemented by a processor, microprocessor or integrated circuit integrated on the chip or chip system.
  • the communication apparatus 3000 may correspond to the network device in the above method embodiments, for example, may be a network device, or a component (such as a circuit, a chip, or a chip system, etc.) configured in the network device.
  • the communication apparatus 3000 may correspond to the network device in the method 700 or the method 800 according to the embodiment of the present application, and the communication apparatus 3000 may include a network device for executing the method 700 in FIG. 15 or the method 800 in FIG. 16 .
  • each unit in the communication apparatus 3000 and the above-mentioned other operations and/or functions are respectively to implement the corresponding flow of the method 700 in FIG. 15 or the method 800 in FIG. 16 .
  • the processing unit 3100 can be used to execute steps 720 and 750 in the method 700
  • the transceiver unit 3200 can be used to execute steps 710 , 730 to 750 in the method 700 . 750. It should be understood that the specific process of each unit performing the above-mentioned corresponding steps has been described in detail in the above-mentioned method embodiments, and for the sake of brevity, it will not be repeated here.
  • the processing unit 3100 can be used to execute the step 850 of the method 800
  • the transceiver unit 3200 can be used to execute the steps 810 , 830 to 850 of the method 800 . It should be understood that the specific process of each unit performing the above-mentioned corresponding steps has been described in detail in the above-mentioned method embodiments, and for the sake of brevity, it will not be repeated here.
  • the transceiver unit 3200 in the communication apparatus 3000 may be implemented by a transceiver, for example, it may correspond to the transceiver 4020 in the communication apparatus 4000 shown in FIG. 20 or the transceiver 4020 in FIG. 22 .
  • the RRU 6100 in the shown base station 6000, the processing unit 3100 in the communication device 3000 may be implemented by at least one processor, for example, may correspond to the processor 4010 in the communication device 4000 shown in FIG. 20 or the processor 4010 shown in FIG. 22
  • the processing unit 6200 or the processor 6202 in the outgoing base station 6000 may be implemented by a transceiver, for example, it may correspond to the transceiver 4020 in the communication apparatus 4000 shown in FIG. 20 or the transceiver 4020 in FIG. 22 .
  • the transceiver unit 3200 in the communication apparatus 3000 may be implemented through input/output interfaces, circuits, etc., and the processing unit 3100 in the communication apparatus 3000 It can be implemented by a processor, microprocessor or integrated circuit integrated on the chip or chip system.
  • FIG. 20 is another schematic block diagram of a communication apparatus 4000 provided by an embodiment of the present application.
  • the communication device 4000 includes a processor 2010 , a transceiver 4020 and a memory 4030 .
  • the processor 4010, the transceiver 4020 and the memory 4030 communicate with each other through an internal connection path, the memory 4030 is used to store instructions, and the processor 4010 is used to execute the instructions stored in the memory 4030 to control the transceiver 4020 to send signals and / or receive signals.
  • the communication apparatus 4000 may correspond to the terminal device in the above method embodiments, and may be used to execute various steps and/or processes performed by the network device or the terminal device in the above method embodiments.
  • the memory 4030 may include read only memory and random access memory and provide instructions and data to the processor. A portion of the memory may also include non-volatile random access memory.
  • the memory 4030 may be a separate device or may be integrated in the processor 4010 .
  • the processor 4010 may be configured to execute the instructions stored in the memory 4030, and when the processor 4010 executes the instructions stored in the memory, the processor 4010 is configured to execute each of the foregoing method embodiments corresponding to the network device or the terminal device steps and/or processes.
  • the communication apparatus 4000 is the terminal device in the foregoing embodiment.
  • the communication apparatus 4000 is the network device in the foregoing embodiment.
  • the transceiver 4020 may include a transmitter and a receiver.
  • the transceiver 4020 may further include antennas, and the number of the antennas may be one or more.
  • the processor 4010, the memory 4030 and the transceiver 4020 may be devices integrated on different chips.
  • the processor 4010 and the memory 4030 may be integrated in the baseband chip, and the transceiver 4020 may be integrated in the radio frequency chip.
  • the processor 4010, the memory 4030 and the transceiver 4020 may also be devices integrated on the same chip. This application does not limit this.
  • the communication apparatus 4000 is a component configured in a terminal device, such as a circuit, a chip, a chip system, and the like.
  • the communication apparatus 4000 is a component configured in a network device, such as a circuit, a chip, a chip system, and the like.
  • the transceiver 4020 may also be a communication interface, such as an input/output interface, a circuit, and the like.
  • the transceiver 4020, the processor 4010 and the memory 4020 can be integrated in the same chip, such as integrated in a baseband chip.
  • FIG. 21 is a schematic structural diagram of a terminal device 5000 provided by an embodiment of the present application.
  • the terminal device 5000 can be applied to the system shown in FIG. 1 to perform the functions of the terminal device in the foregoing method embodiments.
  • the terminal device 5000 includes a processor 5010 and a transceiver 5020 .
  • the terminal device 5000 further includes a memory 5030 .
  • the processor 5010, the transceiver 5020 and the memory 5030 can communicate with each other through an internal connection path to transmit control and/or data signals.
  • the memory 5030 is used to store computer programs, and the processor 5010 is used to retrieve data from the memory 5030.
  • the computer program is called and executed to control the transceiver 5020 to send and receive signals.
  • the terminal device 5000 may further include an antenna 3040 for sending the uplink data or uplink control signaling output by the transceiver 5020 through wireless signals.
  • the above-mentioned processor 5010 and the memory 5030 can be combined into a processing device, and the processor 5010 is configured to execute the program codes stored in the memory 5030 to realize the above-mentioned functions.
  • the memory 5030 may also be integrated in the processor 5010 or independent of the processor 5010 .
  • the processor 5010 may correspond to the processing unit 3100 in FIG. 19 or the processor 4010 in FIG. 20 .
  • the transceiver 5020 described above may correspond to the transceiver unit 3200 in FIG. 19 or the transceiver 4020 in FIG. 20 .
  • the transceiver 5020 may include a receiver (or called receiver, receiving circuit) and a transmitter (or called transmitter, transmitting circuit). Among them, the receiver is used for receiving signals, and the transmitter is used for transmitting signals.
  • the terminal device 5000 shown in FIG. 21 can implement various processes involving the terminal device in the method embodiment shown in FIG. 15 or FIG. 16 .
  • the operations and/or functions of each module in the terminal device 5000 are respectively to implement the corresponding processes in the foregoing method embodiments.
  • the above-mentioned processor 5010 may be used to perform the actions described in the foregoing method embodiments that are implemented inside the terminal device, and the transceiver 5020 may be used to perform the actions described in the foregoing method embodiments that the terminal device sends to or receives from the network device. action.
  • the transceiver 5020 may be used to perform the actions described in the foregoing method embodiments that the terminal device sends to or receives from the network device. action.
  • the above-mentioned terminal device 5000 may further include a power supply 5050 for providing power to various devices or circuits in the terminal device.
  • the terminal device 5000 may further include one or more of an input unit 5060, a display unit 5070, an audio circuit 5080, a camera 5090, a sensor 5100, etc., the audio circuit Speakers 5082, microphones 5084, etc. may also be included.
  • FIG. 22 is a schematic structural diagram of a network device provided by an embodiment of the present application, which may be, for example, a schematic structural diagram of a base station.
  • the base station 6000 can be applied to the system as shown in FIG. 1 to perform the functions of the network device in the foregoing method embodiments.
  • the base station 6000 may include one or more radio frequency units, such as a remote radio unit (RRU) 6100 and one or more baseband units (BBUs) (also referred to as distributed units (DUs) )) 6200.
  • RRU 6100 may be called a transceiver unit, which may correspond to the transceiver unit 3200 in FIG. 19 or the transceiver 4020 in FIG. 20 .
  • the RRU 6100 may also be referred to as a transceiver, a transceiver circuit, or a transceiver, etc., which may include at least one antenna 6101 and a radio frequency unit 6102.
  • the RRU 6100 may include a receiving unit and a sending unit, the receiving unit may correspond to a receiver (or called a receiver, a receiving circuit), and the sending unit may correspond to a transmitter (or called a transmitter, a sending circuit).
  • the part of the RRU 6100 is mainly used for transmitting and receiving radio frequency signals and converting radio frequency signals to baseband signals, for example, for sending indication information to terminal equipment.
  • the part of the BBU 6200 is mainly used to perform baseband processing and control the base station.
  • the RRU 6100 and the BBU 6200 may be physically set together, or may be physically set apart, that is, a distributed base station.
  • the BBU 6200 is the control center of the base station, and can also be called a processing unit, which can correspond to the processing unit 3100 in Figure 19 or the processor 4010 in Figure 20, and is mainly used to complete baseband processing functions, such as channel coding, multiplexing , modulation, spread spectrum, etc.
  • the BBU processing unit
  • the BBU may be used to control the base station to perform the operation procedure of the network device in the foregoing method embodiments, for example, to generate the foregoing indication information and the like.
  • the BBU 6200 may be composed of one or more single boards, and multiple single boards may jointly support a wireless access network (such as an LTE network) of a single access standard, or may respectively support a wireless access network of different access standards.
  • Wireless access network (such as LTE network, 5G network or other network).
  • the BBU 6200 also includes a memory 6201 and a processor 6202.
  • the memory 6201 is used to store necessary instructions and data.
  • the processor 6202 is configured to control the base station to perform necessary actions, for example, to control the base station to perform the operation flow of the network device in the foregoing method embodiments.
  • the memory 6201 and the processor 6202 may serve one or more single boards. That is to say, the memory and processor can be provided separately on each single board. It can also be that multiple boards share the same memory and processor. In addition, necessary circuits may also be provided on each single board.
  • the base station 6000 shown in FIG. 22 can implement various processes involving network devices in the method embodiment shown in FIG. 15 or FIG. 16 .
  • the operations and/or functions of each module in the base station 6000 are respectively to implement the corresponding processes in the foregoing method embodiments.
  • the above-mentioned BBU 6200 may be used to perform the actions described in the foregoing method embodiments that are implemented internally by the network device, while the RRU 6100 may be used to perform the actions described in the foregoing method embodiments that the network device sends to or receives from the terminal device.
  • the RRU 6100 may be used to perform the actions described in the foregoing method embodiments that the network device sends to or receives from the terminal device.
  • the base station 6000 shown in FIG. 22 is only a possible form of network equipment, and should not constitute any limitation to the present application.
  • the method provided in this application can be applied to other forms of network equipment.
  • it includes AAU, may also include CU and/or DU, or includes BBU and adaptive radio unit (ARU), or BBU; may also be customer terminal equipment (customer premises equipment, CPE), may also be
  • AAU adaptive radio unit
  • BBU adaptive radio unit
  • CPE customer premises equipment
  • the CU and/or DU may be used to perform the actions implemented by the network device described in the foregoing method embodiments, and the AAU may be used to execute the network device described in the foregoing method embodiments to send or receive from the terminal device. Actions. For details, please refer to the descriptions in the foregoing method embodiments, which will not be repeated here.
  • the present application further provides a processing apparatus, including at least one processor, where the at least one processor is configured to execute a computer program stored in a memory, so that the processing apparatus executes the method, A method performed by a network device or a method performed by a terminal device.
  • the embodiment of the present application also provides a processing device, which includes a processor and an input and output interface.
  • the input-output interface is coupled to the processor.
  • the input and output interface is used for inputting and/or outputting information.
  • the information includes at least one of instructions and data.
  • the processor is configured to execute a computer program, so that the processing apparatus executes the method executed by the acquisition device, the method executed by the network device, or the method executed by the terminal device in the above method embodiments.
  • Embodiments of the present application further provide a processing apparatus, including a processor and a memory.
  • the memory is used to store a computer program
  • the processor is used to call and run the computer program from the memory, so that the processing device executes the method performed by the training device in the above method embodiments.
  • the above-mentioned processing device may be one or more chips.
  • the processing device may be a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), a system on chip (SoC), or a It is a central processing unit (CPU), a network processor (NP), a digital signal processing circuit (DSP), or a microcontroller (microcontroller unit). , MCU), it can also be a programmable logic device (PLD) or other integrated chips.
  • FPGA field programmable gate array
  • ASIC application specific integrated circuit
  • SoC system on chip
  • MCU microcontroller unit
  • MCU programmable logic device
  • PLD programmable logic device
  • each step of the above-mentioned method can be completed by a hardware integrated logic circuit in a processor or an instruction in the form of software.
  • the steps of the methods disclosed in conjunction with the embodiments of the present application may be directly embodied as executed by a hardware processor, or executed by a combination of hardware and software modules in the processor.
  • the software modules may be located in random access memory, flash memory, read-only memory, programmable read-only memory or electrically erasable programmable memory, registers and other storage media mature in the art.
  • the storage medium is located in the memory, and the processor reads the information in the memory, and completes the steps of the above method in combination with its hardware. To avoid repetition, detailed description is omitted here.
  • the processor in this embodiment of the present application may be an integrated circuit chip, which has a signal processing capability.
  • each step of the above method embodiments may be completed by a hardware integrated logic circuit in a processor or an instruction in the form of software.
  • the aforementioned processors may be general purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components .
  • DSPs digital signal processors
  • ASICs application specific integrated circuits
  • FPGAs field programmable gate arrays
  • the methods, steps, and logic block diagrams disclosed in the embodiments of this application can be implemented or executed.
  • a general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.
  • the steps of the method disclosed in conjunction with the embodiments of the present application may be directly embodied as executed by a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor.
  • the software modules may be located in random access memory, flash memory, read-only memory, programmable read-only memory or electrically erasable programmable memory, registers and other storage media mature in the art.
  • the storage medium is located in the memory, and the processor reads the information in the memory, and completes the steps of the above method in combination with its hardware.
  • the memory in this embodiment of the present application may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memory.
  • the non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically programmable Erase programmable read-only memory (electrically EPROM, EEPROM) or flash memory.
  • Volatile memory may be random access memory (RAM), which acts as an external cache.
  • RAM random access memory
  • DRAM dynamic random access memory
  • SDRAM synchronous DRAM
  • SDRAM double data rate synchronous dynamic random access memory
  • ESDRAM enhanced synchronous dynamic random access memory
  • SLDRAM synchronous link dynamic random access memory
  • direct rambus RAM direct rambus RAM
  • the present application also provides a computer program product, the computer program product includes: computer program code, when the computer program code is run on a computer, the computer is made to execute the embodiment shown in FIG. 6 .
  • the computer program product includes: computer program code, when the computer program code is run on a computer, the computer is made to execute the embodiment shown in FIG. 6 .
  • the present application further provides a computer-readable storage medium, where the computer-readable storage medium stores program codes, and when the program codes are run on a computer, the computer executes the program shown in FIG. 6 .
  • the present application also provides a system, which includes the aforementioned one or more terminal devices and one or more network devices, the one or more terminal devices and the one or more terminal devices
  • the target neural network obtained by the foregoing obtaining device is respectively configured in the network device.
  • the network equipment in each of the above apparatus embodiments completely corresponds to the terminal equipment and the network equipment or terminal equipment in the method embodiments, and corresponding steps are performed by corresponding modules or units.
  • a processing unit processor
  • processor For functions of specific units, reference may be made to corresponding method embodiments.
  • the number of processors may be one or more.
  • the terminal device may be used as an example of a receiving device
  • the network device may be used as an example of a sending device.
  • the sending device and the receiving device may both be terminal devices or the like. This application does not limit the specific types of the sending device and the receiving device.
  • a component may be, but is not limited to, a process running on a processor, a processor, an object, an executable, a thread of execution, a program, and/or a computer.
  • an application running on a computing device and the computing device may be components.
  • One or more components may reside within a process and/or thread of execution, and a component may be localized on one computer and/or distributed between 2 or more computers.
  • these components can execute from various computer readable media having various data structures stored thereon.
  • a component may, for example, be based on a signal having one or more data packets (eg, data from two components interacting with another component between a local system, a distributed system, and/or a network, such as the Internet interacting with other systems via signals) Communicate through local and/or remote processes.
  • data packets eg, data from two components interacting with another component between a local system, a distributed system, and/or a network, such as the Internet interacting with other systems via signals
  • the disclosed system, apparatus and method may be implemented in other manners.
  • the apparatus embodiments described above are only illustrative.
  • the division of the units is only a logical function division. In actual implementation, there may be other division methods.
  • multiple units or components may be combined or Can be integrated into another system, or some features can be ignored, or not implemented.
  • the shown or discussed mutual coupling or direct coupling or communication connection may be through some interfaces, indirect coupling or communication connection of devices or units, and may be in electrical, mechanical or other forms.
  • the units described as separate components may or may not be physically separated, and components displayed as units may or may not be physical units, that is, may be located in one place, or may be distributed to multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution in this embodiment.
  • each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically alone, or two or more units may be integrated into one unit.
  • the functions, if implemented in the form of software functional units and sold or used as independent products, may be stored in a computer-readable storage medium.
  • the technical solution of the present application can be embodied in the form of a software product in essence, or the part that contributes to the prior art or the part of the technical solution.
  • the computer software product is stored in a storage medium, including Several instructions are used to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application.
  • the aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk and other media that can store program codes .

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Signal Processing (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Software Systems (AREA)
  • Evolutionary Computation (AREA)
  • Biophysics (AREA)
  • Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Molecular Biology (AREA)
  • Computer Hardware Design (AREA)
  • Quality & Reliability (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Arrangements For Transmission Of Measured Signals (AREA)
  • Feedback Control In General (AREA)

Abstract

本申请提供了一种获取神经网络的方法和装置，以期减少发射机和接收机下载神经网络结构和参数信息等带来的空口开销。该方法包括：基于标准化的参考数据集、标准化的至少一个参考神经网络和标准化的性能判断准则，确定构建的目标神经网络是否满足预设的输出条件；在满足输出条件的情况下，输出该目标神经网络；在不满足输出条件的情况下，对该目标神经网络进行优化，直至获得的目标神经网络满足输出条件；其中，该目标神经网络是应用于发射机的发端目标神经网络或应用于接收机的收端目标神经网络。

Description

获取神经网络的方法和装置
本申请要求于2020年06月30日提交中国国家知识产权局、申请号为202010615552.9、申请名称为“获取神经网络的方法和装置”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及人工智能领域，并且更具体地，涉及获取神经网络的方法和装置。
背景技术
在通信系统中，发射机用于发送信号，接收机用于接收发射机发送的信号。神经网络可应用于发射机和接收机，对收发两端进行联合优化，以提高整体性能。以信道信息(channel state information，CSI)压缩和重构为例，发射机可基于神经网络对CSI进行压缩，生成空口信息。接收机可基于神经网络对空口信息进行解析，以重构信道。在一种实现方式中，可以采用联合发射端的神经网络与接收端的神经网络设计来实现算法的联合优化，获得最优的压缩和重构算法。
然而，在实际商用系统中，接收机和发射机往往是由不同的厂商提供的。例如，发射机由A厂商提供，接收机由B厂商提供。为了解决联合优化问题，一种实现方式是，发射机和接收机分别下载匹配当前使用场景的神经网络结构和参数信息，再依据下载的神经网络进行收、发两端的处理。这会造成空口的巨大开销，在实际使用中受到较大局限。
发明内容
本申请提供一种获取神经网络的方法和装置，以期减少发射机和接收机下载神经网络结构和参数信息等所带来的巨大的空口开销。
第一方面，提供了一种获取神经网络的方法。该方法包括：基于标准化的参考数据集、一个或多个参考神经网络和性能判断准则，确定构建的目标神经网络是否满足预设的输出条件；在满足所述输出条件的情况下，输出所述目标神经网络，输出的所述目标神经网络是应用于发射机的发端目标神经网络或应用于接收机中的收端目标神经网络。
基于上述技术方案，通过引入标准化的参考数据集、标准化的参考神经网络以及标准化的性能判断准则，使得获取设备可以基于已有的标准化的参考神经网络，对目标神经网络进行评估，在目标神经网络满足输出条件的情况下才将其输出使用。输出的目标神经网络可应用于发射机或接收机中。从而使得发射机或接收机通过空口来下载神经网络结构和参数信息的开销得以避免。同时还可以保证不同厂商的产品之间能够互联互通，在保证性能的前提下，还可以体现不同厂商的产品的差异化设计和竞争力，因此从整体上说，大幅提升了对偶网络实现的可行性。
结合第一方面，在第一方面的某些可能的实现方式中，所述方法还包括：在不满足所 述输出条件的情况下，继续对所述目标神经网络进行优化，直至获得的目标神经网络满足所述输出条件。
即，在不满足输出条件的情况下，还可继续对目标神经网络进行优化。所述优化可以包括但不限于，调整神经网络结构设计、调整参考数据集、调整神经网络训练方法、调整代价函数、损失函数或目标函数的定义、改变神经网络的初始化方法、改变神经网络参数的约束条件、以及改变神经网络的激活函数的定义等。
通过对目标神经网络的优化，所得到的优化后的目标神经网络可继续基于上述性能判断准则进行评估，如此循环，指导优化后的目标神经网络满足输出条件，进而将其输出。
结合第一方面，在第一方面的某些可能的实现方式中，所述发端目标神经网络用于星座调制、信道编码、预编码、或信道信息的提取和压缩；所述收端目标神经网络用于星座解调、信道解码、检测接收、或信道信息的重构。
换言之，该目标神经网络可以应用于通信的各个阶段。发端目标神经网络和收端目标神经网络可以相互耦合，协同工作。例如，协同工作的发端目标神经网络和收端目标神经网络可分别用于星座调制与解调、或信道编码与解码、或预编码与检测接收、或信道信息的提取和压缩与信道信息的重构。
结合第一方面，在第一方面的某些可能的实现方式中，所述方法还包括：从至少一套标准化的配置中确定适用于所述目标神经网络当前应用场景的配置；每套标准化的配置至少包括以下一项或多项：标准化的参考数据集、标准化的参考神经网络、标准化的数据结构、标准化的性能度量函数、标准化的性能判断准则、以及标准化的可实现能力度量值。
其中，所述参考神经网络包括应用于接收机中的收端参考神经网络和/或应用于发射机中的发端参考神经网络；所述数据结构用于生成和/或解析空口信息；所述性能度量函数用于生成性能度量值，所述性能判断准则包括性能指标和所述目标神经网络的输出条件，所述性能判断准则用于基于所述性能度量值和所述性能指标，确定所述目标神经网络是否满足所述输出条件；所述可实现能力度量值包括神经网络用于选择匹配的神经网络和/或匹配的性能指标。
应理解，这里所述的匹配的神经网络，具体可以是指，该神经网络在给定的约束下，能够应用在发射机或接收机中，所述给定的约束可以包括实现复杂度和/或空口开销。匹配的性能指标，具体可以是指，用于获取该神经网络的阶段对该神经网络进行评估所使用的性能指标。
针对不同的应用场景可以定义不同的标准化的配置。例如，在不同的应用场景中，性能判断准则可以不同；或，参考神经网络可以不同，等等。基于这样的设计，可以使得输出的目标神经网络与不同的应用场景相匹配，从而有利于获得较好的系统性能。
可选地，所述性能度量函数包括以下一项或多项：均方误差、归一化均方误差、平均绝对误差、最大绝对误差、相关系数、交叉熵、互信息、误比特率、或误帧率。
上文所列举的性能度量函数仅为示例，不应对本申请构成任何限定。性能度量函数在用于对目标神经网络进行评估时，可以同时使用其中的一项或多项来评估。本申请对此不作限定。
可选地，所述性能指标包括一个或多个性能指标，所述性能判断准则包括基于所述性能度量值与所述一个或多个性能指标的比较，确定是否满足所述输出条件。
应理解，所述性能指标可以为一项或多项。与之对应的性能指标的上界和/或下界也可以为一项或多项。
下文示出了结合上述性能判断准则，对目标神经网络进行评估，以确定其是否满足输出条件的几个示例。
在一种可能的设计中，所述目标神经网络为发端目标神经网络，所述一个或多个参考神经网络为一个或多个收端参考神经网络；以及，所述基于标准化的参考数据集、一个或多个参考神经网络和性能判断准则，确定设计得到的目标神经网络是否满足预设的输出条件，包括：将从所述参考数据集中获取到的发端数据输入至所述发端目标神经网络中，所述发端目标神经网络用于根据预定义的数据结构对所述发端数据进行处理，以生成待发送给所述一个或多个收端参考神经网络的空口信息；从所述一个或多个收端参考神经网络获取一个或多个输出结果，所述一个或多个输出结果由所述一个或多个收端参考神经网络分别基于接收到的所述空口信息得到；将所述一个或多个输出结果和所述发端数据作为所述性能度量函数的输入，以得到一个或多个性能度量值；根据所述一个或多个性能度量值，以及所述性能判断准则，确定所述目标神经网络是否满足所述输出条件。
即，将发端目标神经网络和收端参考神经网络级联来对收端目标神经网络进行评估。
在另一种可能的设计中，所述目标神经网络为发端目标神经网络，所述一个或多个参考神经网络为一个或多个发端参考神经网络；以及，所述基于标准化的参考数据集、一个或多个参考神经网络和性能判断准则，确定设计得到的目标神经网络是否满足预设的输出条件，包括：将从所述参考数据集中获取到的发端数据输入至所述发端目标神经网络中，所述发端目标神经网络用于根据预定义的数据结构对所述发端数据进行处理，以生成目标空口信息；将所述发端数据输入至所述一个或多个发端参考神经网络中，所述一个或多个发端参考神经网络用于根据预定义的数据结构对所述发端数据进行处理，以生成一个或多个参考空口信息；将所述一个或多个参考空口信息和所述目标空口信息作为所述性能度量函数的输入，以得到一个或多个性能度量值；根据所述一个或多个性能度量值，以及所述性能判断准则，确定所述发端目标神经网络是否满足所述输出条件。
即，将发端参考神经网络作为参考，来对发端目标神经网络进行评估。
在又一种可能的设计中，所述目标神经网络为收端目标神经网络，所述一个或多个参考神经网络为一个或多个发端参考神经网络；以及，所述基于标准化的参考数据集、一个或多个参考神经网络和性能判断准则，确定设计得到的目标神经网络是否满足预设的输出条件，包括：将从所述参考数据集中获取到的发端数据输入至所述一个或多个发端参考神经网络中，所述一个或多个发端参考神经网络用于根据预定义的数据结构对所述发端数据进行处理，以得到待发送给所述收端目标神经网络的一个或多个空口信息；从所述收端目标神经网络获取一个或多个输出结果，所述一个或多个输出结果是所述收端目标神经网络根据接收到的所述一个或多个空口信息分别生成的输出结果；将所述一个或多个输出结果和所述发端数据作为所述性能度量函数的输入，以得到一个或多个性能度量值；根据所述一个或多个性能度量值，以及所述性能判断准则，确定所述收端目标神经网络是否满足所述输出条件。
即，将收端目标神经网络和发端参考神经网络级联来对收端目标神经网络进行评估。
在再一种可能的设计中，所述目标神经网络为收端目标神经网络，所述一个或多个参 考神经网络包括一个或多个发端参考神经网络和一个或多个收端参考神经网络；以及，所述基于标准化的参考数据集、一个或多个参考神经网络和性能判断准则，确定设计得到的目标神经网络是否满足预设的输出条件，包括：将从所述参考数据集中获取到的发端数据输入至所述一个或多个发端参考神经网络中，所述一个或多个发端参考神经网络用于根据预定义的数据结构对所述发端数据进行处理，以得到待发送给所述收端目标神经网络和所述一个或多个收端参考神经网络的一个或多个空口信息；从所述收端目标神经网络获取一个或多个目标输出结果，所述一个或多个目标输出结果是所述收端目标神经网络根据接收到的所述一个或多个空口信息分别生成的输出结果；从所述一个或多个收端参考神经网络获取一个或多个参考输出结果，所述一个或多个参考输出结果是由所述一个或多个收端参考神经网络分别根据接收到的所述一个或多个空口信息生成的输出结果；将所述一个或多个目标输出结果和所述一个或多个参考输出结果作为所述性能度量函数的输入，以得到一个或多个性能度量值；根据所述一个或多个性能度量值，以及所述性能判断准则，确定所述收端目标神经网络是否满足所述输出条件。
即，将收端目标神经网络和发端参考神经网络级联，并使用收端参考神经网络作为参考，对收端目标神经网络进行评估。
结合第一方面，在第一方面的某些可能的实现方式中，所述可实现能力度量值包括以下一项或多项：空口信息的开销、神经网络实现所需计算量、神经网络实现所存储的参数量、以及神经网络实现所需的计算精度。
因此，基于不同应用的实现复杂度，可选择匹配的神经网络来工作，也可选择匹配的性能指标来评估。
第二方面，提供了一种通信方法，该方法包括：第一设备基于与第二设备之间的空口传输条件，确定场景模式，所述场景模式用于确定适用于当前空口传输条件的神经网络，所述神经网络是从预先设计得到的多个神经网络中确定的，所述多个神经网络中的每个神经网络与一种或多种场景模式相匹配；所述第一设备向所述第二设备发送所述场景模式的指示信息。
基于上述方法，第一设备可以基于与第二设备之间当前的空口传输条件，选择合适的场景模式，并将该场景模式指示给第二设备，以便于第一设备和第二设备双方基于同一场景模式来确定采用怎样的神经网络来工作。从而有利于获得性能的提升。
结合第二方面，在第二方面的某些可能的实现方式中，所述方法还包括：所述第一设备接收来自所述第二设备的确认信息，所述确认信息用于指示对所述场景模式的指示信息的成功接收。
第二设备可以基于对上述指示信息的成功接收，向第一设备发送确认信息，以便于第一设备根据对该指示信息的接收情况执行后续流程。比如，在成功接收的情况下，采用与所述场景模式匹配神经网络工作；在未成功接收的情况下，重新发送所述场景模式的指示信息，等等。本申请对此不作限定。
结合第二方面，在第二方面的某些可能的实现方式中，所述方法还包括：所述第一设备采用与所述场景模式匹配的神经网络与所述第二设备通信。
第一设备可以基于第二设备对上述指示信息的成功接收，采用与所述场景模式匹配的神经网络来工作。第二设备也可以基于对上述指示信息的成功接收，确定并采用与所指示 的场景模式匹配的神经网络来工作，从而可以实现两个相互耦合的神经网络协同工作，有利于提高系统性能。
第三方面，提供了一种通信方法，该方法包括：第二设备接收来自第一设备的场景模式的指示信息，所述场景模式用于确定适用于当前空口传输条件的神经网络，所述神经网络是从预先设计得到的多个神经网络中确定的，所述多个神经网络中的每个神经网络与一种或多种场景模式相匹配；所述第二设备根据所述场景模式的指示信息，确定所述场景模式；根据所述场景模式，确定适用于所述空口传输条件的神经网络。
基于上述方法，第一设备可以基于与第二设备之间当前的空口传输条件，选择合适的场景模式，并将该场景模式指示给第二设备，以便于第一设备和第二设备双方基于同一场景模式来确定采用怎样的神经网络来工作。从而有利于获得性能的提升。
结合第三方面，在第三方面的某些可能的实现方式中，所述方法还包括：所述第二设备向所述第一设备发送确认信息，所述确认信息用于指示对所述场景模式的指示信息的成功接收。
第二设备可以基于对上述指示信息的成功接收，向第一设备发送确认信息，以便于第一设备根据对该指示信息的接收情况执行后续流程。比如，在成功接收的情况下，采用与所述场景模式匹配神经网络工作；在未成功接收的情况下，重新发送所述场景模式的指示信息，等等。本申请对此不作限定。
结合第三方面，在第三方面的某些可能的实现方式中，所述方法还包括：所述第二设备采用与所述场景模式相匹配的神经网络与所述第一设备通信。
第一设备可以基于第二设备对上述指示信息的成功接收，采用与所述场景模式匹配的神经网络来工作。第二设备也可以基于对上述指示信息的成功接收，确定并采用与所指示的场景模式匹配的神经网络来工作，从而可以实现两个相互耦合的神经网络协同工作，有利于提高系统性能。
结合第二方面或第三方面，在某些可能的实现方式中，所述多个神经网络中的每个神经网络是基于标准化的参考数据集、标准化的参考神经网络以及标准化的性能判断准则获得的神经网络。
即，第一设备和第二设备所使用的神经网络可以是基于上文第一方面中所述的方法所获取的目标神经网络。
结合第二方面或第三方面，在某些可能的实现方式中，所述第一设备为网络设备，所述第二设备为终端设备；或，所述第一设备为终端设备，所述第二设备为网络设备。
即，上述对空口传输条件的确定以及对场景模式的选择可以是由网络设备执行，也可以是由终端设备执行。本申请对此不作限定。
应理解，上述第二方面和第三方面的方法可以与第一方面的方法结合使用，也可以单独使用。本申请对此不作限定。
还应理解，上述示例中采用的是第一设备和第二设备之间通过场景模式的指示信息来确定与场景模式匹配的神经网络。但这不应对本申请构成任何限定。在另一种可能的实现方法中，空口传输条件的确定以及对场景模式的选择可以不通过信令交互来指示，如网络设备和终端设备可以各自基于预设的空口传输条件、场景模式的选择和神经网络的匹配关系进行神经网络的适配。本申请对此也不做限定。
第四方面，提供了一种获取神经网络的装置，包括用于执行第一方面任一种可能实现方式中的方法的各个模块或单元。
第五方面，提供了一种获取神经网络的装置，包括处理器。该处理器与存储器耦合，可用于执行存储器中的指令或者数据，以实现上述第一方面任一种可能实现方式中的方法。可选地，该装置还包括存储器。可选地，该装置还包括输入/输出接口，处理器与输入/输出接口耦合。
第六方面，提供了一种通信装置，包括用于执行第二方面任一种可能实现方式中的方法的各个模块或单元。
第七方面，提供了一种通信装置，包括处理器。该处理器与存储器耦合，可用于执行存储器中的指令或者数据，以实现上述第二方面任一种可能实现方式中的方法。可选地，该通信装置还包括存储器。可选地，该通信装置还包括通信接口，处理器与通信接口耦合。
在一种实现方式中，该通信装置为终端设备。当该通信装置为终端设备时，所述通信接口可以是收发器，或，输入/输出接口。
在另一种实现方式中，该通信装置为配置于终端设备中的芯片。当该通信装置为配置于终端设备中的芯片时，所述通信接口可以是输入/输出接口。
可选地，所述收发器可以为收发电路。可选地，所述输入/输出接口可以为输入/输出电路。
第八方面，提供了一种通信装置，包括用于执行第三方面任一种可能实现方式中的方法的各个模块或单元。
第九方面，提供了一种通信装置，包括处理器。该处理器与存储器耦合，可用于执行存储器中的指令或者数据，以实现上述第三方面任一种可能实现方式中的方法。可选地，该通信装置还包括存储器。可选地，该通信装置还包括通信接口，处理器与通信接口耦合。
在一种实现方式中，该通信装置为网络设备。当该通信装置为网络设备时，所述通信接口可以是收发器，或，输入/输出接口。
在另一种实现方式中，该通信装置为配置于网络设备中的芯片。当该通信装置为配置于网络设备中的芯片时，所述通信接口可以是输入/输出接口。
可选地，所述收发器可以为收发电路。可选地，所述输入/输出接口可以为输入/输出电路。
第十方面，提供了一种处理器，包括：输入电路、输出电路和处理电路。所述处理电路用于通过所述输入电路接收信号，并通过所述输出电路发射信号，使得所述处理器执行第一方面至第三方面中任一种可能实现方式中的方法。
在具体实现过程中，上述处理器可以为一个或多个芯片，输入电路可以为输入管脚，输出电路可以为输出管脚，处理电路可以为晶体管、门电路、触发器和各种逻辑电路等。输入电路所接收的输入的信号可以是由例如但不限于接收器接收并输入的，输出电路所输出的信号可以是例如但不限于输出给发射器并由发射器发射的，且输入电路和输出电路可以是同一电路，该电路在不同的时刻分别用作输入电路和输出电路。本申请实施例对处理器及各种电路的具体实现方式不做限定。
第十一方面，提供了一种处理装置，包括处理器和存储器。该处理器用于读取存储器中存储的指令，并可通过接收器接收信号，通过发射器发射信号，以执行第一方面至第三 方面任一种可能实现方式中的方法。
可选地，所述处理器为一个或多个，所述存储器为一个或多个。
可选地，所述存储器可以与所述处理器集成在一起，或者所述存储器与处理器分离设置。
在具体实现过程中，存储器可以为非瞬时性(non-transitory)存储器，例如只读存储器(read only memory，ROM)，其可以与处理器集成在同一块芯片上，也可以分别设置在不同的芯片上，本申请实施例对存储器的类型以及存储器与处理器的设置方式不做限定。
应理解，相关的数据交互过程例如发送指示信息可以为从处理器输出指示信息的过程，接收能力信息可以为处理器接收输入能力信息的过程。具体地，处理器输出的数据可以输出给发射器，处理器接收的输入数据可以来自接收器。其中，发射器和接收器可以统称为收发器。
上述第十一方面中的处理装置可以是一个或多个芯片。该处理装置中的处理器可以通过硬件来实现也可以通过软件来实现。当通过硬件实现时，该处理器可以是逻辑电路、集成电路等；当通过软件来实现时，该处理器可以是一个通用处理器，通过读取存储器中存储的软件代码来实现，该存储器可以集成在处理器中，可以位于该处理器之外，独立存在。
第十二方面，提供了一种计算机程序产品，所述计算机程序产品包括：计算机程序(也可以称为代码，或指令)，当所述计算机程序被运行时，使得计算机执行上述第一方面至第三方面中任一种可能实现方式中的方法。
第十三方面，提供了一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序(也可以称为代码，或指令)当其在计算机上运行时，使得计算机执行上述第一方面至第三方面中任一种可能实现方式中的方法。
第十四方面，提供了一种通信系统，包括前述的网络设备和终端设备，所述网络设备和所述终端设备可以分别配置有前述的目标神经网络。
附图说明
图1是可使用本申请实施例提供的神经网络进行通信的通信系统；
图2是配置了神经网络的发射机和接收机对信号的处理过程的示意图；
图3是深度神经网络(deep neural network，DNN)的示意图；
图4是卷积神经网络(convolutional neural network，CNN)的示意图；
图5是递归神经网络(recurrent neural network，RNN)的示意图；
图6至图14是本申请实施例提供的通信系统中的神经网络的训练方法的示意性流程图；
图15和图16是本申请实施例提供的通信方法的示意性流程图；
图17和图18是本申请实施例提供的获取神经网络的装置的示意性框图；
图19和图20是本申请实施例提供的通信装置的示意图框图；
图21是本申请实施例提供的终端设备的结构示意图；
图22是本申请实施例提供的网络设备的结构示意图。
具体实施方式
下面将结合附图，对本申请中的技术方案进行描述。
本申请提供的通信方法可以应用于各种通信系统，例如：长期演进(Long Term Evolution，LTE)系统、LTE频分双工(frequency division duplex，FDD)系统、LTE时分双工(time division duplex，TDD)、通用移动通信系统(universal mobile telecommunication system，UMTS)、全球互联微波接入(worldwide interoperability for microwave access，WiMAX)通信系统、未来的第五代(5th Generation，5G)移动通信系统或新无线接入技术(new radio access technology，NR)。其中，5G移动通信系统可以包括非独立组网(non-standalone，NSA)和/或独立组网(standalone，SA)。
本申请提供的通信方法还可以应用于机器类通信(machine type communication，MTC)、机器间通信长期演进技术(Long Term Evolution-machine，LTE-M)、设备到设备(device to device，D2D)网络、机器到机器(machine to machine，M2M)网络、物联网(internet of things，IoT)网络或者其他网络。其中，IoT网络例如可以包括车联网。其中，车联网系统中的通信方式统称为车到其他设备(vehicle to X，V2X，X可以代表任何事物)，例如，该V2X可以包括：车辆到车辆(vehicle to vehicle，V2V)通信，车辆与基础设施(vehicle to infrastructure，V2I)通信、车辆与行人之间的通信(vehicle to pedestrian，V2P)或车辆与网络(vehicle to network，V2N)通信等。
本申请提供的通信方法还可以应用于未来的通信系统，如第六代移动通信系统等。本申请对此不作限定。
本申请实施例中，网络设备可以是任意一种具有无线收发功能的设备。网络设备包括但不限于：演进型节点B(evolved Node B，eNB)、无线网络控制器(radio network controller，RNC)、节点B(Node B，NB)、基站控制器(base station controller，BSC)、基站收发台(base transceiver station，BTS)、家庭基站(例如，home evolved NodeB，或home Node B，HNB)、基带单元(baseband unit，BBU)，无线保真(wireless fidelity，WiFi)系统中的接入点(access point，AP)、无线中继节点、无线回传节点、传输点(transmission point，TP)或者发送接收点(transmission and reception point，TRP)等，还可以为5G，如，NR，系统中的gNB，或，传输点(TRP或TP)，5G系统中的基站的一个或一组(包括多个天线面板)天线面板，或者，还可以为构成gNB或传输点的网络节点，如基带单元(BBU)，或，分布式单元(distributed unit，DU)等。
在一些部署中，gNB可以包括集中式单元(centralized unit，CU)和DU。gNB还可以包括有源天线单元(active antenna unit，AAU)。CU实现gNB的部分功能，DU实现gNB的部分功能，比如，CU负责处理非实时协议和服务，实现无线资源控制(radio resource control，RRC)，分组数据汇聚层协议(packet data convergence protocol，PDCP)层的功能。DU负责处理物理层协议和实时服务，实现无线链路控制(radio link control，RLC)层、介质接入控制(medium access control，MAC)层和物理(physical，PHY)层的功能。AAU实现部分物理层处理功能、射频处理及有源天线的相关功能。由于RRC层的信息最终会变成PHY层的信息，或者，由PHY层的信息转变而来，因而，在这种架构下，高层信令，如RRC层信令，也可以认为是由DU发送的，或者，由DU+AAU发送的。可以理 解的是，网络设备可以为包括CU节点、DU节点、AAU节点中一项或多项的设备。此外，可以将CU划分为接入网(radio access network，RAN)中的网络设备，也可以将CU划分为核心网(core network，CN)中的网络设备，本申请对此不做限定。
网络设备为小区提供服务，终端设备通过网络设备分配的传输资源(例如，频域资源，或者说，频谱资源)与小区进行通信，该小区可以属于宏基站(例如，宏eNB或宏gNB等)，也可以属于小小区(small cell)对应的基站，这里的小小区可以包括：城市小区(metro cell)、微小区(micro cell)、微微小区(pico cell)、毫微微小区(femto cell)等，这些小小区具有覆盖范围小、发射功率低的特点，适用于提供高速率的数据传输服务。
在本申请实施例中，终端设备也可以称为用户设备(user equipment，UE)、接入终端、用户单元、用户站、移动站、移动台、远方站、远程终端、移动设备、用户终端、终端、无线通信设备、用户代理或用户装置。
终端设备可以是一种向用户提供语音/数据连通性的设备，例如，具有无线连接功能的手持式设备、车载设备等。目前，一些终端的举例可以为：手机(mobile phone)、平板电脑(pad)、带无线收发功能的电脑(如笔记本电脑、掌上电脑等)、移动互联网设备(mobile internet device，MID)、虚拟现实(virtual reality，VR)设备、增强现实(augmented reality，AR)设备、工业控制(industrial control)中的无线终端、无人驾驶(self driving)中的无线终端、远程医疗(remote medical)中的无线终端、智能电网(smart grid)中的无线终端、运输安全(transportation safety)中的无线终端、智慧城市(smart city)中的无线终端、智慧家庭(smart home)中的无线终端、蜂窝电话、无绳电话、会话启动协议(session initiation protocol，SIP)电话、无线本地环路(wireless local loop，WLL)站、个人数字助理(personal digital assistant，PDA)、具有无线通信功能的手持设备、计算设备或连接到无线调制解调器的其它处理设备、车载设备、可穿戴设备，5G网络中的终端设备或者未来演进的公用陆地移动通信网络(public land mobile network，PLMN)中的终端设备等。
其中，可穿戴设备也可以称为穿戴式智能设备，是应用穿戴式技术对日常穿戴进行智能化设计、开发出可以穿戴的设备的总称，如眼镜、手套、手表、服饰及鞋等。可穿戴设备即直接穿在身上，或是整合到用户的衣服或配件的一种便携式设备。可穿戴设备不仅仅是一种硬件设备，更是通过软件支持以及数据交互、云端交互来实现强大的功能。广义穿戴式智能设备包括功能全、尺寸大、可不依赖智能手机实现完整或者部分的功能，例如：智能手表或智能眼镜等，以及只专注于某一类应用功能，需要和其它设备如智能手机配合使用，如各类进行体征监测的智能手环、智能首饰等。
此外，终端设备还可以是物联网(internet of things，IoT)系统中的终端设备。IoT是未来信息技术发展的重要组成部分，其主要技术特点是将物品通过通信技术与网络连接，从而实现人机互连，物物互连的智能化网络。IoT技术可以通过例如窄带(narrow band，NB)技术，做到海量连接，深度覆盖，终端省电。
此外，终端设备还可以包括智能打印机、火车探测器、加油站等传感器，主要功能包括收集数据(部分终端设备)、接收网络设备的控制信息与下行数据，并发送电磁波，向网络设备传输上行数据。
为便于理解本申请实施例，首先结合图1详细说明适用于本申请实施例提供的通信方法的通信系统。如图所示，该通信系统100可以包括至少一个网络设备，如图1中所示的 网络设备101；该通信系统100还可以包括至少一个终端设备，如图1中所示的终端设备102至107。其中，该终端设备102至107可以是移动的或固定的。网络设备101和终端设备102至107中的一个或多个均可以通过无线链路通信。每个网络设备可以为特定的地理区域提供通信覆盖，并且可以与位于该覆盖区域内的终端设备通信。例如，网络设备可以向终端设备发送配置信息，终端设备可以基于该配置信息向网络设备发送上行数据；又例如，网络设备可以向终端设备发送下行数据。因此，图1中的网络设备101和终端设备102至107构成一个通信系统。
可选地，终端设备之间可以直接通信。例如可以利用D2D技术等实现终端设备之间的直接通信。如图中所示，终端设备105与106之间、终端设备105与107之间，可以利用D2D技术直接通信。终端设备106和终端设备107可以单独或同时与终端设备105通信。
终端设备105至107也可以分别与网络设备101通信。例如可以直接与网络设备101通信，如图中的终端设备105和106可以直接与网络设备101通信；也可以间接地与网络设备101通信，如图中的终端设备107经由终端设备105与网络设备101通信。
应理解，图1示例性地示出了一个网络设备和多个终端设备，以及各通信设备之间的通信链路。可选地，该通信系统100可以包括多个网络设备，并且每个网络设备的覆盖范围内可以包括其它数量的终端设备，例如更多或更少的终端设备。本申请对此不做限定。
上述各个通信设备，如图1中的网络设备101和终端设备102至107，可以配置多个天线。该多个天线可以包括至少一个用于发送信号的发射天线和至少一个用于接收信号的接收天线。另外，各通信设备还附加地包括发射机和接收机，本领域普通技术人员可以理解，它们均可包括与信号发送和接收相关的多个部件(例如处理器、调制器、复用器、解调器、解复用器或天线等)。因此，网络设备与终端设备之间可通过多天线技术通信。
可选地，该无线通信系统100还可以包括网络控制器、移动管理实体等其他网络实体，本申请实施例不限于此。
应理解，在图1所示的通信系统100中，网络设备101与终端设备102至107中的任意一个之间可以应用于以下多种可能的场景：调制与解调、编码与解码、预编码与检测接收以及信道信息的提取和压缩与信道信息的重构等。应理解，上文所列举的场景仅为示例，不应对本申请构成任何限定。本申请包含但不限于此。
具体来说，网络设备中的发射机可用于对信号进行调制，终端设备中的接收机可用于对调制后的信号进行解调；网络设备中的发射机可用于对信号进行编码，终端设备中的接收机可用于对编码后的信号进行解码；网络设备中的发射机可用于对信号进行预编码，终端设备中的接收机可用于对预编码后的信号进行检测；终端设备中的发射机可用于对信道信息进行压缩，网络设备中的接收机可用于基于压缩后的信道信息重构信道。
本申请实施例提供的通信方法中，网络设备和终端设备均可以采用神经网络来工作。网络设备和终端设备所采用的神经网络可以是基于本申请实施例提供的获取神经网络的方法所获得的。例如，将通过本申请实施例提供的获取神经网络的方法所获得的发端神经网络应用于网络设备的发射机中，将通过本申请实施例提供的获取神经网络的方法所获得的收端神经网络应用于终端设备的接收机中；和/或，将通过本申请实施例提供的获取神经网络的方法所获得的发端神经网络应用于终端设备的接收机中，通过本申请实施例提供 的获取神经网络的方法所获得的收端神经网络应用于网络设备的发射机中。
由此，网络设备中的发射机与终端设备中的接收机可基于神经网络进行联合优化，终端设备中的发射机和网络设备中的接收机也可基于神经网络进行联合优化。从而可以获得通信系统性能的提升。
图2示出了分别配置了神经网络的发射机和接收机对信号的处理过程。图2以信道信息的提取和压缩与信道信息的重构为例示出了采用了神经网络的发射机和采用了神经网络的接收机对信号的处理过程。
如图2所示，用于进行信道信息的提取和压缩的发端神经网络可以包括前处理模块、信息提取模块和后处理模块。但这不应对本申请构成任何限定。用于进行信道信息的提取和压缩的发端神经网络也可以不包括前处理模块，而包括信息提取模块和后处理模块，或者，还可以不包括前处理模块和后处理模块，而包括信息提取模块。应理解，发端神经网络具体包括的模块可以根据基于不同的设计而定义，本申请对此不作限定。
用于进行信道信息的重构的收端神经网络可以包括逆前处理模块、逆信息提取模块和逆后处理模块。但这不应对本申请构成任何限定。用于进行信道信息的重构的收端神经网络也可以不包括逆前处理模块，而可以包括逆信息提取模块和逆后处理模块，或者，还可以不包括逆前处理模块和逆后处理模块，而包括逆信息提取模块。应理解，收端神经网络具体包括的模块可以基于不同的设计而定义，本申请对此不作限定。
下文列举了发端神经网络和收端神经网络的几种可能的结构。
用于进行信道信息的提取和压缩的发端神经网络的第一种可能的结构是，包括前处理模块、信息提取模块和后处理模块。该前处理模块可用于基于标准化的数据结构对输入该发端目标神经网络的原始数据进行预处理。该信息提取模块可用于从经过预处理的数据进行信息提取，以便对预处理后的数据进行压缩，以得到压缩后的数据。该后处理模块可用于基于标准化的数据结构，对压缩后的数据进行量化和比特映射，以生成空口信息。
用于进行信道信息的提取和压缩的发端神经网络的第二种可能的结构是，包括信息提取模块和后处理模块。换言之，该发端目标神经网络可以不包括前处理模块。该前处理模块可以配置在发射机中，可用于基于标准化的数据结构对输入发射机的原始数据进行预处理。经过预处理后的数据被输入至发端神经网络中。该发端神经网络中的信息提取模块可用于对经过预处理的数据进行信息提取，以便对预处理后的数据进行压缩，以得到压缩后的数据。该后处理模块可用于基于标准化的数据结构，对压缩后的数据进行量化和比特映射，以生成空口信息。
用于进行信道信息的提取和压缩的发端神经网络的第三种可能的结构是，包括信息提取模块。换言之，该发端神经网络可以不包括前处理模块和后处理模块。该前处理模块和后处理模块可以配置在发射机中。前处理模块可用于基于标准化的数据结构对输入发射机的原始数据进行预处理。经过预处理后的数据被输入至发端神经网络中。该发端神经网络中的信息提取模块用于从经过预处理的数据进行信息提取，以便对预处理后的数据进行压缩，以得到的压缩后的数据。该压缩后的数据可输入至后处理模块中，后处理模块可用于基于标准化的数据结构，对压缩后的数据进行量化和比特映射，生成空口信息。用于进行信道信息的重构的收端神经网络的第一种可能的结构是，包括逆后处理模块、逆信息提取模块和逆前处理模块。该逆后处理模块可用于基于标准化的数据结构，对输入至该收端目 标神经网络的空口信息进行解析，以得到压缩后的数据。该逆信息提取模块可用于对压缩后的数据进行处理，以得到压缩前的数据。该逆前处理模块可用于对压缩前的数据进行处理，以得到原始数据。
用于进行信道信息的重构的收端神经网络的第二种可能的结构是，包括逆信息提取模块和逆后处理模块。换言之，该收端神经网络可以不包括逆前处理模块。该逆前处理模块可以配置在接收机中。该收端神经网络中的逆后处理模块可用于基于标准化的数据结构，对输入至该收端目标神经网络的空口信息进行解析，以得到压缩后的数据。该逆信息提取模块可用于对压缩后的数据进行处理，以得到压缩前的数据。该压缩前的数据可以被输入至逆前处理模块中，逆前处理模块可用于对压缩前的数据进行处理，以获取原始数据。
用于进行信道信息的重构的收端目标神经网络的第三种可能的结构是，包括逆信息提取模块。换言之，该收端神经网络可以不包括逆前处理模块和逆后处理模块。该逆前处理模块和逆后处理模块可以配置在接收机中。逆后处理模块可用于基于标准化的数据结构，对输入至该接收机的空口信息进行解析，以得到压缩后的数据。压缩后的数据可以被输入至该收端神经网络中。该收端神经网络中的逆信息提取模块可用于对压缩后的数据进行处理，以得到压缩前的数据。该压缩后的数据可基于标准化的数据结构对接收到的空口信息进行处理得到。该压缩前的数据可以被输入至逆前处理模块，该逆前处理模块可用于对压缩前的数据进行处理，以获取原始数据。
应理解，上文所列举的几种可能的结构仅为示例，不应对本申请构成任何限定。
应理解，上述前处理、信息处理和后处理都需要基于标准化的数据结构进行，与之对应，逆后处理、逆信息提取和逆前处理也都需要基于与发端相一致的标准化的数据结构进行，其目的是保证收端提供空口信息的可靠解析。
其中，逆前处理模块可视为前处理模块的逆处理模块，逆信息提取模块可视为信息提取模块的逆处理模块，逆后处理模块可视为后处理模块的逆处理模块。换言之，发射机中的各模块与接收机中的各模块的行为是互逆的。
还应理解，上文所列举的发端神经网络和收端神经网络所包括的模块仅为便于理解而命名，不应对本申请构成任何限定。例如，前处理模块也可以称为预处理模块，后处理模块也可以称为量化模块，等等。
图2是信道信息的提取与压缩和信道信息的重构的示意图。图2所示的信道信息为信道H，发射机获取到该信道H后，可以先将该信道H输入至前处理模块中，前处理模块对该信道H进行前处理，得到经过前处理后的信道特征信息V。例如，对信道H进行奇异值分解，得到特征向量，或者对信道H进行空频域变换得到角度时延谱。V可理解为经过前处理后的信道。经过前处理后的信道特征信息V被输入至信息提取模块，信息提取模块可以基于预先设计好的算法对该信道特征信息V进行信息提取，得到压缩后的信道V'。例如，基于函数f en(V，Θ en)得到压缩后的信道特征信息V'。压缩后的信道特征信息V'被输入至后处理模块，后处理模块可以基于预先定义的数据结构对该压缩后的信道特征信息V'进行量化和比特映射，以生成空口信息S。
接收机接收到该空口信息S后，可以将该空口信息S输入至该接收机中的逆后处理模块。逆后处理模块可基于预先定义的数据结构对空口信息S进行处理，以恢复出压缩后的信道特征信息
Figure PCTCN2021102079-appb-000001
压缩后的信道特征信息
Figure PCTCN2021102079-appb-000002
被输入至逆信息提取模块，逆信息提取模块 可以基于与发射机中信息提取模块相应的算法对该信道特征信息
Figure PCTCN2021102079-appb-000003
进行处理，以恢复出压缩前的信道特征信息
Figure PCTCN2021102079-appb-000004
例如，基于函数f de(f en(V,Θ en),Θ de)恢复出信道特征信息
Figure PCTCN2021102079-appb-000005
信道特征信息
Figure PCTCN2021102079-appb-000006
被输入至逆前处理模块，逆前处理模块对该信道特征信息
Figure PCTCN2021102079-appb-000007
进行处理后，便可恢复出信道
Figure PCTCN2021102079-appb-000008
应理解，由此恢复出的信道
Figure PCTCN2021102079-appb-000009
为上文输入至发射机中的信道H的估计值。如图所示，上述发端神经网络和收端神经网络对输入的数据的处理可以通过如下函数来表示：
Figure PCTCN2021102079-appb-000010
其中Θ＝{Θ en,Θ de}。
需要说明的是，上述过程中，前处理模块和逆前处理模块需要基于一致的规则来工作，信息提取模块和逆信息提取模块需要基于一致的规则来工作，后处理模块和逆后处理模块需要基于一致的规则来工作，保证发射端的神经网络和接收端的神经网络能够协同工作。这里的一致是指，二者可以基于一致的规则来设计，但并不代表双方使用相同的算法、相同的数据结构等来处理数据。基于一致的规则来工作，旨在强调双方在用于对数据进行处理时，所使用的算法、数据结构等是相对应的。
比如，发射机中的后处理模块基于某一数据结构，对压缩后的信道V'进行量化和比特映射，生成空口信息S；接收机中的逆后处理模块需要基于对应的某一数据结构根据该空口信息S最大程度恢复出信道特征信息
Figure PCTCN2021102079-appb-000011
又如，发射机中的信息提取模块基于某一算法对信道特征信息V进行压缩，得到压缩后的信道特征信息V'；接收机中的逆信息提取模块需要基于对应的某一算法对压缩后的信道特征信息V'进行处理，最大程度地恢复出压缩前的信道特征信息
Figure PCTCN2021102079-appb-000012
应理解，图2所示例的信道H作为信道信息的一例，仅为便于理解而示例，不应对本申请构成任何限定。上文所列举的信道H、信道特征信息V都可以作为信道信息的一例。由于对神经网络所包含的模块的不同设计，输入至发端神经网络的信道信息可能是信道H，也可能是信道特征信息V；该发端神经网络的输出可能是压缩后的信道特征信息V'，也可能是空口信息S。收端神经网络的输入可能是空口信息S，也可能是基于空口信息S恢复出的压缩后的信道特征信息
Figure PCTCN2021102079-appb-000013
从收端神经网络输出的信道信息可能是恢复出的信道特征信息
Figure PCTCN2021102079-appb-000014
也可能是恢复出的信道
Figure PCTCN2021102079-appb-000015
并且，由于收端神经网络和发端神经网络的结构并不一定完全对称，故对信道信息的定义也可能不同。比如，输入发端神经网络的信道信息可能是信道H，而由收端神经网络输出的信道信息可能是恢复出的信道特征信息
Figure PCTCN2021102079-appb-000016
基于信道信息的不同，上文结合图2描述的函数
Figure PCTCN2021102079-appb-000017
可以作出相应的变形。例如，可以表示为，Y＝f(X,Θ)＝f de(f en(X,Θ en),Θ de)。其中，X表示输入发端神经网络的信道信息，Y表示由收端神经网络输出的信道信息。如，X可为H或V的任意一个，Y可为
Figure PCTCN2021102079-appb-000018
或
Figure PCTCN2021102079-appb-000019
的任意一个。
还应理解，上文结合图2所描述的发端神经网络是用于进行信道信息的提取和压缩的神经网络，其中所包含的模块与信道信息的提取和压缩过程相关。上文结合图2所描述的收端神经网络是用于进行信道信息的重构的神经网络，其中所包含的模块也与信道信息的重构过程相关。在用于信号的其他处理阶段，如编码与解码、星座调制与解调、预编码与检测接收等处理阶段时，发端神经网络和收端神经网络所包含的模块可能会有所不同。为了简洁，这里不一一举例说明。
在实际商用过程中，发射机和接收机可能是由不同的厂商提供的。若要达到上述要求，可能需要发射机或接收机从对端下载匹配当前场景的神经网络结构和参数信息，或者，发 射机和接收机从某一网络虚拟处理节点下载匹配当前场景的神经网络结构和参数信息。比如，下载生成和解析空口信息的规则；又比如，下载发射机对信道H进行信息提取和压缩的神经网络的参数和结构描述等。这可能会带来较大的空口信令开销。
有鉴于此，本申请提供一种通信系统中的神经网络获取方法，以期减小发射机和接收机下载神经网络结构和参数信息所带来的巨大的空口开销。
为了更好地理解本申请实施例，下面对本文中涉及到的术语做简单说明。
1、神经网络(neural network，NN)：作为人工智能的重要分支，是一种模仿动物神经网络行为特征进行信息处理的网络结构。神经网络的结构由大量的节点(或称神经元)相互联接构成，基于特定运算模型通过对输入信息进行学习和训练达到处理信息的目的。一个神经网络包括输入层、隐藏层及输出层，输入层负责接收输入信号，输出层负责输出神经网络的计算结果，隐藏层负责特征表达等复杂的功能，隐藏层的功能由权重矩阵和对应的激活函数来表征。
深度神经网络(deep neural network，DNN)一般为多层结构。增加神经网络的深度和宽度，可以提高他的表达能力，为复杂系统提供更强大的信息提取和抽象建模能力。
DNN可以有多种构建方式，例如可以包括但不限于，递归神经网络(recurrent neural network，RNN)、卷积神经网络(convolutional neural network，CNN)以及全连接神经网络等。
图3前馈神经网络(feedforward neural network，FNN)的示意图。如图所示，不同方框表示不同的层。比如，从左往右看第一层为输入层(input layer)，第二层为隐含层(hidden layer)，第三层为输出层(output layer)。DNN一般还包含多个隐含层，其作用为不同程度地提取特征信息。输出层可用于从提取的特征信息中映射出所需要的输出信息。图中每个黑色圆圈可以表示一个神经元，不同方框内的神经元表示不同层的神经元。每层的神经元连接方式和采用的激活函数将决定神经网络的表达函数。
2、卷积神经网络(CNN)：一种多层的神经网络，每层有多个二维平面组成，而每个平面由多个独立神经元组成，每个平面的多个神经元共享权重，通过权重共享可以降低神经网络中的参数数目。目前，在卷积神经网络中，处理器进行卷积操作通常是将输入信号特征与权重的卷积，转换为信号矩阵与权重矩阵之间的矩阵乘运算。在具体矩阵乘运算时，对信号矩阵和权重矩阵进行分块处理，得到多个分形(fractional)信号矩阵和分形权重矩阵，然后对多个分形信号矩阵和分形权重矩阵进行矩阵乘和累加运算。
图4是CNN中隐含层的示意图。图中黑色方框为每层的通道，或者称特征地图。不同通道可以对应相同或不同的卷积核。对于CNN来说，并不是所有上下层神经元都能直接相连，而是通过所述“卷积核”作为中介。在该网络中，待处理信息通过卷积操作后仍然可以保留原先的位置关系。如图中所示，左侧第m-1层的信息经过卷积操作后，得到右侧第m层。
应理解，图4仅为便于理解而示例，关于CNN的更详细的说明可以参看现有技术。为了简洁，这里不做详述。
3、递归神经网络(RNN)：具有树状阶层结构且网络节点按其顺序对输入信息进行递归的人工神经网络。在RNN中，神经元的输出可以在下一时刻直接作用到自身，即第i层神经元在m时刻的输入，除了(i-1)层神经元在该时刻的输出外，还包括其自身在(m-1) 时刻的输出。
图5是RNN的示意图。图中每个黑色圆圈表示一个神经元。可以看到，隐含层2的神经元的输入除了隐含层1的神经元在该时刻的输出(如图中直线结构所示)外，还包括其自身在上一时刻的输出(如图中半圆形箭头所示)。
应理解，图5仅为便于理解而示例。关于RNN的更详细的说明可以参看现有技术。为了简洁，这里不做详述。
4、性能判断准则：在本申请实施例中，为了获取具有较优性能的目标神经网络，可以通过性能判断准则来对预先构建好的目标神经网络进行评估。
在一种可能的设计中，该性能判断准则可以包括性能指标和输出条件。对目标神经网络的评估例如可以通过标准化的性能度量函数来获得性能度量值，并将性能度量值输入至性能判断准则中，与一个或多个性能指标进行比较，以判断其是否满足输出条件。
作为示例而非限定，性能度量函数可以包括但不限于以下一项或多项：均方误差(mean square error，MSE)、归一化最小均方误差(normalized mean square error，NMSE)、平均绝对误差(mean absolute error，MAE)、最大绝对误差(也称，绝对误差界)、相关系数、交叉熵、互信息、误比特率、或误码率等。
其中，MSE、NMSE、MAE、最大绝对误差、误比特率、误码率等可用于表征差异程度，相关性系数、交叉熵、互信息等可用于表征相似程度。但本领域的技术人员可以理解，在两个值差异程度较大的情况下，二者的相似程度也就越小；在两个值差异程度较小的情况下，二者的相似程度也就越大。因此，相似程度和差异程度之间是相对应的。
性能指标可以包括一个或多个性能指标的上界和/或一个或多个性能指标的下界。性能指标例如可以通过上文性能度量函数中所列举的多项中的一项或多项来表征。对于不同的性能度量函数，可以定义不同的性能指标。
比如，对应于用于表征相似程度的性能度量函数来说，相似程度越高，性能越好。以相关系数为例，可以定义性能指标的上界为0.99999，性能指标的下界为0.9。
又比如，对应于用于表征差异程度的性能度量函数来说，差异程度越小，性能越好。以MSE为例，可以定义性能指标的上界为0.001，性能指标的下界为0.1。
应理解，性能指标是在用于判断目标神经网络是否满足输出条件而设定的指标，换言之，达到性能指标的目标神经网络在应用于发射机或接收机中后，能够到达的性能的可接受范围。例如，以系统吞吐为例，如果相似系数小于0.9，则可认为性能太差；若相关系数大于或等于0.9，则可认为性能可接受；若相关系数大于0.99999，则可认为性能没有更大的增益，或者增益很小。因此相关系数的可接受范围在[0.9，0.99999]之间。因此，对于相关系数来说，达到性能指标的下界可以是指大于或等于0.9，未超出性能指标的上界可以是指小于或等于0.99999。
又例如，以误码率为例，如果MSE大于0.1，则可认为性能太差；如果MSE小于或等于0.1，则可认为性能可接受；如果MSE小于0.001，则可认为性能没有更大的增益，因此MSE的可接受范围可以是在[0.001，0.1]之间。因此，对于MSE来说，达到性能指标的下界可以是指小于或等于0.1，未超出性能指标的上界可以是指大于或等于0.001。
由上文示例可以看到，在本申请实施例中，若性能度量函数用于表征相似程度，则由性能度量函数输出的性能度量值越高，表示相似程度越高，目标神经网络的性能也就越好， 因此性能指标的下界小于上界；若性能度量函数用于表征差异程度，则由性能度量函数输出的性能度量值越低，表示差异程度越小，表示目标神经网络的性能也就越好，因此，性能指标的上界小于性能指标的下界。换言之，性能指标的上界和下界是基于性能的好与差而定义的，因此，对于不同的性能指标，其上界或下界与具体取值的大小关系并不是一成不变的。
此外，本申请实施例中，当涉及到“达到性能指标的下界”的相关描述时，可以是指相似程度大于或等于某一阈值(为便于区分和说明，例如记作第一阈值)；也可以是指差异程度小于或等于某一阈值(为便于区分和说明，例如记作第二阈值)。当涉及到“未超出性能指标的上界”的相关描述时，可以是指相似程度小于或等于某一阈值(为便于区分和说明，例如记作第三阈值)；也可以是指差异程度大于或等于某一阈值(为便于区分和说明，例如记作第四阈值)。其中，对于相似程度而言，所对应的第一阈值小于第三阈值；对于差异程度而言，所对应的第二阈值大于第四阈值。
性能判断准则在用于判断某一目标神经网络是否满足输出条件时，可以根据一个或多个性能指标来判断。该一个或多个性能指标例如可以包括针对一个性能度量函数给定的一个或多个性能指标，如包括性能指标的上界和/或下界；该一个或多个性能指标也可以包括针对多个不同的性能度量函数中的每个性能度量函数分别给定的一个或多个性能指标。本申请对此不作限定。
一示例，基于相关系数来确定目标神经网络是否满足输出条件。协议可以定义相关系数的上界和下界，并可定义输出条件为：性能度量值均达到性能指标的下界，且未超出性能指标的上界。如前文示例，下界为0.9，上界为0.99999。由目标神经网络与参考神经网络输出的结果被输入至性能度量函数，可以获得一个或多个性能度量值。将该一个或多个性能度量值分别与上界和下界比较，看是否都落入[0.9，0.99999]的范围内。若是，则满足输出条件；若不是，则不满足输出条件。
又一示例，仍然基于相关系数来确定目标神经网络是否满足输出条件。协议可以定义相关系数的上界和下界，并可定义输出条件为：性能度量值均达到性能指标的下界，且90％的性能度量值未超出性能指标的上界。如前文示例，下界为0.9，上界为0.99999。则可将由性能度量函数输出的一个或多个性能度量值分别与上界和下界比较，看是否都达到下界0.9，且是否有90％以上的性能度量值未超出上界0.99999。若是，则满足输出条件；若不是，则不满足输出条件。
另一示例，基于相关系数和误码率来确定目标神经网络是否满足输出条件。协议可以定义性能指标的上界包括相关系数的上界和下界，以及误码率的下界；并可定义输出条件为：性能度量值分别达到相关系数的下界和误码率的下界，且未超过相关系数的上界。如前文示例，相关系数的下界为0.9，上界为0.99999，误码率的下界为0.1。则可将由性能度量函数输出的一个或多个表征相关系数的性能度量值分别与相关系数的上界和下界比较，看是否都达到下界0.9，且都未超出上界0.99999；并将由性能度量函数输出的一个或多个表征误码率的性能度量值是否都达到下界。
此外，性能判断准则还可包括针对不同的业务优先级而设置的性能指标的上界和/或下级，以及输出条件。
例如，对于语音通话、视频通话等业务，所对应的性能指标可以较高，或输出条件较 严苛；对于短信息等业务，所对应的性能指标可以较低，或输出条件较宽松。
比如，相关系数的下界为0.9相对于相关系数的下界为0.8来说，前者的性能指标较高。又如，MSE的下界为0.1相对于MSE的下界为0.01来说，后者的性能指标较高。再如，输出条件为全部的性能度量值均达到性能指标的下界相对于部分性能度量值达到性能指标的下界而言，前者较为严苛。
在对目标神经网络进行评估时，可以根据所应用的业务的优先级来采用相应的性能指标和输出条件对其进行评估。
应理解，上文结合多个性能指标和输出条件的示例，对性能判断准则做了详细说明。但应理解，这些示例仅为便于理解而示出，不应对本申请构成任何限定。
5、对偶网络(或称对偶架构)：用于描述通信网络中收发两端(如基站侧和终端侧)联合优化的神经网络结构。对偶网络具体可以包括发端神经网络和收端神经网络。对偶网络包括但不限于基于自编码器(Autoencoder)结构或者自编码器结构的各种变化形式，或者以其他神经网络结构组合联合构建的收、发端神经网络。在自编码器结构下，发端神经网络为编码器神经网络(encoder neural network)，收端神经网络为译码器神经网络(decoder neural network)。两者间相互约束，协同工作。
需注意，这里所述的编码器和译码器并不同于上文所述的用于信道编、解码的编码器和译码器。
在本申请实施例中，发端神经网络和收端神经网络可以是一对一、一对多、多对一的组合，以适配多种系统性能和实现复杂度的要求。相互耦合的一对神经网络，即，可协同工作的一个发端神经网络和一个收端神经网络。
下面将结合附图详细说明本申请实施例提供的获取神经网络的方法。应理解，该方法可以包括获取应用于发射机中的发端神经网络的过程和获取应用于接收机中的收端神经网络的过程。下文中结合不同的实施例来描述获取发端神经网络和收端神经网络的流程。其中，为便于区分和描述，下文中将要获取的神经网络称为目标神经网络。将应用于发射机的目标神经网络称为发端目标神经网络，将应用于接收机的目标神经网络称为收端目标神经网络。
还应理解，对发端目标神经网络和收端目标神经网络的获取例如可以是线下(offline)完成的。发端目标神经网络和收端目标神经网络例如可以是从同一个设备获取得到，也可以是从不同设备获取得到，本申请对此不作限定。用于获取目标神经网络的设备例如可以是通信设备，也可以是不同于通信设备的计算设备。本申请对此不作限定。下文中为方便说明，将用于获取目标神经网络的设备称为获取设备。
图6是本申请实施例提供的获取神经网络的方法600的示意性流程图。如图6所示，该方法600包括：步骤610至步骤630。下面详细说明方法600中的各个步骤。
在步骤610中，基于标准化的参考数据集、标准化的一个或多个参考神经网络和标准化的性能判断准则，确定构建的目标神经网络是否满足预设的输出条件。
该目标神经网络可以为某个应用的专用神经网络。该目标神经网络可应用于通信系统中对信号的各个处理阶段。例如编码与解码、星座调制与解调、预编码与检测接收、信道信息的提取和压缩与信道信息的重构。具体而言，发端神经网络可用于编码，收端神经网络可用于解码；发端神经网络可用于星座调制，收端神经网络可用于解调；发端神经网络 可用于预编码，收端神经网络可用于检测接收；发端神经网络可用于进行信道信息的提取和压缩，收端可用于进行信道信息的重构。
应理解，基于本申请实施例提供的方法所获取的目标神经网络还可用作其他功能，本申请对此不作限定。
该目标神经网络可以构建在已有的神经网络结构上，例如，在CNN、RNN等基础网络结构上拓展设计；也可以采用新的网络结构设计，基于标准化的参考数据集和/或该应用相关的其他数据集对其进行训练，可以获得目标神经网络。
其中，关于神经网络结构的相关描述在上文已经结合图3至图5做了详细说明，为了简洁，这里不再重复。
为了使得神经网络的输出尽可能地接近真正想要预测的值，获取设备可以对目标神经网络进行训练。示例性地，可以通过比较当前网络的预测值和真正想要的目标值，再根据两者之间的差异情况来更新每一层神经网络的权重向量。比如，网络的预测值高了，就可以调整权重向量让它预测低一些，通过不断的调整，直到神经网络能够预测出真正想要的目标值。因此，就需要预先定义“如何比较预测值和目标值之间的差异”，也即损失函数(loss function)、代价函数(cost function)或目标函数(objective function)，它们是用于衡量预测值和目标值的差异的重要方程。其中，以损失函数举例，损失函数的输出值(loss)越高表示差异越大，那么神经网络的训练就变成了尽可能缩小这个输出值的过程。
上述对目标神经网络的训练可以基于标准化的参考数据集和/或该应用相关的其他数据集来实现。
示例性地，标准化的参考数据集具体可以包括：用于训练目标神经网络的参考数据集、用于对目标神经网络进行性能评估的参考数据集等。标准化可以是指协议定义。各厂商可以基于上述方法构建目标神经网络，并基于标准化的参考数据集来评估该目标神经网络。
由于上述目标神经网络是应用在发射机或接收机中的神经网络。为了在通信系统中获得较优的性能，可以基于标准化的参考数据集、标准化的参考神经网络和标准化的性能判断准则，对该目标神经网络进行评估，以确定其是否满足输出条件。
参考数据集、参考神经网络和性能判断准则都可以是标准化的，或者说，协议预定义的。对应于不同的应用场景，协议可以预定义多套不同的标准化的配置。在基于不同的应用场景对目标神经网络进行评估时，可以采用与所应用的场景匹配的一套标准化的配置进行评估。具体来说，在不同的应用场景下，可以采用相匹配的参考数据集、参考神经网络、性能度量函数以及性能判断准则来对目标神经网络进行评估，以获取应用于不同场景的目标神经网络。
这里所说的应用场景具体可以是指基于不同的空口传输条件而区分的场景。例如，上述应用场景可以包括但不限于，室外密集城区、室外普通城区、室外乡村、室外山区、室内办公、室内厂房等的不同信道场景，单干扰源、多干扰源、强弱不同的多种干扰信号类型，中近点高信噪比、远点低信噪比的不同信噪比条件，步行低速、城区车载中速、高铁超高速等不同移动条件等。由于不同的应用场景下空口传输条件不同，且不同的应用场景对传输可靠性的要求也不同。与应用场景相匹配的标准化的配置中可以包括以下一项或多项：标准化的参考数据集、标准化的参考神经网络、标准化的数据结构、标准化的性能度量函数以及标准化的性能判断准则等。
其中，参考神经网络可以包括收端参考神经网络和/或发端参考神经网络。且，当参考神经网络包括收端参考神经网络和发端参考神经网络时，该收端参考神经网络和发端参考神经网络是一对一耦合的。
数据结构可用于生成和/或解析空口信息。更具体地说，数据结构可用于发端神经网络生成空口信息，也可用于收端神经网络解析空口信息。该数据结构也可理解为是空口信息的生成规则，或对空口信息的解读规则。
应理解，数据结构是一个上位的概括，可用于在对数据的不同处理阶段指导对数据的处理。
以上述信道信息的压缩和提取和信道信息的重构为例。发端可以基于某一数据结构(例如记为数据结构1)对输入的原始数据进行前处理，例如对输入的原始数据进行重排；并可基于另一数据结构(例如记为数据结构2)对重排后的数据进行信息提取，例如确定采用怎样的张量表达；还可基于又一数据结构(例如记为数据结构3)生成空口信息，例如确定如何进行量化和比特映射。收端可以基于与数据结构1相应的数据结构对空口信息进行解析，并可基于用于与数据结构2相应的数据结构对解析得到的数据进行处理，还可基于与数据结构3相应的数据结构，获取原始数据。
本申请对于数据结构的具体内容不作限定。
标准化的可实现能力度量值可理解为神经网络用于实现某一应用时所需要的能力参数，可用于选择匹配的神经网络和/或匹配的性能指标。这里所述的匹配的神经网络，具体可以是指，该神经网络在给定的约束下，能够应用在发射机或接收机中，所述给定的约束可以包括实现复杂度和/或空口开销。匹配的性能指标，具体可以是指，用于获取该神经网络的阶段对该神经网络进行评估所使用的性能指标。
可实现能力度量值例如可以包括但不限于以下一项或多项：空口信息的开销、神经网络实现所需的计算量、神经网络实现所存储的参数量、以及神经网络实现所需的计算精度。
示例性地，空口信息的开销例如可以指比特开销；神经网络实现所需计算量例如可以指每秒浮点计算次数(floating-point operations per second，Flops)；神经网络实现所需的计算精度要求例如可以是2比特定点、或4比特定点、或8比特定点、或16比特定点、或16比特浮点、或32比特浮点、或64比特浮点等。本申请对此不作限定。
各神经网络可通过与上文所列举相应的参数来表征各自的能力。由于不同应用的实现复杂度可能不同，对神经网络能力的要求也不同。获取设备可以基于当前应用对实现复杂度的支持能力，选择匹配的神经网络。
需要说明的是，基于不同神经网络的可实现能力度量值，不同神经网络所能够达到的性能的上限和/或下限可能会有所不同。因此，同一性能指标可能具有一个或多个性能指标的阈值(如包括上界和/或下界)，与一种或多种不同的能力相对应。在后续对神经网络进行评估时，可以根据所选择的可实现能力度量值，选择相应的性能指标进行评估。
应理解，与不同的应用场景匹配的标准化的配置可以是互不相同的，也可以是部分相同的。比如，与不同的应用场景匹配的标准化的配置中，可以共用同一个参考数据集，或者，共用同一个数据结构，等等。本申请对此不作限定。
下面将结合具体的实现方式来详细说明对目标神经网络的评估过程。
在一种实现方式中，各厂商可以基于标准化的参考数据集，将构建的目标神经网络与 标准化的一个或多个参考神经网络级联，并对级联后的输出结果进行评估，根据标准化的性能判断准则，确定该目标神经网络是否满足输出条件。在满足输出条件的情况下，则可以执行步骤620，将该目标神经网络输出；在不满足输出条件的情况下，可以执行步骤630，对该目标神经网络进行优化，直至获得的目标神经网络满足输出条件，将其输出为止。
在另一种实现方式中，各厂商可以将取自标准化的参考数据集的同一个或多个数据输入至构建的目标神经网络以及标准化的一个或多个参考神经网络中，将目标神经网络输出的结果与参考神经网络输出的结果进行对比，根据标准化的性能判断准则，确定该目标神经网络是否满足输出条件。在满足输出条件的情况下，则可以执行步骤620，将该目标神经网络输出；在不满足输出条件的情况下，可以执行步骤630，对该目标神经网络进行优化，直至获得的目标神经网络满足输出条件，将其输出为止。
下面将结合图7至图14的具体示例来对上述两种实现方式做更详细的说明。换言之，下文示出的多个示例是对步骤610的更详细的说明。为方便理解和说明，后文中在未作出特别说明的情况下，发端神经网络可以包括前处理模块、信息提取模块和后处理模块，收端神经网络可以包括逆后处理模块、逆信息处理模块和逆前处理模块。
图7和图8是本申请实施例提供的获取神经网络的方法的示意图。图7和图8所示的获取神经网络的方法为将目标神经网络和一个标准化的参考神经网络级联，并对级联后的输出结果进行评估的示例。其中，图7中所示的目标神经网络可应用于发射机，该目标神经网络为发端参考神经网络；参考神经网络可应用于接收机，该参考神经网络为收端参考神经网络。图8中所示的目标神经网络可应用于接收机，该目标神经网络为收端目标神经网络，参考神经网络可应用于发射机，该参考神经网络为发端参考神经网络。
如图7所示，获取设备可以从参考数据集中获取一个或多个发端数据。比如，将其中的一个发端数据记为X，X被输入至发端目标神经网络。发端目标神经网络对该发端数据X进行前处理和信息提取后，可得到压缩后的数据X'，发端目标神经网络可进一步根据标准化的数据结构对压缩后的数据X'进行处理，以生成空口信息S。此后，空口信息S可以被输入至收端参考神经网络中。收端神经网络根据标准化的数据结构对该空口信息S进行解析后，可恢复出压缩后的数据
Figure PCTCN2021102079-appb-000020
收端参考神经网络，可以进一步基于该压缩后的数据
Figure PCTCN2021102079-appb-000021
得到输出结果Y。
如前所述，发端目标神经网络可以应用于对信号的不同处理阶段中。以信道信息的提取和压缩与信道信息的重构为例。发端数据X例如可以是信道H。发端神经网络可以基于图2中所列举的发端神经网络的操作流程，对信道H先后进行前处理、信息提取和后处理，生成空口信息S。
收端参考神经网络可以基于接收到的空口信息S进行信道H的重构。收端参考神经网络可以基于图2中所列举的收端神经网络的操作流程，对空口信息S先后进行逆后处理、逆信息提取和逆前处理，以恢复出信道
Figure PCTCN2021102079-appb-000022
收端参考神经网络所输出的
Figure PCTCN2021102079-appb-000023
是信道H的估计值
Figure PCTCN2021102079-appb-000024
应理解，该估计值也即输出结果。
此后，获取设备可以将输出结果Y与输入的发端数据X输入至性能度量函数中，以获得性能度量值。应理解，获取设备可以将该输出结果Y与发端数据X输入至一个或多个性能度量函数中，获得一个或多个性能度量值。
还应理解，获取设备还可从参考数据集中获得的一个或多个发端数据按照上述流程来 处理，可以获得多对输出结果与发端数据。每对输出结果和发端数据可以被输入至一个或多个性能度量函数中，获得一个或多个性能度量值。多对输出结果与发端数据可获得多个性能度量值。
该获取设备可以进一步根据标准化的性能判断准则，将输出的性能度量值与标准化的性能指标进行比较，再根据比较结果确定该发端目标神经网络是否满足预设的输出条件。
一示例，该输出条件可以是，性能度量值均达到性能指标的下界。该性能度量函数例如是MSE，该性能指标的下界例如可以是MSE的下界。将一对或多对输出结果与发端数据输入至性能度量函数可得到一个或多个MSE的值。例如输出结果
Figure PCTCN2021102079-appb-000025
与发端数据H输入至性能度量函数可得到二者的MSE，即
Figure PCTCN2021102079-appb-000026
获取合并可以进一步判断获得的一个或多个性能度量值是否达到标准化的MSE的下界。若都达到，则满足输出条件；若部分或全部未达到，则不满足输出条件。
应理解，若该收端参考神经网络是根据可实现能力度量值选择出来的可匹配的神经网络，则上述性能指标可基于该可实现能力度量值对应的性能指标来评估。
还应理解，如前所述，基于对神经网络所包含的模块的不同设计，发端神经网络的输入和输出以及收端神经网络的输入和输出并不一定和上文示例相同。比如，发端神经网络的输入可能是信道H，收端神经网络的输出却可能是特征向量(应理解，特征向量可以理解为上文所述信道特征信息V的一例)。这里，收端输出的特征向量是恢复得到的特征向量，例如记作
Figure PCTCN2021102079-appb-000027
但这并不影响性能判断准则的使用。由于特征向量与信道之间的关系，U＝eig(H)(eig()表示用于求解特征向量的函数)，可以由特征向量
Figure PCTCN2021102079-appb-000028
进一步恢复出信道
Figure PCTCN2021102079-appb-000029
或者由信道H可以进一步确定特征向量U，进而确定二者的MSE。即，
Figure PCTCN2021102079-appb-000030
也可能表现为
Figure PCTCN2021102079-appb-000031
或者说，
Figure PCTCN2021102079-appb-000032
与
Figure PCTCN2021102079-appb-000033
之间是可替换的。下文中多处涉及到性能度量函数的描述，为了简洁，后文中省略对相同或相似情况的说明。
另一示例，该输出条件可以是，90％的性能度量值达到性能指标的下界。该性能度量函数例如是相关系数的计算函数，该性能指标的下界例如是相关系数的下界。将一对或多对输出结果与发端数据输入至性能度量函数可得到一个或多个相关系数。例如输出结果
Figure PCTCN2021102079-appb-000034
与发端数据H的相关系数可表示为
Figure PCTCN2021102079-appb-000035
或
Figure PCTCN2021102079-appb-000036
Corr()表示求相关系数的函数。获取设备可以进一步判断获得的一个或多个性能度量值是否达到标准化的相关系数的下界。若90％以上的性能度量值都达到相关系数的下界，则满足输出条件；若性能度量值达到相关系数的下界的个数不满90％，则不满足输出条件。
应理解，MSE和相关系数仅为示例，本申请对于获取设备确定目标神经网络是否满足输出条件所使用的性能指标及其相关的性能度量函数均不作限定。还应理解，达到性能指标的下界可以作为目标神经网络的输出条件的一例，但不应对本申请构成任何限定。
由于上文已经结合多个示例对性能判断准则做了详细说明，这里为了简洁，不再列举更多的输出条件，并针对不同的输出条件一一举例详述。
如图8所示获取设备可以从参考数据集中获取一个或多个发端数据。比如，将其中的一个发端数据记为X，X被输入至发端参考神经网络。发端参考神经网络对该发端数据X进行前处理和信息提取后，可得到压缩后的数据X'，发端参考神经网络可进一步根据标准化的数据结构对压缩后的数据X'进行处理，以生成空口信息S。此后，空口信息S可以 被输入至收端目标神经网络中。收端目标神经网络可以基于标准化的数据结构对该空口信息S进行解析，恢复出压缩后的数据
Figure PCTCN2021102079-appb-000037
收端目标神经网络可以进一步基于该压缩后的数据
Figure PCTCN2021102079-appb-000038
得到输出结果Y。
仍以信道信息的提取和压缩与信道信息的重构为例。发端数据X可以是信道H。发端参考神经网络可以基于图2中所列举的发端神经网络的操作流程，对信道H先后进行前处理、信息提取和后处理，生成空口信息S。
收端目标神经网络可以基于接收到的空口信息S进行信道H的重构。收端目标神经网络可以基于图2中所列举的收端神经网络的操作流程，对空口信息线先后进行逆后处理、逆信息提取和逆前处理，以恢复出信道
Figure PCTCN2021102079-appb-000039
收端目标神经网络所输出的
Figure PCTCN2021102079-appb-000040
是信道H的估计值
Figure PCTCN2021102079-appb-000041
应理解，该估计值也即输出值。
此后，获取设备可以将输出结果Y与输入的发端数据X输入至性能度量函数中，以获得性能度量值。应理解，获取设备可以将该输出结果Y与发端数据X输入至一个或多个性能度量函数中，获得一个或多个性能度量值。
还应理解，获取设备还可从参考数据集中获得的一个或多个发端数据按照上述流程来处理，可以获得多对输出结果与发端数据。每对输出结果和发端数据可以被输入至一个或多个性能度量函数中，获得一个或多个性能度量值。多对输出结果与发端数据可获得多个性能度量值。
该获取设备可以进一步根据标准化的性能判断准则，将输出的性能度量值与标准化的性能指标进行比较，再根据比较结果确定该发端目标神经网络是否满足预设的输出条件。
由于上文已经结合多个示例对性能判断准则做了详细说明，这里为了简洁，不再列举更多的输出条件，并针对不同的输出条件一一举例详述。
图9和图10是本申请实施例提供的神经网络的获取方法的示意图。图9和图10所示的神经网络的训练方法将一个标准化的参考神经网络作为参考，对训练得到的目标神经网络进行评估的示例。其中，图9中所示的目标神经网络和参考神经网络可应用于发射机，该目标神经网络是发端目标神经网络，该参考神经网络是发端参考神经网络。图10所示的目标神经网络和参考神经网络可应用于接收机，该目标神经网络是收端目标神经网络，该参考神经网络是收端参考神经网络。
如图9所示，获取设备可以从参考数据集中获取一个或多个发端数据。每个发端数据可以被输入至发端目标神经网络和发端参考神经网络中，以获得从发端目标神经网络输出的目标空口信息和从发端参考神经网络生成的参考空口信息。
以其中的一个发端数据X为例。发端数据X被输入至发端目标神经网络和发端参考神经网络中。发端参考神经网络可以基于预先设计好的算法对该发端数据X进行前处理和信息提取，得到压缩后的数据X 0'，发端参考神经网络可基于标准化的数据结构对该压缩后的数据X 0'进行处理，以获得由发端参考神经网络生成的参考空口信息S 0。发端目标神经网络可以基于与收端神经网络相应的算法对该发端数据X进行前处理和信息提取，得到压缩后的数据X 1'，发端目标神经网络可以基于标准化的数据结构对该压缩后的数据X 1'进行处理，以获得由发端目标神经网络生成的目标空口信息S 1。
其中，用于对发端数据X进行前处理的前处理模块可以是发端参考神经网络和发端目标神经网络共用的模块，即，对发端数据X进行前处理的过程是由同一个前处理模块执行 的。或者，用于对发端数据X进行前处理的前处理模块可以是分别包含在发端参考神经网络和发端目标神经网络中的模块，即，对发端数据X进行前处理的过程由发端参考神经网络和发端目标神经网络分别执行。
当然，前处理模块也可以不包含在发端参考神经网络和发端目标神经网络中，作为前处理模块单独存在于发射机中。前处理模块可将经过前处理后的数据分别输入至发端参考神经网络和发端目标神经网络。此情况下，发端数据X经由前处理模块的处理后被输入至发端目标神经网络和发端参考神经网络。
仍以信道信息的提取和压缩与信道信息的重构为例。发端数据X可以是信道H。在一种实现方式中，信道H可以被分别输入至发端参考神经网络和发端目标神经网络中，发端参考神经网络和发端目标神经网络分别可以按照如图2中所列举的发端神经网络的操作流程，对信道H先后进行前处理、信息提取和后处理，分别得到参考空口信息S 0和目标空口信息S 1。
此后，获取设备可以将参考空口信息S 0与目标空口信息S 1输入至性能度量函数中，以获得性能度量值。应理解，获取设备可以将该参考空口信息S 0与目标空口信息S 1输入至一个或多个性能度量函数中，获得一个或多个性能度量值。
还应理解，获取设备还可从参考数据集中获得的一个或多个发端数据按照上述流程来处理，可以获得多对参考空口信息与目标空口信息。由同一个发端数据所获得的一对参考空口信息与目标空口信息可以被输入至一个或多个性能度量函数中，获得一个或多个性能度量值。多对参考空口信息与目标空口信息可获得多个性能度量值。
该获取设备可以进一步根据标准化的性能判断准则，将输出的性能度量值与标准化的性能指标进行比较，再根据比较结果确定该发端目标神经网络是否满足预设的输出条件。
由于上文已经结合多个示例对性能判断准则做了详细说明，这里为了简洁，不再列举更多的输出条件，并针对不同的输出条件一一举例详述。
如图10所示，获取设备可以将同一个或多个空口信息输入至收端参考神经网络和收端目标神经网络。该一个或多个空口信息可以是直接从参考数据集中获取到的，也可以是将从参考数据集中获取到的发端数据输入至某一发射机之后，从该发射机获取到的空口信息。本申请对此不做限定。只要每次输入至收端参考神经网络和收端目标神经网络的空口信息是相同的空口信息即可。
在一种可能的设计中，输入至收端参考神经网络和收端目标神经网络的空口信息可以是通过如下方式获得的：将参考数据集中获取到的发端数据输入至一个发端参考神经网络中，该一个发端参考神经网络可根据预定义的数据结构对输入的发端数据进行处理，可以得到待发送给收端目标神经网络和收端参考神经网络的空口信息。
其中，发端参考神经网络输出的空口信息可以被输入至与之耦合的收端参考神经网络。换言之，本实施例中可通过相互耦合的一对发端参考神经网络和收端参考神经网络来对收端目标神经网络进行评估。
以其中的一个空口信息S为例。空口信息S可以被输入至收端目标神经网络和收端参考神经网络中。收端参考神经网络可以基于标准化的数据结构对该空口信息S进行解析，并可进一步基于预先设计好的算法对解析结果进行处理，以获得由收端参考神经网络输出的参考输出结果Y 0。收端目标神经网络可以基于标准化的数据结构对空口信息S进行解 析，并可进一步基于收端参考神经网络相应的算法对解析结果进行处理，以获得由收端目标神经网络输出的目标输出结果Y 1。收端参考神经网络和收端目标神经网络
仍以信道信息的提取和压缩与信道信息的重构为例。输入收端目标神经网络和收端参考神经网络的空口信息均为S。收端参考神经网络在接收到空口信息S后，可以基于图2中所列举的收端神经网络的操作流程，对空口信息S先后进行逆后处理、逆信息提取、逆前处理，以获得信道H的估计值
Figure PCTCN2021102079-appb-000042
收端目标神经网络在接收到空口信息S后，可以基于图2中所列举的收端神经网络的操作流程，对空口信息S先后进行逆后处理、逆信息提取、逆前处理，以获得信道H的估计值
Figure PCTCN2021102079-appb-000043
应理解，参考神经网络输出的估计值
Figure PCTCN2021102079-appb-000044
也即参考输出结果。目标神经网络输出的估计值
Figure PCTCN2021102079-appb-000045
也即目标输出结果。
此后，获取设备可以将参考输出结果
Figure PCTCN2021102079-appb-000046
与目标输出结果
Figure PCTCN2021102079-appb-000047
输入至性能度量函数中，以获得性能度量值。应理解，获取设备可以将该参考输出结果
Figure PCTCN2021102079-appb-000048
与目标输出结果
Figure PCTCN2021102079-appb-000049
输入至一个或多个性能度量函数中，获得一个或多个性能度量值。
还应理解，获取设备可将一个或多个空口信息按照上述流程来处理，以获得从收端目标神经网络输出的多个目标输出结果和从参考神经网络输出的多个参考输出结果。由同一个空口信息所获得的一对目标输出结果和参考输出结果可以被输入至一个或多个性能度量函数中，获得一个或多个性能度量值。多对目标输出结果和参考输出结果可获得多个性能度量值。
该获取设备可以进一步根据标准化的性能判断准则，将输出的性能度量值与标准化的性能指标进行比较，再根据比较结果确定该发端目标神经网络是否满足预设的输出条件。
由于上文已经结合多个示例对性能判断准则做了详细说明，这里为了简洁，不再列举更多的输出条件，并针对不同的输出条件一一举例详述。
上文结合图7至图10主要描述了获取设备结合一个标准化的发端参考神经网络和/或收端参考神经网络来对目标神经网络进行评估的具体过程。但应理解，上文所列举的示例不应对本申请构成任何限定。
例如，该获取设备还可以使用两个不同的参考神经网络对某一目标神经网络进行评估。该两个参考神经网络可以基于同一个输入生成不同的两个输出结果。该两个输出结果可以分别是从获取不同的性能下限出发而设计的两个参考神经网络的输出结果，或者，该两个输出结果也可以分别是从获取性能的上限和下限出发而设计的两个参考神经网络的输出结果。因此，该两个输出结果均可用于对目标神经网络是否满足输出条件做判别。
更具体地说，若该两个输出结果分别是从获取不同的性能下限出发而设计的两个不同的参考神经网络的输出结果，该两个输出结果可以分别与发端数据输入至不同的性能度量函数中，以得到不同的性能度量值，该性能度量值可与相应的性能指标的下界比较。在该两个输出结果均超出性能指标的下界的情况下，确定该目标神经网络满足输出条件。从而可以保证该目标神经网络在使用时性能不低于下限。
若该两个输出结果分别是从获取性能的上限和下限出发而设计的两个参考神经网络的输出结果，则该两个输出结果中，从获取性能的上限出发而设计的参考神经网络的输出结果可与发端数据被输入至性能度量函数中，以得到性能度量值，该性能度量值可用于和性能指标的上界比较，从而可以避免在目标神经网络在使用时性能超出上限；从获取性能的下限出发而设计的参考神经网络的输出结果可与发端数据被输入至性能度量函数中，以 得到另一性能度量值，该性能度量值可用于和性能指标的下界比较，从而可以保证该目标神经网络在使用时性能不低于下限。
下文图11至图14示出了获取设备结合多个标准化的发端参考神经网络和/或收端参考神经网络来对目标神经网络进行评估的具体过程。
图11示出了获取设备结合两个标准化的参考神经网络来判别目标神经网络是否满足输出条件的具体过程。在图11中，目标神经网络可与参考神经网络级联，目标神经网络可以用于发射机中，为发端目标神经网络；参考神经网络可以用于接收机中，为收端参考神经网络。
图11所示的两个收端参考神经网络中的一个(例如图中的收端参考神经网络1)可以是从获取某一性能(如MSE)的下限出发而设计的神经网络，另一个(例如图中的收端参考神经网络2)是从获取另一性能(如相关系数)的下限出发而设计的神经网络。由发端目标神经网络输出的空口信息S分别被输入至收端参考神经网络1和收端参考神经网络2。收端参考神经网络1和收端参考神经网络2可分别基于图2中所列举的收端神经网络的操作流程，获得输出结果，例如，由收端参考神经网络1可以得到输出结果Y 01，由收端参考神经网络2可以得到输出结果Y 02。基于上文所述的方法，获取设备可以将输入至发端目标神经网络的发端数据X分别与收端参考神经网络生成的输出结果Y 01和Y 02输入至性能度量函数中，以获得性能度量值，并进一步根据性能判断准则和性能度量值，确定该发端目标神经网络是否满足输出条件。
图11所示的性能度量函数包括MSE和相关系数。图11中由发端数据X和输出结果Y 01可以得到性能度量值MSE(X，Y 01)；由发端数据X和输出结果Y 02可以得到性能度量值相关系数(X，Y 02)。获取设备进一步结合标准化的MSE的下界和相关系数的下界，确定该发端目标神经网络是否满足输出条件。
可选地，输出条件为：由收端参考神经网络1生成的输出结果Y 01与发端数据的性能度量值达到MSE的下界，且，由收端参考神经网络2生成的输出结果Y 02与发端数据的性能度量值达到相关系数的下界。由于输出上述两个输出结果的两个收端参考神经网络是分别是基于不同的性能下限设计而得到的，因此所获得的性能度量值可分别与不同的性能指标的下界比较。
在本实施例中，若MSE(X，Y 01)达到性能指标的下界，相关系数(X，Y 02)达到相关系数的下界，则该发端目标神经网络满足输出条件。
若上述条件中有一个未满足，则该目标神经网络不满足输出条件。比如，对于输出条件1，若MSE(X，Y 01)未达到MSE的下界，或，相关系数(X，Y 02)未达到相关系数的下界，则可判定该发端目标神经网络不满足输出条件。
应理解，上文MSE的下界和相关系数的下界可以该目标神经网络所满足的能力相匹配的下界。例如，获取设备可以根据所选择的目标神经网络所满足的可实现能力度量值选择所对应的性能指标来对该目标神经网络进行评估。
还应理解，上述两个输出结果Y 01和Y 02可以是由两个不同的参考神经网络对同一空口信息进行处理得到的，例如一个为高复杂深度卷积网络，另一个为低复杂轻型卷积网络，本申请对此不作限定。
应理解，用于与该发端目标神经网络级联的收端参考神经网络可以为两个，也可以为 更多个。本申请对此不作限定。
例如，将更多个收端参考神经网络与该发端目标神经网络级联。这些收端参考神经网络中的一部分(例如记为收端参考神经网络集合1)可以是从获取性能的下限出发而设计的神经网络，另一部分(例如记为收端参考神经网络集合2)可以是从获取性能的上限出发而设计的神经网络。由此可以得到更多个收端神经网络的输出结果。这些输出结果也可分别与输入至发端目标神经网络的发端数据输入至性能度量函数中，以获得多个性能度量值。
此情况下，该输出条件例如可以为：由收端参考神经网络集合1输出的结果与发端数据的性能度量值均达到性能指标的下界，由收端参考神经网络集合2输出的结果与发端数据的性能度量值均未超出性能指标的上界。或者，该输出条件还可以为：由收端参考神经网络集合1输出的结果与发端数据的性能度量值均达到性能指标的下界，由收端参考神经网络集合2输出的结果与发端数据的性能度量值中90％以上的值未超出性能指标的上界，等等。为了简洁，这里不再列举。应理解，本申请对于输出条件的具体内容不做限定。
图12示出了获取设备结合两个标准化的参考神经网络来判别目标神经网络是否满足输出条件的具体过程。图12中的目标神经网络可与参考神经网络级联，目标神经网络应用于接收机中，为收端目标神经网络；参考神经网络应用于发射机，为发端参考神经网络。图12所示的两个参考神经网络中的一个(例如图中的参考神经网络1)可以是从获取某性能(如MSE)的下限出发而设置的神经网络，另一个(例如图中的参考神经网络2)是从获取另一性能(如相关系数)的下限出发而设置的神经网络。由发端参考神经网络1可以得到空口信息S 01，由发端参考神经网络2可以得到空口信息S 02。收端目标神经网络可以基于图2中所示的收端神经网络的操作流程，基于空口信息S 01和S 02分别获得输出结果Y 1和Y 2。基于上文所述的方法，获取设备可以将输入至发端参考神经网络的发端数据X与收端目标神经网络基于空口信息S 01和S 02分别得到的输出结果Y 1和Y 2输入性能度量函数中，以获得性能度量值，并进一步根据性能判断准则和性能度量值，确定该目标神经网络是否满足输出条件。
图12所示的性能度量函数包括MSE和相关系数。由发端数据X和输出结果Y 2可以得到性能度量值相关系数(X，Y 2)。获取设备进一步结合标准化的MSE的下界和相关系数的下界，确定该收端目标神经网络是否满足输出条件。由于上文已经结合性能判断准则对该判别过程做了详细说明，为了简洁，这里不再赘述。
通过性能指标的上界和下界来判别目标神经网络是否满足输出条件并不限于上文所列举的将该目标神经网络与一个参考神经网络级联的方式。训练设备还可以将同一个或多个发端数据输入至发端目标神经网络和发端参考神经网络，并将发端目标神经网络的输出结果与发端参考神经网络的输出结果进行比对，以确定该发端目标神经网络是否满足输出条件。
图13示出了获取设备结合两个标准化的参考神经网络来判别目标神经网络是否满足输出条件的具体过程。目标神经网络和参考神经网络都可应用于发射机中，分别为发端目标神经网络和发端参考神经网络。
图13所示的两个发端参考神经网络中的一个(例如图中的发端参考神经网络1)可以是从获取某一性能(如MSE)的下限出发而设置的神经网络，另一个(例如图中的发 端参考神经网络2)是从获取另一性能(如相关系数)的下限出发而设置的神经网络。同一发端数据X被分别输入至发端参考神经网络1、发端参考神经网络2和发端目标神经网络。发端参考神经网络1、发端参考神经网络2和发端目标神经网络可分别对该发端数据X进行处理，例如按照图2中所列举的发端神经网络的操作流程对发端数据X进行处理，以生成不同的空口信息。由发端参考神经网络1可以生成参考空口信息S 01，由发端参考神经网络2可以生成参考空口信息S 02，由发端目标神经网络可以生成目标空口信息S 1。基于上文所述的方法，获取设备将目标空口信息S 1、参考空口信息S 01和S 02输入至性能度量函数中，以获得性能度量值，并可进一步根据性能判断准则和性能度量值，确定该目标神经网络是否满足输出条件。由于上文已经结合性能判断准则对该判别过程做了详细说明，为了简洁，这里不再赘述。
图14示出了获取设备结合两个标准化的参考神经网络来判别目标神经网络是否满足输出条件的具体过程。图14中的参考神经网络包括应用于发射机中的发端参考神经网络和应用于接收机中的收端参考神经网络；目标神经网络应用于接收机，为收端目标神经网络。
获取设备可将从参考数据集中获取的发端数据输入至发端参考神经网络。发端参考神经网络可基于标准化的数据接口，生成空口信息S。该空口信息S可被输入至收端目标神经网络和两个收端参考神经网络中。其中，该两个收端参考神经网络中的一个(例如图中的收端参考神经网络1)可以是从获取某一性能(如MSE)的下限出发而设计的神经网络，另一个(例如图中的收端参考神经网络2)可以是从获取另一性能(如相关系数)的下限出发而设计的神经网络。收端参考神经网络1、收端参考神经网络2和发端目标神经网络可分别基于标准化的数据结构，对该空口信息S进行解析，得到不同的输出结果。由收端参考神经网络1可以得到参考输出结果Y 01，由收端参考神经网络2可以得到参考输出结果Y 02，由收端目标神经网络可以得到目标输出结果Y 1。基于上文所述的方法，获取设备可以将目标输出结果Y 1、参考输出结果Y 01和Y 02输入至性能度量函数中，以获得性能度量值，并可进一步根据性能判断准则和性能度量值，确定该收端目标神经网络是否满足输出条件。由于上文已经结合性能判断准则对该判别过程做了详细说明，为了简洁，这里不再赘述。
应理解，上文仅为便于理解，结合图7至图14详细说明了上述步骤610的具体实现过程。但这些示例仅为便于理解而示出，不应对本申请构成任何限定。基于相同的构思，本领域的技术人员可以对上述步骤中的部分或全部进行等价替换或简单变形，以达到相同的效果。因此，这些等价替换或简单变形均应落入本申请的保护范围内。
还应理解，图7至图14主要以图2所示的发端神经网络和收端神经网络所包括的模块为例来描述了步骤610的具体过程。但这不应对本申请构成任何限定。在对发端神经网络和收端神经网络所包含的模块做出不同设计的情况下，发端神经网络的输入和输出可能与图中所示不同，收端神经网络的输入和输出也可能与图中所示不同。
此后，获取设备基于对目标神经网络是否满足输出条件，可以执行步骤620或步骤630中的一项。若目标神经网络满足输出条件，则可执行步骤620，输出该目标神经网络；若目标神经网络不满足输出条件，则可执行步骤630，继续对该目标神经网络进行优化。
获取设备对该目标神经网络进行优化的过程具体可以包括但不限于，调整神经网络结 构设计、调整参考数据集、调整神经网络训练方法、调整代价函数、损失函数或目标函数的定义、改变神经网络的初始化方法、改变神经网络参数的约束条件、以及改变神经网络的激活函数的定义等等。
通过对上文列举的一项或多项的执行，获取设备可以获得优化的目标神经网络。获取设备可以进一步对优化的目标神经网络进行评估，以确定其是否满足输出条件。获取设备对优化的目标神经网络进行评估的具体过程可以参考上文步骤610中的相关描述，为了简洁，这里不再赘述。
基于上述技术方案，通过引入标准化的参考数据集、标准化的参考神经网络以及标准化的性能判断准则，使得获取设备可以基于已有的标准化的参考神经网络，对目标神经网络进行评估，在不满足输出条件的情况下，可对该目标神经网络进行优化，直到目标神经网络满足输出条件。从而使得通过空口来下载神经网络结构和参数信息的开销得以避免。同时还可以保证不同厂商的产品之间能够互联互通，在保证性能的前提下，还可以体现不同厂商的产品的差异化设计和竞争力，因此从整体上说，大幅提升了对偶网络实现的可行性。
应理解，上文仅为便于理解，以信道信息的提取和压缩与信道信息的重构为例对上述方法做了进一步详尽的说明。但这不应对本申请构成任何限定。在上文所列举的信息的其他处理阶段，本申请实施例所提供的获取神经网络的方法也仍然适用。为了简洁，这里不一一举例说明。
基于上文所述的方法获取的目标神经网络，可以应用于通信网络中。例如，发端目标神经网络可应用于网络设备中，收端目标神经网络可应用于终端设备中。网络设备可以包括如前文所述的发射机，终端设备可以包括如前文所述的接收机。又例如，发端目标神经网络可应用终端设备中，收端目标神经网络可应用于网络设备中。终端设备可以包括如前文所述的发射机，网络设备可以包括如前文所述的接收机。本申请对此不作限定。
由于网络设备与终端设备之间的空口传输条件因所处环境的不同而不同，在选择目标神经网络应用于网络设备和终端设备中时，可以根据当前空口传输条件来选择相匹配的场景模式。其中，空口传输条件可以包括但不限于，信道状态、干扰状态等。
网络设备和终端设备中的一个可以基于空口传输条件来确定场景模式，以便确定用于适用于当前空口传输条件的目标神经网络。
图15是本申请实施例提供的通信方法的示意性流程图。图15所示的方法700可以包括步骤710至步骤750。下面详细说明方法700中的各个步骤。
在步骤710中，终端设备发送终端设备的能力信息。相应地，网络设备接收终端设备的能力信息。
具体地，该能力信息可以用于指示可实现能力度量值。该可实现能力度量值具体可以是指终端设备可以采用的神经网络的可实现能力度量值。网络设备可以根据终端设备上报的可实现能力度量值，以及由与该终端设备之间的空口传输条件确定的场景模式，来选择相匹配的神经网络。
关于可实现能力度量值的具体内容在上文方法600中已经做了详细说明，为了简洁，这里不再赘述。
在步骤720中，网络设备基于与终端设备之间的空口传输条件，确定场景模式。
网络设备例如可以持续地或周期性地获取与终端设备之间的空口传输条件，例如通过对信道的测量、干扰的测量等来获取该空口传输信息。该周期的具体时长可以是预定义值。本申请对于网络设备获取空口传输条件的实际行为不作限定。
网络设备可以基于与终端设备之间的空口传输条件，如信道状态、干扰状态等，确定当前的场景模式。该场景模式例如可以包括但不限于：室外密集城区、室外普通城区、室外乡村、室外山区、室内办公、室内厂房等的不同信道场景，单干扰源、多干扰源、强弱不同的多种干扰信号类型，中近点高信噪比、远点低信噪比的不同信噪比条件，步行低速、城区车载中速、高铁超高速等不同移动条件等。由于不同的场景对空口传输条件可能会有不同的影响，故可以将不同的场景模式与不同的神经网络相匹配，以将与当前空口传输条件相匹配的神经网络应用到网络设备和终端设备中。在步骤730中，网络设备向终端设备发送场景模式的指示信息。相应地，终端设备接收来自网络设备的场景模式的指示信息。
网络设备在确定了场景模式之后，可以通过信令向终端设备指示该场景模式的指示信息，以便于终端设备选择与网络设备一对一耦合的神经网络来协同工作。
网络设备指示该场景模式的具体方式可以有多种。例如，可以预先定义多个指示比特与多种场景模式的对应关系，通过不同的指示比特来指示不同的场景模式。
可选地，该场景模式的指示信息携带在高层信令中。该高层信令例如为无线资源控制(radio resource control，RRC)消息或介质接入控制(medium access control，MAC)控制元素(control element，CE)。
即，该场景模式可以是静态配置，或半静态配置的。
可选地，该场景模式的指示信息携带在物理层信令中。该物理层信令例如为下行控制信息(downlink control information，DCI)。
即，该场景模式可以是动态配置的。
在步骤740中，网络设备接收来自终端设备的确认信息，该确认信息用于指示对场景模式的指示信息的成功接收。
终端设备基于对上述指示信息的成功接收，可以向网络设备发送确认信息。
例如，终端设备可以在成功接收的情况下发送该确认信息，在未成功接收的情况下不发送该确认信息。网络设备可以在发送上述场景模式的指示信息后开启计时器，该计时器的时长可以是预定义值。若在计时器超时前接收到该确认信息，则可确定终端设备成功接收到场景模式的指示信息；否则，认为该终端设备未成功接收到该场景模式的指示信息。换言之，步骤740为可选的步骤。若终端设备未成功接收到该指示信息，网络设备可能无法接收到来自终端设备的确认信息。
又例如，该确认信息可以是一个指示比特。例如“0”或“空(null)”表示未成功接收；“1”表示成功接收。
应理解，该确认信息的具体形式可以为很多种，本申请对此不做限定。
在步骤750中，网络设备采用与该场景模式、终端设备的能力信息匹配的神经网络与终端设备通信。相应地，终端设备也采用与该场景模式、该终端设备的能力信息匹配的神经网络与网络设备通信。
网络设备可以基于该确认信息，确定终端设备成功接收到上述指示信息。网络设备和终端设备自此可以采用与该场景模式、终端设备的能力信息匹配的发端神经网络和收端神 经网络工作。
例如，网络设备可以采用与该场景模式、该终端设备的能力信息匹配的发端神经网络工作，终端设备可以采用与该场景模式、终端设备的能力信息匹配的收端神经网络工作；又如，网络设备可以采用与该场景模式、终端设备的能力信息匹配的收端神经网络工作，终端设备可以采用与该场景模式、终端设备的能力信息匹配的发端神经网络工作；再如，网络设备可以采用在与某一场景模式(如记作场景模式1)、终端设备的能力信息匹配的发端神经网络工作的同时，还可采用与另一场景模式(如记作场景模式2)、终端设备的能力信息匹配的收端神经网络工作，终端设备可以在采用与场景模式1、终端设备的能力信息匹配的收端神经网络工作的同时，还可采用与场景模式2、终端设备的能力信息匹配的发端神经网络工作。本申请对此不做限定。
应理解，在同一种场景模式下，可以有一个发端神经网络以及一个收端神经网络与之匹配。可以理解的是，每个发端神经网络都存在至少一个可与之协同工作的收端神经网络；每个收端神经网络都存在至少一个可与之协同工作的发端神经网络。
还应理解，发端神经网络和收端神经网络都可以是基于上文方法600中的方法获取到的。
若网络设备确定终端设备未成功接收到上述指示信息，网络设备可以重新发送该指示信息，例如可以重新执行上述流程，直到接收到该终端设备的确认信息。
还应理解，网络设备可以在发现传输空口条件发生变化时，重新确定场景模式，并在场景模式发生变化的情况下，将重新确定的场景模式指示给终端设备，以便于网络设备和终端设备基于新确定的场景模式采用相匹配的神经网络工作。即，上述步骤720至步骤750可重复执行。
基于上述方法，网络设备可以基于与终端设备之间当前的空口传输条件，选择合适的场景模式，并将该场景模式指示给终端设备，以便于网络设备和终端设备双方基于同一场景模式以及终端设备的能力信息来确定采用怎样的神经网络来工作。从而有利于获得性能的提升。
图16是本申请另一实施例提供的通信方法的示意性流程图。图16所示的方法800可以包括步骤810至步骤850。下面详细说明方法800中的各个步骤。
在步骤810中，终端设备上报终端设备的能力信息。相应地，网络设备接收终端设备的能力信息。在步骤820中，终端设备基于与网络设备之间的空口传输条件，确定场景模式。
在步骤830中，终端设备向网络设备发送场景模式的指示信息。相应地，网络设备接收终端设备的场景模式的指示信息。
在步骤840中，终端设备接收来自网络设备的确认信息，该确认信息用于指示对场景模式的指示信息的成功接收。
在步骤850中，终端设备采用与该场景模式、终端设备的能力信息匹配的神经网络与网络设备通信。相应地，网络设备采用与该场景模式、终端设备的能力信息匹配的神经网络与终端设备通信。
应理解，步骤820至步骤850的具体过程与上文方法700中的步骤720至步骤750的具体过程相似，所不同的是，在方法800中，由终端设备来根据空口传输条件确定场景模 式，并向网络设备指示该场景模式。网络设备基于对该指示信息的成功接收回复确认信息。此后，网络设备和终端设备都可以采用与该场景模式、终端设备的能力信息相匹配的神经网络工作。
由于上文方法700中已经对步骤720至步骤750的具体过程做了详细说明，为了简洁，这里不做详述。
应理解，终端设备可以在发现传输空口条件发生变化时，重新确定场景模式，并在场景模式发生变化的情况下，将重新确定的场景模式指示给网络设备，以便于网络设备和终端设备基于新确定的场景模式采用相匹配的神经网络工作。即，上述步骤810至步骤840也可以重复执行。
基于上述方法，终端设备可以基于与网络设备之间当前的空口传输条件，选择合适的场景模式，并将该场景模式指示给网络设备，以便于网络设备和终端设备双方基于同一场景模式以及终端设备的能力信息来确定采用怎样的神经网络来工作。从而有利于获得性能的提升。
以上，结合图6至图16详细说明了本申请实施例提供的通信网络中的神经网络的训练方法。以下，结合图17至图20详细说明本申请实施例提供的通信网络中的神经网络的训练装置。
图17是本申请实施例提供的获取神经网络的装置1000的示意性框图。如图17所示，该训练装置1000包括处理单元1100和输入输出单元1200。其中，处理单元1100可用于基于标准化的参考数据集、一个或多个参考神经网络和性能判断准则，确定构建的目标神经网络是否满足预设的输出条件；输入输出单元1200可用于在满足所述输出条件的情况下，输出所述目标神经网络，输出的所述目标神经网络是应用于发射机的发端神经网络或应用于接收机中的收端神经网络。
应理解，该装置1000可对应于图6至图14所示实施例中的获取设备，可以包括用于执行图6至图14中的方法600实施例中获取设备执行的方法的单元。该处理单元1100可用于执行上文方法600中的步骤610和步骤620，该输入输出单元1200可用于执行上文方法600中的步骤630。
其中，处理单元1100所执行的步骤可以由一个或多个处理器执行相应的程序来实现。例如对目标神经网络的训练可以由专用于训练神经网络的处理器来实现。该处理单元1100可对应于图18中所示的处理器2010。
输入输出单元1200所执行的步骤例如可以由输入输出接口、电路等来实现。该输入输出单元1200例如可对应于图18中所示的输入输出接口2020。
在一种可能的设计中，该装置1000可以部署在芯片上。
图18是本申请实施例提供的获取神经网络的装置2000的示意性框图。如图18所示，该装置2000包括处理器2100、输入输出接口2200和存储器2030。其中，处理器2010、输入输出接口2020和存储器2030通过内部连接通路互相通信，该存储器2030用于存储指令，该处理器2010用于执行该存储器2030存储的指令，以控制该输入输出接口2020发送信号和/或接收信号。
应理解，该装置2000可以对应于图6至图14所示实施例中的获取设备，可以包括用于执行图6至图14中的方法600实施例中获取设备执行的方法的单元。可选地，该存储 器2030可以包括只读存储器和随机存取存储器，并向处理器提供指令和数据。存储器的一部分还可以包括非易失性随机存取存储器。存储器2030可以是一个单独的器件，也可以集成在处理器2010中。该处理器2010可以用于执行存储器2030中存储的指令，并且当该处理器2010执行存储器中存储的指令时，该处理器2010用于执行上述与获取设备对应的方法实施例的各个步骤和/或流程。
图19是本申请实施例提供的通信装置的示意性框图。如图19所示，该通信装置3000可以包括处理单元3100和收发单元3200。
可选地，该通信装置3000可对应于上文方法实施例中的终端设备，例如，可以为终端设备，或者配置于终端设备中的部件(如电路、芯片或芯片系统等)。
应理解，该通信装置3000可对应于根据本申请实施例的方法700或方法800中的终端设备，该通信装置3000可以包括用于执行图15中的方法700或图16中的方法800中终端设备执行的方法的单元。并且，该通信装置3000中的各单元和上述其他操作和/或功能分别为了实现图15中的方法700或图16中的方法800的相应流程。
其中，当该通信装置3000用于执行图15中的方法700时，处理单元3100可用于执行方法700中的步骤750，收发单元3200可用于执行方法700中的步骤710、步骤730至步骤750。应理解，各单元执行上述相应步骤的具体过程在上述方法实施例中已经详细说明，为了简洁，在此不再赘述。
当该通信装置3000用于执行图16中的方法800时，处理单元3100可用于执行方法800中的步骤820和步骤850，收发单元3200可用于执行方法800中的步骤810、步骤830至步骤850。应理解，各单元执行上述相应步骤的具体过程在上述方法实施例中已经详细说明，为了简洁，在此不再赘述。
还应理解，该通信装置3000为终端设备时，该通信装置3000中的收发单元3200可以通过收发器实现，例如可对应于图20中示出的通信装置4000中的收发器4020或图21中示出的终端设备5000中的收发器5020，该通信装置3000中的处理单元3100可通过至少一个处理器实现，例如可对应于图20中示出的通信装置4000中的处理器4010或图21中示出的终端设备5000中的处理器5010。
还应理解，该通信装置3000为配置于终端设备中的芯片或芯片系统时，该通信装置3000中的收发单元3200可以通过输入/输出接口、电路等实现，该通信装置3000中的处理单元3100可以通过该芯片或芯片系统上集成的处理器、微处理器或集成电路等实现。
可选地，该通信装置3000可对应于上文方法实施例中的网络设备，例如，可以为网络设备，或者配置于网络设备中的部件(如电路、芯片或芯片系统等)。
应理解，该通信装置3000可对应于根据本申请实施例的方法700或方法800中的网络设备，该通信装置3000可以包括用于执行图15中的方法700或图16中的方法800中网络设备执行的方法的单元。并且，该通信装置3000中的各单元和上述其他操作和/或功能分别为了实现图15中的方法700或图16中的方法800的相应流程。
其中，当该通信装置3000用于执行图15中的方法700时，处理单元3100可用于执行方法700中的步骤720和步骤750，收发单元3200可用于执行方法700中的步骤710步骤730至步骤750。应理解，各单元执行上述相应步骤的具体过程在上述方法实施例中已经详细说明，为了简洁，在此不再赘述。
当该通信装置3000用于执行图16中的方法800时，处理单元3100可用于执行方法800中的步骤850，收发单元3200可用于执行方法800中的步骤810、步骤830至步骤850。应理解，各单元执行上述相应步骤的具体过程在上述方法实施例中已经详细说明，为了简洁，在此不再赘述。
还应理解，该通信装置3000为网络设备时，该通信装置3000中的收发单元3200可以通过收发器实现，例如可对应于图20中示出的通信装置4000中的收发器4020或图22中示出的基站6000中的RRU 6100，该通信装置3000中的处理单元3100可通过至少一个处理器实现，例如可对应于图20中示出的通信装置4000中的处理器4010或图22中示出的基站6000中的处理单元6200或处理器6202。
还应理解，该通信装置3000为配置于网络设备中的芯片或芯片系统时，该通信装置3000中的收发单元3200可以通过输入/输出接口、电路等实现，该通信装置3000中的处理单元3100可以通过该芯片或芯片系统上集成的处理器、微处理器或集成电路等实现。
图20是本申请实施例提供的通信装置4000的另一示意性框图。如图6所示，该通信装置4000包括处理器2010、收发器4020和存储器4030。其中，处理器4010、收发器4020和存储器4030通过内部连接通路互相通信，该存储器4030用于存储指令，该处理器4010用于执行该存储器4030存储的指令，以控制该收发器4020发送信号和/或接收信号。
应理解，该通信装置4000可以对应于上述方法实施例中的终端设备，并且可以用于执行上述方法实施例中网络设备或终端设备执行的各个步骤和/或流程。可选地，该存储器4030可以包括只读存储器和随机存取存储器，并向处理器提供指令和数据。存储器的一部分还可以包括非易失性随机存取存储器。存储器4030可以是一个单独的器件，也可以集成在处理器4010中。该处理器4010可以用于执行存储器4030中存储的指令，并且当该处理器4010执行存储器中存储的指令时，该处理器4010用于执行上述与网络设备或终端设备对应的方法实施例的各个步骤和/或流程。
可选地，该通信装置4000是前文实施例中的终端设备。
可选地，该通信装置4000是前文实施例中的网络设备。
其中，收发器4020可以包括发射机和接收机。收发器4020还可以进一步包括天线，天线的数量可以为一个或多个。该处理器4010和存储器4030与收发器4020可以是集成在不同芯片上的器件。如，处理器4010和存储器4030可以集成在基带芯片中，收发器4020可以集成在射频芯片中。该处理器4010和存储器4030与收发器4020也可以是集成在同一个芯片上的器件。本申请对此不作限定。
可选地，该通信装置4000是配置在终端设备中的部件，如电路、芯片、芯片系统等。
可选地，该通信装置4000是配置在网络设备中的部件，如电路、芯片、芯片系统等。
其中，收发器4020也可以是通信接口，如输入/输出接口、电路等。该收发器4020与处理器4010和存储器4020都可以集成在同一个芯片中，如集成在基带芯片中。
图21是本申请实施例提供的终端设备5000的结构示意图。该终端设备5000可应用于如图1所示的系统中，执行上述方法实施例中终端设备的功能。如图所示，该终端设备5000包括处理器5010和收发器5020。可选地，该终端设备5000还包括存储器5030。其中，处理器5010、收发器5020和存储器5030之间可以通过内部连接通路互相通信，传递控制和/或数据信号，该存储器5030用于存储计算机程序，该处理器5010用于从该存 储器5030中调用并运行该计算机程序，以控制该收发器5020收发信号。可选地，终端设备5000还可以包括天线3040，用于将收发器5020输出的上行数据或上行控制信令通过无线信号发送出去。
上述处理器5010可以和存储器5030可以合成一个处理装置，处理器5010用于执行存储器5030中存储的程序代码来实现上述功能。具体实现时，该存储器5030也可以集成在处理器5010中，或者独立于处理器5010。该处理器5010可以与图19中的处理单元3100或图20中的处理器4010对应。
上述收发器5020可以与图19中的收发单元3200或图20中的收发器4020对应。收发器5020可以包括接收器(或称接收机、接收电路)和发射器(或称发射机、发射电路)。其中，接收器用于接收信号，发射器用于发射信号。
应理解，图21所示的终端设备5000能够实现图15或图16所示方法实施例中涉及终端设备的各个过程。终端设备5000中的各个模块的操作和/或功能，分别为了实现上述方法实施例中的相应流程。具体可参见上述方法实施例中的描述，为避免重复，此处适当省略详细描述。
上述处理器5010可以用于执行前面方法实施例中描述的由终端设备内部实现的动作，而收发器5020可以用于执行前面方法实施例中描述的终端设备向网络设备发送或从网络设备接收的动作。具体请见前面方法实施例中的描述，此处不再赘述。
可选地，上述终端设备5000还可以包括电源5050，用于给终端设备中的各种器件或电路提供电源。
除此之外，为了使得终端设备的功能更加完善，该终端设备5000还可以包括输入单元5060、显示单元5070、音频电路5080、摄像头5090和传感器5100等中的一个或多个，所述音频电路还可以包括扬声器5082、麦克风5084等。
图22是本申请实施例提供的网络设备的结构示意图，例如可以为基站的结构示意图。该基站6000可应用于如图1所示的系统中，执行上述方法实施例中网络设备的功能。如图所示，该基站6000可以包括一个或多个射频单元，如远端射频单元(remote radio unit，RRU)6100和一个或多个基带单元(BBU)(也可称为分布式单元(DU))6200。所述RRU 6100可以称为收发单元，可以与图19中的收发单元3200或图20中的收发器4020对应。可选地，该RRU 6100还可以称为收发机、收发电路、或者收发器等等，其可以包括至少一个天线6101和射频单元6102。可选地，RRU 6100可以包括接收单元和发送单元，接收单元可以对应于接收器(或称接收机、接收电路)，发送单元可以对应于发射器(或称发射机、发射电路)。所述RRU 6100部分主要用于射频信号的收发以及射频信号与基带信号的转换，例如用于向终端设备发送指示信息。所述BBU 6200部分主要用于进行基带处理，对基站进行控制等。所述RRU 6100与BBU 6200可以是物理上设置在一起，也可以物理上分离设置的，即分布式基站。
所述BBU 6200为基站的控制中心，也可以称为处理单元，可以与图19中的处理单元3100或图20中的处理器4010对应，主要用于完成基带处理功能，如信道编码，复用，调制，扩频等等。例如所述BBU(处理单元)可以用于控制基站执行上述方法实施例中关于网络设备的操作流程，例如，生成上述指示信息等。
在一个示例中，所述BBU 6200可以由一个或多个单板构成，多个单板可以共同支持 单一接入制式的无线接入网(如LTE网)，也可以分别支持不同接入制式的无线接入网(如LTE网，5G网或其他网)。所述BBU 6200还包括存储器6201和处理器6202。所述存储器6201用以存储必要的指令和数据。所述处理器6202用于控制基站进行必要的动作，例如用于控制基站执行上述方法实施例中关于网络设备的操作流程。所述存储器6201和处理器6202可以服务于一个或多个单板。也就是说，可以每个单板上单独设置存储器和处理器。也可以是多个单板共用相同的存储器和处理器。此外每个单板上还可以设置有必要的电路。
应理解，图22所示的基站6000能够实现图15或图16所示方法实施例中涉及网络设备的各个过程。基站6000中的各个模块的操作和/或功能，分别为了实现上述方法实施例中的相应流程。具体可参见上述方法实施例中的描述，为避免重复，此处适当省略详细描述。
上述BBU 6200可以用于执行前面方法实施例中描述的由网络设备内部实现的动作，而RRU 6100可以用于执行前面方法实施例中描述的网络设备向终端设备发送或从终端设备接收的动作。具体请见前面方法实施例中的描述，此处不再赘述。
应理解，图22所示出的基站6000仅为网络设备的一种可能的形态，而不应对本申请构成任何限定。本申请所提供的方法可适用于其他形态的网络设备。例如，包括AAU，还可以包括CU和/或DU，或者包括BBU和自适应无线单元(adaptive radio unit，ARU)，或BBU；也可以为客户终端设备(customer premises equipment，CPE)，还可以为其它形态，本申请对于网络设备的具体形态不做限定。
其中，CU和/或DU可以用于执行前面方法实施例中描述的由网络设备内部实现的动作，而AAU可以用于执行前面方法实施例中描述的网络设备向终端设备发送或从终端设备接收的动作。具体请见前面方法实施例中的描述，此处不再赘述。
本申请还提供了一种处理装置，包括至少一个处理器，所述至少一个处理器用于执行存储器中存储的计算机程序，以使得所述处理装置执行上述方法实施例中获取设备所执行的方法、网络设备所执行的方法或终端设备所执行的方法。
本申请实施例还提供了一种处理装置，包括处理器和输入输出接口。所述输入输出接口与所述处理器耦合。所述输入输出接口用于输入和/或输出信息。所述信息包括指令和数据中的至少一项。所述处理器用于执行计算机程序，以使得所述处理装置执行上述方法实施例中获取设备所执行的方法、网络设备所执行的方法或终端设备所执行的方法。
本申请实施例还提供了一种处理装置，包括处理器和存储器。所述存储器用于存储计算机程序，所述处理器用于从所述存储器调用并运行所述计算机程序，以使得所述处理装置执行上述方法实施例中训练设备所执行的方法。
应理解，上述处理装置可以是一个或多个芯片。例如，该处理装置可以是现场可编程门阵列(field programmable gate array，FPGA)，可以是专用集成芯片(application specific integrated circuit，ASIC)，还可以是系统芯片(system on chip，SoC)，还可以是中央处理器(central processor unit，CPU)，还可以是网络处理器(network processor，NP)，还可以是数字信号处理电路(digital signal processor，DSP)，还可以是微控制器(micro controller unit，MCU)，还可以是可编程控制器(programmable logic device，PLD)或其他集成芯片。
在实现过程中，上述方法的各步骤可以通过处理器中的硬件的集成逻辑电路或者软件形式的指令完成。结合本申请实施例所公开的方法的步骤可以直接体现为硬件处理器执行完成，或者用处理器中的硬件及软件模块组合执行完成。软件模块可以位于随机存储器，闪存、只读存储器，可编程只读存储器或者电可擦写可编程存储器、寄存器等本领域成熟的存储介质中。该存储介质位于存储器，处理器读取存储器中的信息，结合其硬件完成上述方法的步骤。为避免重复，这里不再详细描述。
应注意，本申请实施例中的处理器可以是一种集成电路芯片，具有信号的处理能力。在实现过程中，上述方法实施例的各步骤可以通过处理器中的硬件的集成逻辑电路或者软件形式的指令完成。上述的处理器可以是通用处理器、数字信号处理器(DSP)、专用集成电路(ASIC)、现场可编程门阵列(FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件。可以实现或者执行本申请实施例中的公开的各方法、步骤及逻辑框图。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。结合本申请实施例所公开的方法的步骤可以直接体现为硬件译码处理器执行完成，或者用译码处理器中的硬件及软件模块组合执行完成。软件模块可以位于随机存储器，闪存、只读存储器，可编程只读存储器或者电可擦写可编程存储器、寄存器等本领域成熟的存储介质中。该存储介质位于存储器，处理器读取存储器中的信息，结合其硬件完成上述方法的步骤。
可以理解，本申请实施例中的存储器可以是易失性存储器或非易失性存储器，或可包括易失性和非易失性存储器两者。其中，非易失性存储器可以是只读存储器(read-only memory，ROM)、可编程只读存储器(programmable ROM，PROM)、可擦除可编程只读存储器(erasable PROM，EPROM)、电可擦除可编程只读存储器(electrically EPROM，EEPROM)或闪存。易失性存储器可以是随机存取存储器(random access memory，RAM)，其用作外部高速缓存。通过示例性但不是限制性说明，许多形式的RAM可用，例如静态随机存取存储器(static RAM，SRAM)、动态随机存取存储器(dynamic RAM，DRAM)、同步动态随机存取存储器(synchronous DRAM，SDRAM)、双倍数据速率同步动态随机存取存储器(double data rate SDRAM，DDR SDRAM)、增强型同步动态随机存取存储器(enhanced SDRAM，ESDRAM)、同步连接动态随机存取存储器(synchlink DRAM，SLDRAM)和直接内存总线随机存取存储器(direct rambus RAM，DR RAM)。应注意，本文描述的系统和方法的存储器旨在包括但不限于这些和任意其它适合类型的存储器。
根据本申请实施例提供的方法，本申请还提供一种计算机程序产品，该计算机程序产品包括：计算机程序代码，当该计算机程序代码在计算机上运行时，使得该计算机执行图6所示实施例中获取设备执行的方法、图15或图16所示实施例中的终端设备执行的方法或网络设备执行的方法。
根据本申请实施例提供的方法，本申请还提供一种计算机可读存储介质，该计算机可读存储介质存储有程序代码，当该程序代码在计算机上运行时，使得该计算机执行图6所示实施例中获取设备执行的方法、图15或图16所示实施例中的终端设备执行的方法或网络设备执行的方法。
根据本申请实施例提供的方法，本申请还提供一种系统，其包括前述的一个或多个终端设备以及一个或多个网络设备，所述一个或多个终端设备和所述一个或多个网络设备中 分别配置了通过前述获取设备获取的目标神经网络。
上述各个装置实施例中网络设备与终端设备和方法实施例中的网络设备或终端设备完全对应，由相应的模块或单元执行相应的步骤，例如通信单元(收发器)执行方法实施例中接收或发送的步骤，除发送、接收外的其它步骤可以由处理单元(处理器)执行。具体单元的功能可以参考相应的方法实施例。其中，处理器可以为一个或多个。
上述实施例中，终端设备可以作为接收设备的一例，网络设备可以作为发送设备的一例。但这不应对本申请构成任何限定。例如，发送设备和接收设备也可以均为终端设备等。本申请对于发送设备和接收设备的具体类型不作限定。
在本说明书中使用的术语“部件”、“模块”、“系统”等用于表示计算机相关的实体、硬件、固件、硬件和软件的组合、软件、或执行中的软件。例如，部件可以是但不限于，在处理器上运行的进程、处理器、对象、可执行文件、执行线程、程序和/或计算机。通过图示，在计算设备上运行的应用和计算设备都可以是部件。一个或多个部件可驻留在进程和/或执行线程中，部件可位于一个计算机上和/或分布在2个或更多个计算机之间。此外，这些部件可从在上面存储有各种数据结构的各种计算机可读介质执行。部件可例如根据具有一个或多个数据分组(例如来自与本地系统、分布式系统和/或网络间的另一部件交互的二个部件的数据，例如通过信号与其它系统交互的互联网)的信号通过本地和/或远程进程来通信。
本领域普通技术人员可以意识到，结合本文中所公开的实施例描述的各示例的单元及算法步骤，能够以电子硬件、或者计算机软件和电子硬件的结合来实现。这些功能究竟以硬件还是软件方式来执行，取决于技术方案的特定应用和设计约束条件。专业技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能，但是这种实现不应认为超出本申请的范围。
所属领域的技术人员可以清楚地了解到，为描述的方便和简洁，上述描述的系统、装置和单元的具体工作过程，可以参考前述方法实施例中的对应过程，在此不再赘述。
在本申请所提供的几个实施例中，应该理解到，所揭露的系统、装置和方法，可以通过其它的方式实现。例如，以上所描述的装置实施例仅仅是示意性的，例如，所述单元的划分，仅仅为一种逻辑功能划分，实际实现时可以有另外的划分方式，例如多个单元或组件可以结合或者可以集成到另一个系统，或一些特征可以忽略，或不执行。另一点，所显示或讨论的相互之间的耦合或直接耦合或通信连接可以是通过一些接口，装置或单元的间接耦合或通信连接，可以是电性，机械或其它的形式。
所述作为分离部件说明的单元可以是或者也可以不是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部单元来实现本实施例方案的目的。
另外，在本申请各个实施例中的各功能单元可以集成在一个处理单元中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个单元中。
所述功能如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本申请的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案的部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质中，包括若干指令用以使得一台计算机设备(可以是个人计 算机，服务器，或者网络设备等)执行本申请各个实施例所述方法的全部或部分步骤。而前述的存储介质包括：U盘、移动硬盘、只读存储器(read-only memory，ROM)、随机存取存储器(random access memory，RAM)、磁碟或者光盘等各种可以存储程序代码的介质。
以上所述，仅为本申请的具体实施方式，但本申请的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本申请揭露的技术范围内，可轻易想到变化或替换，都应涵盖在本申请的保护范围之内。因此，本申请的保护范围应以所述权利要求的保护范围为准。

Claims (43)

  1. 一种获取神经网络的方法，其特征在于，包括：
    基于标准化的参考数据集、一个或多个参考神经网络和性能判断准则，确定构建的目标神经网络是否满足预设的输出条件；
    在满足所述输出条件的情况下，输出所述目标神经网络，输出的所述目标神经网络是应用于发射机的发端目标神经网络或应用于接收机中的收端目标神经网络。
  2. 如权利要求1所述的方法，其特征在于，所述方法还包括：
    在不满足所述输出条件的情况下，继续对所述目标神经网络进行优化，直至获得的目标神经网络满足所述输出条件。
  3. 如权利要求1或2所述的方法，其特征在于，所述发端目标神经网络用于星座调制、信道编码、预编码、或信道信息的提取和压缩；所述收端目标神经网络用于星座解调、信道解码、检测接收、或信道信息的重构。
  4. 如权利要求1至3中任一项所述的方法，其特征在于，所述方法还包括：
    从至少一套标准化的配置中确定适用于所述目标神经网络当前应用场景的配置；每套标准化的配置包括以下一项或多项：标准化的参考数据集、标准化的参考神经网络、标准化的数据结构、标准化的性能度量函数、标准化的性能判断准则、以及标准化的可实现能力度量值；
    其中，所述参考神经网络包括应用于接收机中的收端参考神经网络和/或应用于发射机中的发端参考神经网络；所述数据结构用于生成和/或解析空口信息；所述性能度量函数用于生成性能度量值，所述性能判断准则包括性能指标和所述目标神经网络的输出条件，所述性能判断准则用于基于所述性能度量值和所述性能指标，确定所述目标神经网络是否满足所述输出条件；所述可实现能力度量值用于选择匹配的神经网络和/或匹配的性能指标。
  5. 如权利要求4所述的方法，其特征在于，所述目标神经网络为发端目标神经网络，所述一个或多个参考神经网络为一个或多个收端参考神经网络；以及
    所述基于标准化的参考数据集、一个或多个参考神经网络和性能判断准则，确定设计得到的目标神经网络是否满足预设的输出条件，包括：
    将从所述参考数据集中获取到的发端数据输入至所述发端目标神经网络中，所述发端目标神经网络用于根据预定义的数据结构对所述发端数据进行处理，以生成待发送给所述一个或多个收端参考神经网络的空口信息；
    从所述一个或多个收端参考神经网络获取一个或多个输出结果，所述一个或多个输出结果由所述一个或多个收端参考神经网络分别基于接收到的所述空口信息得到；
    将所述一个或多个输出结果和所述发端数据作为所述性能度量函数的输入，以得到一个或多个性能度量值；
    根据所述一个或多个性能度量值，以及所述性能判断准则，确定所述目标神经网络是否满足所述输出条件。
  6. 如权利要求4所述的方法，其特征在于，所述目标神经网络为发端目标神经网络， 所述一个或多个参考神经网络为一个或多个发端参考神经网络；以及
    所述基于标准化的参考数据集、一个或多个参考神经网络和性能判断准则，确定设计得到的目标神经网络是否满足预设的输出条件，包括：
    将从所述参考数据集中获取到的发端数据输入至所述发端目标神经网络中，所述发端目标神经网络用于根据预定义的数据结构对所述发端数据进行处理，以生成目标空口信息；
    将所述发端数据输入至所述一个或多个发端参考神经网络中，所述一个或多个发端参考神经网络用于根据预定义的数据结构对所述发端数据进行处理，以生成一个或多个参考空口信息；
    将所述一个或多个参考空口信息和所述目标空口信息作为所述性能度量函数的输入，以得到一个或多个性能度量值；
    根据所述一个或多个性能度量值，以及所述性能判断准则，确定所述目标神经网络是否满足所述输出条件。
  7. 如权利要求4所述的方法，其特征在于，所述目标神经网络为收端目标神经网络，所述一个或多个参考神经网络为一个或多个发端参考神经网络；以及
    所述基于标准化的参考数据集、一个或多个参考神经网络和性能判断准则，确定设计得到的目标神经网络是否满足预设的输出条件，包括：
    将从所述参考数据集中获取到的发端数据输入至所述一个或多个发端参考神经网络中，所述一个或多个发端参考神经网络用于根据预定义的数据结构对所述发端数据进行处理，以得到待发送给所述收端目标神经网络的一个或多个空口信息；
    从所述收端目标神经网络获取一个或多个输出结果，所述一个或多个输出结果是所述收端目标神经网络根据接收到的所述一个或多个空口信息分别生成的输出结果；
    将所述一个或多个输出结果和所述发端数据作为所述性能度量函数的输入，以得到一个或多个性能度量值；
    根据所述一个或多个性能度量值，以及所述性能判断准则，确定所述目标神经网络是否满足所述输出条件。
  8. 如权利要求4所述的方法，其特征在于，所述目标神经网络为收端目标神经网络，所述一个或多个参考神经网络包括一个或多个发端参考神经网络和一个或多个收端参考神经网络；以及
    所述基于标准化的参考数据集、一个或多个参考神经网络和性能判断准则，确定设计得到的目标神经网络是否满足预设的输出条件，包括：
    将从所述参考数据集中获取到的发端数据输入至所述一个或多个发端参考神经网络中，所述一个或多个发端参考神经网络用于根据预定义的数据结构对所述发端数据进行处理，以得到待发送给所述收端目标神经网络和所述一个或多个收端参考神经网络的一个或多个空口信息；
    从所述收端目标神经网络获取一个或多个目标输出结果，所述一个或多个目标输出结果是所述收端目标神经网络根据接收到的所述一个或多个空口信息分别生成的输出结果；
    从所述一个或多个收端参考神经网络获取一个或多个参考输出结果，所述一个或多个参考输出结果是由所述一个或多个收端参考神经网络分别根据接收到的所述一个或多个 空口信息生成的输出结果；
    将所述一个或多个目标输出结果和所述一个或多个参考输出结果作为所述性能度量函数的输入，以得到一个或多个性能度量值；
    根据所述一个或多个性能度量值，以及所述性能判断准则，确定所述目标神经网络是否满足所述输出条件。
  9. 如权利要求4至8中任一项所述的方法，其特征在于，所述性能度量函数包括以下一项或多项：均方误差、归一化均方误差、平均绝对误差、最大绝对误差、相关系数、交叉熵、互信息、误比特率、或误帧率。
  10. 如权利要求4至9中任一项所述的方法，其特征在于，所述可实现能力度量值包括以下一项或多项：空口信息的开销、神经网络实现所需计算量、神经网络实现所存储的参数量、以及神经网络实现所需的计算精度。
  11. 如权利要求4至10中任一项所述的方法，其特征在于，所述性能指标包括一个或多个性能指标，所述性能判断准则包括基于所述性能度量值与所述一个或多个性能指标的比较，确定是否满足所述输出条件。
  12. 一种通信方法，其特征在于，包括：
    第一设备基于与第二设备之间的空口传输条件，确定场景模式，所述场景模式用于确定适用于当前空口传输条件的神经网络，所述神经网络是从预先设计得到的多个神经网络中确定的，所述多个神经网络中的每个神经网络与一种或多种场景模式相匹配；
    所述第一设备向所述第二设备发送所述场景模式的指示信息。
  13. 如权利要求12所述的方法，其特征在于，所述多个神经网络中的每个神经网络是基于标准化的参考数据集、标准化的参考神经网络以及标准化的性能判断准则获得的神经网络。
  14. 如权利要求12或13所述的方法，其特征在于，所述方法还包括：
    所述第一设备接收来自所述第二设备的确认信息，所述确认信息用于指示对所述场景模式的指示信息的成功接收。
  15. 如权利要求12至14中任一项所述的方法，其特征在于，所述方法还包括：
    所述第一设备采用与所述场景模式匹配的神经网络与所述第二设备通信。
  16. 如权利要求12至15中任一项所述的方法，其特征在于，所述第一设备为网络设备，所述第二设备为终端设备；或，所述第一设备为终端设备，所述第二设备为网络设备。
  17. 一种通信方法，其特征在于，包括：
    第二设备接收来自第一设备的场景模式的指示信息，所述场景模式用于确定适用于当前空口传输条件的神经网络，所述神经网络是从预先设计得到的多个神经网络中确定的，所述多个神经网络中的每个神经网络与一种或多种场景模式相匹配；
    所述第二设备根据所述场景模式的指示信息，确定所述场景模式；
    所述第二设备根据所述场景模式，确定适用于所述空口传输条件的神经网络。
  18. 如权利要求17所述的方法，其特征在于，所述多个神经网络中的每个神经网络是基于标准化的参考数据集、标准化的参考神经网络以及标准化的性能判断准则获得的神经网络。
  19. 如权利要求17或18所述的方法，其特征在于，所述方法还包括：
    所述第二设备向所述第一设备发送确认信息，所述确认信息用于指示对所述场景模式的指示信息的成功接收。
  20. 如权利要求17至19中任一项所述的方法，其特征在于，所述方法还包括：
    所述第二设备采用与所述场景模式相匹配的神经网络与所述第一设备通信。
  21. 如权利要求17至20中任一项所述的方法，其特征在于，所述第一设备为网络设备，所述第二设备为终端设备；或，所述第一设备为终端设备，所述第二设备为网络设备。
  22. 一种获取神经网络的装置，其特征在于，包括：
    处理单元，用于基于标准化的参考数据集、一个或多个参考神经网络和性能判断准则，确定设计得到的目标神经网络是否满足预设的输出条件；
    输入输出单元，用于在满足所述输出条件的情况下，输出所述目标神经网络，输出的所述目标神经网络是应用于发射机的发端目标神经网络或应用于接收机中的收端目标神经网络。
  23. 如权利要求22所述的装置，其特征在于，所述处理单元还用于在不满足所述输出条件的情况下，继续对所述目标神经网络进行优化，直至获得的目标神经网络满足所述输出条件。
  24. 如权利要求22或23所述的装置，其特征在于，所述发端目标神经网络用于星座调制、信道编码、预编码、或信道信息的提取和压缩；所述收端目标神经网络用于星座解调、信道解码、检测接收、或信道信息的重构。
  25. 如权利要求22至24中任一项所述的装置，其特征在于，所述处理单元还用于从至少一套标准化的配置中确定适用于所述目标神经网络当前应用场景的配置；每套标准化的配置包括以下一项或多项：标准化的参考数据集、标准化的参考神经网络、标准化的数据结构、标准化的性能度量函数、标准化的性能判断准则、以及标准化的可实现能力度量值；
    其中，所述参考神经网络包括应用于接收机中的收端参考神经网络和/或应用于发射机中的发端参考神经网络；所述数据结构用于生成和/或解析空口信息；所述性能度量函数用于生成性能度量值，所述性能判断准则包括性能指标和所述目标神经网络的输出条件，所述性能判断准则用于基于所述性能度量值和所述性能指标，确定所述目标神经网络是否满足所述输出条件；所述可实现能力度量值用于选择匹配的神经网络和/或匹配的性能指标。
  26. 如权利要求25所述的装置，其特征在于，所述目标神经网络为收端目标神经网络，所述一个或多个参考神经网络包括一个或多个发端参考神经网络和一个或多个收端参考神经网络；
    所述输入输出单元还用于：
    将从所述参考数据集中获取到的发端数据输入至所述发端目标神经网络中，所述发端目标神经网络用于根据预定义的数据结构对所述发端数据进行处理，以生成待发送给所述一个或多个收端参考神经网络的空口信息；
    从所述一个或多个收端参考神经网络获取一个或多个输出结果，所述一个或多个输出结果由所述一个或多个收端参考神经网络分别基于接收到的所述空口信息得到；
    将所述一个或多个输出结果和所述发端数据作为所述性能度量函数的输入，以得到一 个或多个性能度量值；
    所述处理单元具体用于：
    根据所述一个或多个性能度量值，以及所述性能判断准则，确定所述目标神经网络是否满足所述输出条件。
  27. 如权利要求25所述的装置，其特征在于，所述目标神经网络为发端目标神经网络，所述一个或多个参考神经网络为一个或多个发端参考神经网络；
    所述输入输出单元还用于：
    将从所述参考数据集中获取到的发端数据输入至所述发端目标神经网络中，所述发端目标神经网络用于根据预定义的数据结构对所述发端数据进行处理，以生成目标空口信息；
    将所述发端数据输入至所述一个或多个发端参考神经网络中，所述一个或多个发端参考神经网络用于根据预定义的数据结构对所述发端数据进行处理，以生成一个或多个参考空口信息；
    将所述一个或多个参考空口信息和所述目标空口信息作为所述性能度量函数的输入，以得到一个或多个性能度量值；
    所述处理单元具体用于：
    根据所述一个或多个性能度量值，以及所述性能判断准则，确定所述目标神经网络是否满足所述输出条件。
  28. 如权利要求25所述的装置，其特征在于，所述目标神经网络为收端目标神经网络，所述一个或多个参考神经网络为一个或多个发端参考神经网络；
    所述输入输出单元还用于：
    将从所述参考数据集中获取到的发端数据输入至所述一个或多个发端参考神经网络中，所述一个或多个发端参考神经网络用于根据预定义的数据结构对所述发端数据进行处理，以得到待发送给所述收端目标神经网络的一个或多个空口信息；
    从所述收端目标神经网络获取一个或多个输出结果，所述一个或多个输出结果是所述收端目标神经网络根据接收到的所述一个或多个空口信息分别生成的输出结果；
    将所述一个或多个输出结果和所述发端数据作为所述性能度量函数的输入，以得到一个或多个性能度量值；
    所述处理单元具体用于：
    根据所述一个或多个性能度量值，以及所述性能判断准则，确定所述目标神经网络是否满足所述输出条件。
  29. 如权利要求25所述的装置，其特征在于，所述目标神经网络为收端目标神经网络，所述一个或多个参考神经网络包括一个或多个发端参考神经网络和一个或多个收端参考神经网络；
    所述输入输出单元还用于：
    将从所述参考数据集中获取到的发端数据输入至所述一个或多个发端参考神经网络中，所述一个或多个发端参考神经网络用于根据预定义的数据结构对所述发端数据进行处理，以得到待发送给所述收端目标神经网络和所述一个或多个收端参考神经网络的一个或多个空口信息；
    从所述收端目标神经网络获取一个或多个目标输出结果，所述一个或多个目标输出结果是所述收端目标神经网络根据接收到的所述一个或多个空口信息分别生成的输出结果；
    从所述一个或多个收端参考神经网络获取一个或多个参考输出结果，所述一个或多个参考输出结果是由所述一个或多个收端参考神经网络分别根据接收到的所述一个或多个空口信息生成的输出结果；
    将所述一个或多个目标输出结果和所述一个或多个参考输出结果作为所述性能度量函数的输入，以得到一个或多个性能度量值；
    所述处理单元具体用于：
    根据所述一个或多个性能度量值，以及所述性能判断准则，确定所述目标神经网络是否满足所述输出条件。
  30. 如权利要求25至29中任一项所述的装置，其特征在于，所述性能度量函数包括以下一项或多项：均方误差、归一化均方误差、平均绝对误差、最大绝对误差、相关系数、交叉熵、互信息、误比特率、或误帧率。
  31. 如权利要求25至30中任一项所述的装置，其特征在于，所述可实现能力度量值包括以下一项或多项：空口信息的开销、神经网络实现所需计算量、神经网络实现所存储的参数量、以及神经网络实现所需的计算精度。
  32. 如权利要求25至31中任一项所述的装置，其特征在于，所述性能指标包括一个或多个性能指标，所述性能判断准则包括基于所述性能度量值与所述一个或多个性能指标的比较，确定是否满足所述输出条件。
  33. 一种通信装置，其特征在于，包括：
    处理单元，用于基于与第二设备之间的空口传输条件，确定场景模式，所述场景模式用于确定适用于所述空口传输条件的神经网络，所述神经网络是从预先设计得到的多个神经网络中确定的，所述多个神经网络中的每个神经网络与一种或多种场景模式相匹配；
    收发单元，用于向所述第二设备发送所述信道状态模式的指示信息。
  34. 如权利要求33所述的装置，其特征在于，所述多个神经网络中的每个神经网络是基于标准化的参考数据集、标准化的参考神经网络以及标准化的性能判断准则获得的神经网络。
  35. 如权利要求33或34所述的装置，其特征在于，所述收发单元还用于接收来自所述第二设备的确认信息，所述确认信息用于指示对所述场景模式的指示信息的成功接收。
  36. 如权利要求33至35中任一项所述的装置，其特征在于，所述处理单元还用于采用与所述场景模式匹配的神经网络与所述第二设备通信。
  37. 如权利要求33至36中任一项所述的装置，其特征在于，所述装置为网络设备，所述第二设备为终端设备；或，所述装置为终端设备，所述第二设备为网络设备。
  38. 一种通信装置，其特征在于，包括：
    收发单元，用于接收来自第一设备的场景模式的指示信息，所述场景模式用于确定适用于当前空口传输条件的神经网络，所述神经网络是从预先设计得到的多个神经网络中确定的，所述多个神经网络中的每个神经网络与一种或多种场景模式相匹配；
    处理单元，用于根据所述场景模式的指示信息，确定所述场景模式；并用于根据所述场景模式，确定适用于所述空口传输条件的神经网络。
  39. 如权利要求38所述的装置，其特征在于，所述多个神经网络中的每个神经网络是基于标准化的参考数据集、标准化的参考神经网络以及标准化的性能判断准则获得的神经网络。
  40. 如权利要求38或39所述的装置，其特征在于，所述收发单元还用于向所述第一设备发送确认信息，所述确认信息用于指示对所述场景模式的指示信息的成功接收。
  41. 如权利要求38至40中任一项所述的装置，其特征在于，所述处理单元还用于采用与所述场景模式相匹配的神经网络与所述第一设备通信。
  42. 如权利要求38至41中任一项所述的装置，其特征在于，所述装置为终端设备，所述第一设备为网络设备，或，所述装置为网络设备，所述第一设备为终端设备。
  43. 一种计算机可读存储介质，其特征在于，包括计算机程序，当所述计算机程序在计算机上运行时，使得所述计算机执行如权利要求1至21中任一项所述的方法。
PCT/CN2021/102079 2020-06-30 2021-06-24 获取神经网络的方法和装置 WO2022001822A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
EP21833854.9A EP4163827A4 (en) 2020-06-30 2021-06-24 METHOD AND DEVICE FOR ACQUIRING AN ARTIFICIAL NEURAL NETWORK
US18/147,111 US20230136416A1 (en) 2020-06-30 2022-12-28 Neural network obtaining method and apparatus

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010615552.9A CN113873539A (zh) 2020-06-30 2020-06-30 获取神经网络的方法和装置
CN202010615552.9 2020-06-30

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US18/147,111 Continuation US20230136416A1 (en) 2020-06-30 2022-12-28 Neural network obtaining method and apparatus

Publications (1)

Publication Number Publication Date
WO2022001822A1 true WO2022001822A1 (zh) 2022-01-06

Family

ID=78981477

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/102079 WO2022001822A1 (zh) 2020-06-30 2021-06-24 获取神经网络的方法和装置

Country Status (4)

Country Link
US (1) US20230136416A1 (zh)
EP (1) EP4163827A4 (zh)
CN (1) CN113873539A (zh)
WO (1) WO2022001822A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115114395B (zh) * 2022-04-15 2024-03-19 腾讯科技（深圳）有限公司 内容检索及模型训练方法、装置、电子设备和存储介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108229647A (zh) * 2017-08-18 2018-06-29 北京市商汤科技开发有限公司 神经网络结构的生成方法和装置、电子设备、存储介质
WO2019197710A1 (en) * 2018-04-09 2019-10-17 Nokia Technologies Oy Content-specific neural network distribution
US20200026983A1 (en) * 2018-07-19 2020-01-23 Rohde & Schwarz Gmbh & Co. Kg Method and apparatus for signal matching
WO2020026741A1 (ja) * 2018-08-03 2020-02-06 ソニー株式会社 情報処理方法、情報処理装置及び情報処理プログラム
CN111340216A (zh) * 2020-02-24 2020-06-26 上海商汤智能科技有限公司 网络融合方法、网络融合装置、电子设备及存储介质

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111091175A (zh) * 2018-10-23 2020-05-01 北京嘀嘀无限科技发展有限公司 神经网络模型训练方法、分类方法、装置和电子设备
CN111105029B (zh) * 2018-10-29 2024-04-16 北京地平线机器人技术研发有限公司 神经网络的生成方法、生成装置和电子设备
CN109902186B (zh) * 2019-03-12 2021-05-11 北京百度网讯科技有限公司 用于生成神经网络的方法和装置

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108229647A (zh) * 2017-08-18 2018-06-29 北京市商汤科技开发有限公司 神经网络结构的生成方法和装置、电子设备、存储介质
WO2019197710A1 (en) * 2018-04-09 2019-10-17 Nokia Technologies Oy Content-specific neural network distribution
US20200026983A1 (en) * 2018-07-19 2020-01-23 Rohde & Schwarz Gmbh & Co. Kg Method and apparatus for signal matching
WO2020026741A1 (ja) * 2018-08-03 2020-02-06 ソニー株式会社 情報処理方法、情報処理装置及び情報処理プログラム
CN111340216A (zh) * 2020-02-24 2020-06-26 上海商汤智能科技有限公司 网络融合方法、网络融合装置、电子设备及存储介质

Also Published As

Publication number Publication date
EP4163827A1 (en) 2023-04-12
CN113873539A (zh) 2021-12-31
EP4163827A4 (en) 2023-11-08
US20230136416A1 (en) 2023-05-04

Similar Documents

Publication Publication Date Title
WO2022012257A1 (zh) 通信的方法及通信装置
WO2021217519A1 (zh) 用于调整神经网络的方法和装置
WO2022033456A1 (zh) 一种信道状态信息测量反馈方法及相关装置
CN109873665B (zh) 数据传输的方法和设备
WO2021142605A1 (zh) 用于信道测量的方法和装置
US20240030980A1 (en) Data transmission method and apparatus
CN115917980A (zh) 无线通信系统的终端、基站以及由终端和基站执行的方法
WO2021254119A1 (zh) 信道编译码方法和通信装置
WO2022001822A1 (zh) 获取神经网络的方法和装置
WO2022012256A1 (zh) 通信的方法及通信装置
CN114492784A (zh) 神经网络的测试方法和装置
WO2023092307A1 (zh) 通信方法、模型训练方法和设备
WO2023060503A1 (zh) 信息处理方法、装置、设备、介质、芯片、产品及程序
WO2023092310A1 (zh) 信息处理方法、模型生成方法及设备
WO2023115254A1 (zh) 处理数据的方法及装置
WO2023015499A1 (zh) 无线通信的方法和设备
WO2022267633A1 (zh) 信息传输的方法和装置
WO2024032775A1 (zh) 量化方法和装置
WO2023207783A1 (zh) 一种通信方法、装置及系统
WO2024061093A1 (zh) 一种上行预编码的指示方法及通信装置
WO2024046215A1 (zh) 一种通信方法及装置
WO2024027682A1 (zh) 特征信息传输方法、转换信息确定方法、装置和通信设备
WO2023231933A1 (zh) 一种通信方法及装置
WO2024008004A1 (zh) 一种通信方法及装置
WO2021207895A1 (zh) 一种用于传输上行信号的方法和通信装置

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21833854

Country of ref document: EP

Kind code of ref document: A1

DPE1 Request for preliminary examination filed after expiration of 19th month from priority date (pct application filed from 20040101)
ENP Entry into the national phase

Ref document number: 2021833854

Country of ref document: EP

Effective date: 20230103

NENP Non-entry into the national phase

Ref country code: DE